There is also a third big difference: Mars lacks a magnetic field. Mars' core cooled 4 billion years ago3, which turned off its magnetic field and allowed solar wind to erode the atmosphere. Don’t we need to restart or create a magnetic field to make Mars livable? If so, terraforming Mars may not be possible with today’s technology because restarting convection within a planetary core is something that nobody has a serious proposal for, and building an artificial magnetosphere would require enormous investment4.

Surprisingly, no, you don’t need a magnetic field to live outside on a terraformed Mars. Atmospheric loss is extremely slow, and other problems like radiation are more easily solved in other ways. There are many reasons why terraforming Mars would be difficult, but the lack of a magnetic field isn’t one of them!

Atmospheric loss is geologically slow

Without a magnetic field, Mars ‘leaks’ atmosphere to space very slowly. Fortunately, it leaks so slowly that it probably doesn’t matter.

One possible endstate for a terraformed Mars5 is to have a thin, breathable, oxygen-rich atmosphere that weighs about 6 × 10¹⁷ kg. Mars loses about 1 kg of atmosphere per second6. That means it would take 100 million years to lose about 1% of your terraformed Mars’ atmosphere7.

To put that in perspective, here are some things that will happen faster:

• 100,000 years - The next Hawaiian Island breaches the surface8

• 20 million years - Faultlines move Los Angeles adjacent to San Francisco9

• 25 million years - The East African rift opens up into a new ocean10

• 50 million years - The Mediterranean Sea disappears as Africa merges with Europe11

Why is this so slow? The young sun produced much more solar wind and high-energy radiation than it does today12, and still took hundreds of millions of years to strip the atmosphere. Today’s sun is much calmer. The engine that drove ancient atmospheric loss has largely switched off.

If we terraform Mars, the atmosphere loss levels would change, but not because of increased pressure. Atmospheric loss scales with the cross-section of the planet, or (r + z)², where r is the planetary radius, and z is the effective atmospheric thickness13. A thicker atmosphere adds negligibly to Mars’ 3,390 km radius. The solar wind doesn’t strip noticeably more from a thick atmosphere than from the thin one Mars has today.

It is true that changes in the composition of Mars’ atmosphere might significantly alter the loss rate. A warmer Mars would have more water vapor in the atmosphere, and ultraviolet light splits water to release hydrogen, which escapes much more easily than heavier atoms. As far as we’re aware, nobody has modeled hydrogen escape from a warm, oxygenated Mars. The rate would depend on details like cloud formation, which is also poorly understood and would benefit from further study. But even pessimistic estimates of the loss rate from a wetter, warmer atmosphere are in the range of kilograms per second, and the resulting water depletion would still only matter on geologic timescales, not human ones.

These timelines are so long that it is difficult to project what humanity’s technology might look like, and how small a problem atmospheric leakage might be. Once in-orbit refueling is possible, a Starship will be able to take 100,000kg to Mars, meaning we could restore the 1kg/s with about one Starship a day, a large but not unreasonable number of flights with a reusable rocket. “Topping up” Mars’ atmosphere, if we care to do so, will just get easier as technology advances.

The atmosphere is the radiation shield

A planetary magnetic field protects you from radiation. But it’s not the only way to do so. The atmosphere itself is good at blocking radiation because it puts a lot of mass between the surface and space. Quite modest increases in Mars’ atmospheric pressure are likely sufficient to block most radiation.

The amount of shielding you need depends on the type of radiation.

Solar energetic particles, which are the acute radiation hazard during solar storms, are effectively blocked by any atmosphere with surface pressure above about 12 mbar14. Mars’ low-elevation regions already have 12 mBar, and any proposed terraforming would push surface pressure well beyond it.

Ultraviolet (UV) radiation in an oxygen-rich atmosphere would generate an ozone layer through photochemistry, just as Earth’s atmosphere does. How much shielding you’d get from different levels of oxygen is not well understood, but Earth atmospheric models suggest that even 20 mbar of oxygen would be enough for most of the UV shielding15. This is a topic that would benefit from more research, since the Mars ozone layer doesn’t behave like the Earth ozone layer16. UV radiation can also be easily blocked by other kinds of shielding, and small amounts of water or soil reduce it to biologically trivial levels17.

The most challenging form of radiation to block are Galactic cosmic rays (GCRs). The proposed18 thin, breathable, oxygen-rich atmosphere would reduce GCRs to roughly ISS-interior levels19. That’s still more than on Earth, but GCRs are only 20% of the radiation exposure of ISS astronauts20. That’s still above Earth-surface background, but well within the range addressable by modest additional shielding or biological interventions.

Possible Showstoppers in Terraforming

There remain significant unknowns that might make terraforming Mars impossible. But the lack of a magnetic field is not one of them. Here are some of the outstanding possible showstoppers we’ve identified.

• The amount of nitrogen on Mars is limited, and biological nitrogen fixation would have to adapt to lower nitrogen concentrations to sustain a nitrogen cycle.

• There might not be enough electron acceptors on Mars to sink all of the hydrogen necessary to sustain a biologically generated oxygen atmosphere.

• Liquid water may migrate into the deep subsurface or high-altitude, though recent work suggests this is less of a concern21!

• A warmer and denser atmosphere might kick up more dust, blocking sunlight and cooling the planet.

Get Involved

Want to help us write a more formal paper on this topic? We’re using this blog post as a way to find our future co-authors! If you want to help us write this up, let us know. Email us magneticfield [ at ] pioneer-labs.org or comment below.

We’re especially interested to hear from you if: 1) You disagree with us! Let’s get it right. 2) You are well positioned to do more detailed ozone modeling or 3) you are well positioned to do hydrogen escape calculations or 4) there’s something else obviously missing you’d like to contribute to. Let us know!

Subscribe now

1

Laid out initially in Ansari, S., Kite, E. S., Ramirez, R., Steele, L. J. & Mohseni, H. Feasibility of keeping Mars warm with nanoparticles. Sci. Adv. 10, eadn4650 (2024). and followed up with subsequent publications. For broader discussion of this topic see the Mars Terraforming Research substack:

Mars Terraforming Research

Fostering a world-scale biosphere on Mars: a research agenda

Applied astrobiology…

Read more

10 months ago · 8 likes · Mars Terraforming Research

We’re excited to share a new technical report detailing an approach we’ve developed at Pioneer that we call BioBloom, a method for Barcoded Saturating Mutagenesis of an entire bacterial genome. BioBloom creates a tube of bacteria in which every possible single-base mutation to a genome is explored, and each bacterium carries a little DNA name tag (or “barcode”) labelling which mutation they are trying out. This tests a broader and more complete set of mutations than classic approaches where bacteria evolve the old-fashioned way, and allows us to measure the relative growth of millions of different mutants simultaneously with DNA sequencing. Altogether, BioBloom produces datasets where we’ve given every single mutation to a bacterial genome a shot, and measured whether it helps a strain grow in a new environment, and by how much.

Every genome sits in a vast space of possible variations, and BioBloom systematically maps all single steps in that vast space of variation, defining where the roads are, and where they appear to lead. The example above is a small slice of the data we obtained from growth in salty broth, showing two genes where mutations improved salt tolerance. The most successful mutations are clustered around catalytic sites and regulatory motifs, marked with orange and purple arrows.

To learn more, you can read our technical report on BioBloom, and its open peer reviews. If you want the actual E.coli libraries, you can get them from Addgene as a kit. Here we’ll take a bigger-picture view of why we created this technique and where we think it could be headed in the future.

Domesticating microbes, faster

BioBloom can help us adapt microbes to new tasks and environments faster than existing approaches because we don’t have to wait for mutations to occur or for one to “win”. This is important and timely for biotechnology, with microbes being used today to turn CO^₂ into food, replace fertilizer for crops, digest waste into renewable energy, mine precious metals, remediate toxic waste, and countless other goals.

In all these applications, we’re “domesticating” previously wild microbes. We give them tasks within their latent abilities, but with unfamiliar temperature, conditions, community, nutrients, environment, and we ask them to do their new job fast. Humankind has domesticated countless different plants and animals, painstakingly stumbling upon variation, curating variants, and breeding them over thousands of years. Microbes are too small to see, which makes it harder. We can pick out a corn ancestor that matures at the right time or produces more grain, but it’s difficult to pick out a bacterium that’s doing unusually well from a big fermentor, the roots under a field of crops, or a copper mine.

BioBloom helps explore and detect these elusive improvements to domesticate microbes faster, and pick the best microbe more often. By creating nearly all single mutations in a bacterial genome, we no longer have to wait for mutations to arise. DNA nametags allow us to quickly determine who in that pool of trillions of cells is doing well in this new environment, and by how much. Where existing approaches like Adaptive Laboratory Evolution (ALE) have succeeded at lab scale, they often have to wait for specific mutants to take over the whole culture before they become detectable. This is not practical for many scaled applications. There is also no guarantee that you end up with the best mutation, as random mutations occur over time, so a middling mutation that occurs first can easily win. BioBloom avoids many of these pitfalls and could be applied to improve microbes for all kinds of scaled applications, exposing the juiciest improved mutations rather than leaving it to chance.

A Depth of Data

But there’s more to this data than just picking the winners. The millions of measurements in a BioBloom dataset also tell you something about how the genome of the bacterium works, what’s in it, and how this microbe could be performing the given task better. This kind of data has been available for proteins for some time now, and “saturating mutagenesis” data has paid dividends: identifying important regions, clarifying enzyme mechanisms, and even uncovering entire unknown genes hidden inside others. Extending this approach from one gene to an entire genome extends these kinds of insights to whole cells. It uncovers components of the bacterial genome and important mechanisms we didn’t know existed. Better yet, it very concretely shows how they can be improved for a specific environment.

Saturation mutagenesis datasets for single proteins have been used to train AI models, and these models are successful at improving proteins- with this “labeled” data from experiments significantly outperforming models trained only on “unlabeled” data- sequences alone. At the same time, we are seeing some of the first AI models applied at the whole genome scale. Whole-genome datasets like Tn-Seq have been used to train AI models to improve understanding and improve synthetic designs. There are some unknowns when bridging this gap, but we’re optimistic that BioBloom can help us model and generate dramatically improved strains.

The Future of BioBloom

We’ve made our BioBloom libraries in E. coli available publicly on Addgene (Kit 1000000273) so that other researchers can use them. We see this as a successful example of an FRO-style nonprofit research project: BioBloom required some up-front risk and cost that made it unlikely that any academic lab was going to create it, but it was speculative enough that industrial biotech was unlikely to try it either. Producing it at Pioneer and sharing it is a bit more evidence for the value that nonprofit research entities can provide.

We believe BioBloom might replace traditional, labor-intensive laboratory evolution for some applications, and help deconvolute the mutations that result from others. When combined with other techniques, BioBloom can chart a course for how to make drastic changes to our favorite synthetic biology workhorse, including Genome Recoding, integrating new metabolism for fixing CO₂, using carbon-neutral feedstocks, or testing new-to-nature pathways altogether. Applying selection schemes like Biosensors or anti-nutrients to these libraries can more effectively target specific phenotypes for bio-production, rather than whole cell survival/growth. BioBloom could also be extended to study combinations of mutations, and some of the core technical requirements for this have been worked out.

E. coli is often where a new technique is proven first, but it should only be the beginning for BioBloom. The retron editing technology underlying BioBloom has been shown to work in other bacteria, so, with further development, the same approach can function across a broader set of organisms and be applied to a broader range of applications. We’d love to see BioBloom helping researchers understand which variants might help their strains perform in applied, scaled-up conditions, such as the production of food protein, replacing fertilizer, treating wastewater, mining precious metals, and remediating toxic waste. If you’re considering one of these projects or are interested in supporting any of them, reach out! We’d like to help make it happen.

For most of history, evolution was something that happened at its own pace. Every genetics textbook defines Forward Genetics as characterizing mutant organisms to understand their underlying genetics, and Reverse Genetics as making specific changes to see what happens. Now, with the ability to make almost any change we can imagine and read off their impact millions at a time, the distinction between forward and the reverse are blurred. BioBloom fully maps a single evolutionary jump, turning a blind search into a mapped landscape. And once you have the map, you can get a lot further!

By measuring how life adapts — to salt, to stress, to new chemistry, even to Mars — then we can begin to build organisms that help us take care of this planet, and maybe one day thrive on another.

Modern liquid handlers are very capable machines. Yet most of them are programmed using proprietary graphical user interfaces (GUIs) or niche scripting languages, making it difficult to reuse code, collaborate across teams, or integrate automation into the broader computational ecosystem of modern biology. When you compare these to the luxuries of modern languages, it’s obvious that lab automation should be programmed in real programming languages.

We run our automation using PyLabRobot, which is an open-source, hardware-agnostic Python interface for laboratory automation devices. Using PyLabRobot, we control our Hamilton STARlet liquid handler with Python code rather than the vendor’s VENUS software. This approach gives us access to the entire Python ecosystem, from version control and testing frameworks to data science and machine learning tools, as well as more direct control over the liquid handling robot and anything we choose to integrate with it.

The result is automation that is faster to develop, easier to maintain, and far more flexible than traditional approaches.

💬Open-Source Software Is Better for Lab Automation

By writing our automation in Python with PyLabRobot instead of proprietary vendor tools, we get

• Faster development and debugging. Interactive Jupyter notebooks let us test individual robot commands without running full protocols.

• Real version control. Git-based workflows with line-by-line diffs, pull requests, branching, and rollbacks (no more versioning with USB drives!)

• Easier onboarding. New team members only need basic Python skills to get up and running on everything. Plus, Jupyter notebooks provide a great gateway for those less experienced by combining code with Markdown text.

• Free scheduling. With integrated drivers, PyLabRobot itself can coordinate multi-instrument protocols, replacing the need for expensive scheduling software.

• Modular, reusable code. Small functions compose into complex workflows, rather than monolithic method files common in automation.

• AI compatibility. LLMs can generate and assist with Python-based automation far more effectively than proprietary GUI-based scripting languages.

• Community-driven improvement. Open-source development is distributed across many labs, not locked behind one vendor’s roadmap.

We’ve put this into practice across several workflows, including setting up inhibition assays and Illumina NGS libraries1, and contributed labware definitions and tooling back to the community2. Our goal is not just to automate our own lab; we want to help build a shared ecosystem of open automation tools.

🪨💸⛓️Here’s What a Normal Automation Stack Looks Like

A traditional lab automation setup typically involves multiple layers of proprietary software. Each instrument (liquid handler, plate reader, centrifuge) comes with its own vendor-specific control software, usually a GUI-based application. To coordinate across multiple instruments, you then need an additional scheduling layer (Green Button Go, Agilent VWorks, Cellario, OVERLORD, etc.). These can have an upfront cost of tens or even hundreds of thousands of dollars, depending on the complexity of your system, as well as annual service contracts.

The result is a complex software stack that is hard to learn. Protocols are also generally saved in proprietary binary or complex XML formats, making it difficult to compare changes across files. “Version control” in this case often means a file named FINAL_v3_ACTUALLY_FINAL on a shared thumb drive. Debugging failed runs requires re-running the entire method or creating a separate testing method just to investigate a simple change. Integrating with data analysis pipelines or modern computational tools requires either creating custom connectors between different software components or incurring the overhead of manually exporting .csvs.

This is the status quo. It works, but it’s expensive, slow to iterate on, and difficult to scale.

🏗️ 🧠 Our Automation Philosophy: Open-Source Modular Workflows

We replace all of those layers with Python. Using PyLabRobot, we write reusable functions that follow paradigms familiar to any software engineer. This approach mirrors how much of modern software is built. Larger, more complex systems are constructed using simpler, reusable components.

Development is faster this way. We run automation interactively through Jupyter notebooks, sending individual commands to the robot during development. Testing a small change, such as adjusting an aspiration height, changing mixing parameters, or tweaking a plate’s pick-and-place position, takes seconds. This dramatically speeds up debugging compared to vendor languages, and it means scientists who already use Python for data analysis can quickly learn to write, edit, and execute automation workflows.

Because our automation workflows are written in Python, Python itself becomes the scheduling software layer. This means the same script does both individual device control (liquid handler, plate reader, colony picker, etc.) as well as the higher-level coordination between instruments. Not only is this simpler, it is also avoids the need to purchase additional software solutions.

PyLabRobot also gives us access to real version control with git, which is generally unavailable for commercial systems. This allows us to track changes line-by-line, review changes to our protocols, and quickly rollback to previous versions in the event that something breaks. Collaboration is also much smoother. Instead of emailing (or using a USB to transfer) exported method files, teams can push and pull code directly from a shared repository. This simplifies working across multiple instruments or sites, and allows folks to edit methods simultaneously across instruments without losing any changes.

Our analysis tools also exist in Python notebooks, making it very easy to connect downstream pipelines directly to experimental workflows. An automated system can read plate absorbance measurements, calculate dilution steps to normalize wells, generate a liquid handler worklist, and execute the transfers all in a single script. AI coding assistants and LLMs are also far more effective at generating and debugging Python code than proprietary scripting languages, making this entire development loop even faster.

Finally, PyLabRobot is open-source. Open-source software projects often advance faster than proprietary tools because development is distributed across a community of contributors. Instead of one company (which may pivot or go out of business!) maintaining every feature, improvements come from many independent labs solving their own problems. This transparency makes it much easier for us to extend PyLabRobot’s functionality or debug issues than it would be with closed-source tools.

🧑‍🔬Join the PyLabRobot Automation Community

We believe the future of lab automation will look more like modern software development: open, collaborative, and built on widely used programming languages. With open-source solutions, we get to directly participate in improving the tools we use. That’s why we are contributing to the PyLabRobot community and sharing the workflows we build at Pioneer Labs.

If you’re interested in laboratory automation (or if you’re a programmer who wants to help improve experimental science 👀) we encourage you to explore PyLabRobot and join the growing community around open laboratory automation.

Subscribe now

Stay tuned to our Substack for some exciting upcoming science on:

• A technique for whole-genome saturating mutagenesis

• How to fully simulate soluble Mars soil chemistry in a lab

• The first microbes we engineered for better growth on Mars

In developing a roadmap for a green Mars, we considered climate, radiation, water, oxygen, and biology. We did not, until now, consider the maximum acceptable dimensions for a dragonfly.

In An Introduction to Mars Terraforming, we describe a low-pressure, oxygen-rich atmosphere of about 150 mBar that could support human life. Then we worked backward to the technologies and biological processes that might get us there. But we forgot to mention that such a high-oxygen environment might also support enormous insects, similar to the high oxygen atmosphere of the Carboniferous period on Earth.

This is an April Fools post, but it is only partly a joke. Mars is, first and foremost, too cold and dry for life.1 But once you start talking seriously about a living Mars, you need to consider a long tail of questions. Some of them are grand and philosophical, like the ethics of spreading life between planets. Others are much more specific, like whether a future Martian ecosystem comes with arthropods of alarming size.

🫁🐞 Why the bugs might be massive

We know from the fossil record that giant insects once roamed the Earth, and it’s been proposed that the oxygen-rich atmosphere at the time helped make that possible. Insects breathe through a tracheal system rather than lungs, so oxygen delivery has long been discussed as one factor that could constrain body size.2 In Drosophila and other model systems, changing atmospheric oxygen changes body size and tracheal development.3

A lot of this is quite theoretical. There is probably not a simple slider where 21% oxygen gives normal flies and 50% gives monsters. But once you start proposing unusual oxygen atmospheres, weird insects become an annoyingly legitimate thing to think about. During high-oxygen periods, Earth hosted some arthropod scales that most modern readers would prefer remain firmly in the fossil record.

🔥🌬️ Mars is not Carboniferous Earth

Our proposed Green Mars is not just “ancient Earth, but on Mars.” The ancient earth atmosphere that supported big bugs was dense and warm, while a green Mars would be lower pressure, lower temperature and lower gravity. If Mars ever gets insects, they would take millions of years to fully adapt and would likely be strange in specifically Martian ways, not just oversized Earth ones. Think enormous wings to bite into the lower density atmosphere and multi-layer insulative shells to keep them warm through cold nights.

The more immediate atmosphere-design concern is not giant insects at all. It is fire. Oxygen-rich atmospheres can be breathable, but they also raise flammability, a lesson NASA learned tragically in the Apollo 1 fire. In its official investigation, NASA identified the high-pressure pure-oxygen cabin atmosphere as one of the major factors in the disaster.4 Our proposed Martian atmosphere is much lower-pressure, which should ameliorate the risks.5 But this is the broader lesson here: once you redesign an atmosphere, and certainly once you redesign a planet, you inherit second-order consequences whether you want them or not. Giant dragonflies are more fun to imagine, but fire safety is the more immediate engineering problem.

🌱🪐 Building a biosphere is complex

Those second-order consequences are really the point of this post. The insect question matters not because Mars is destined to be overrun by giant bugs, but because it illustrates the many ecological and engineering details in need of consideration. Our own work so far sits much earlier on that ladder. In How to make a microbe for Mars, A Biologist’s Guide to Mars Dirt, and Our first target microbe for Mars, we’ve focused on the first practical steps for Mars-ready microbes, useful regolith media, and indoor, infrastructure-heavy bioprocesses.

As recent work on Mars terraforming emphasizes,6 the first problems are still planetary: warming enough for liquid water, enabling oxygenic photosynthesis, dealing with CO₂ availability, water and ice behavior, oxygen buildup, crustal redox capacity, and the sheer industrial scale required to change a planet. Big bugs are a convenient ambassador for the broader weirdness of all of the ways in which a Green Mars would not be Earth-but-less-blue. And some of its consequences will sound like jokes right up until they need a research program.

🔬🦠 Real things we’re actually working on

To be completely clear: we are not currently engineering Martian dragonflies. If you’re curious about what we’re actually doing, keep your eyes peeled for upcoming technical reports on:

• A technique for whole-genome saturating mutagenesis

• How to fully simulate soluble Mars soil chemistry in a lab

• The first microbes we engineered for better growth on Mars

• Open source lab automation and how we use it

As well as a roadmap for Pioneer’s next year of research!

Thanks to Niko McCarty for providing input on this piece!

Now we’ve collated the full calendar here, along with several honorable mentions that didn’t quite make the list the first time. We had great conversations and commiseration the first time, with people sharing stories of similar failures and offering helpful suggestions, and we’d love to hear more reactions! Diving into the nitty-gritty like this was genuinely fun, and we’d love to do (and see) more of this kind of open, behind-the-scenes science in the future.

Subscribe now

The Calendar 🗓️⛔

⛔ result 1: We’re building an assay to measure perchlorate ☠️ in our Mars media so we can evolve microbes to clean it up 🧫 ❤️‍🩹 But the assay uses dichloromethane to extract the dye… which dissolves the bottom of polystyrene plates 😅🕳️ Switching to polypropylene fixed the issue. 🧪✅

This is us adapting a published assay to expand the dynamic range of a colorimetric perchlorate assay. Somebody suggested that polyurethane would work just as well as polypropylene - something for us to keep in mind! We also had some problems transferring dichloromethane with a multichannel pipette - it tends to drip.

⛔ result 2: We want to feed our Mars microbes acetate as a carbon source, since you can make it from CO₂ + water⚡️🌖 BUT when bugs metabolize acetate, the pH goes up instead of down like with sugar.🧪We buffered in the wrong direction and watched pH shoot from 8 → 10😅

This one is exciting. If acetate consumption increases pH, could you feed acetic acid to a bioreactor to keep the pH neutral and supply additional nutrients for fed-batch or continuous growth?

⛔ result 3: We used high salt as our first stressor to shake down our library selection & data collection pipeline🧂🧫But pellets from high-salt cultures are either rocks or snot globs. Our robot couldn’t resuspend them even after 300 pipetting steps with wide-bore tips🤖💪🚫

Someone in the comments mentioned this is probably due to excess EPS production and that we might be able to fix it by washing in PBS a few times first. Something to try!

⛔ result 4: Different LB mixes are NOT the same😅For our first year, we used HiMedia, then had to switch to Difco, which turned cloudy when we added salt and autoclaved🌫️🧂Now we’re on Fisher LB, which stays clear. But the level of salt inhibition is different in every media😒🧫

The X-axis here is % w/v salt added to LB-miller, which already has 1% salt in it, and the Y-axis is a 24-hour endpoint OD. All lines represent the same strain across five different days using different media. They vary in salt tolerance by almost 2%!

⛔ result 5: Because of our LB drama, we switched to measuring salt tolerance in M9, a defined minimal medium🧫➡️📏 And then…😅Strains that looked great in high-salt LB showed no improved salt tolerance in M9. Turns out “salt-tolerant” depends a lot on the exact media🧂🤷‍♂️

Someone shared their experience evolving a different, fast-growing E. coli strain in a turbidostat and sent us growth curves in LB + 12% salt for comparison with our results! Handy to help us understand what a truly evolved strain looks like.

⛔ result 6: Our Mars Media is meant to mimic mixing 40 g of regolith into 1 L of water, shake + filter 🚀🪨But when we try to make a concentrated stock, the calcium sulfate (gypsum) crashes out… and might be dragging the phosphate with it 😵‍💫🧪 #marsmilk

We don’t like to filter this media because then we don’t know the exact composition. To make higher mixes and represent putting less water into more regolith we’ve made altered recipes with reduced calcium & sulfate content. Media formulation is complicated!

⛔ result 7: Turns out trace metals really matter in minimal media 🧫🧪Once we started washing our starter cultures to remove carryover nutrients, they just… stopped growing unless we added trace metals 🤦‍♀️🧬 Super important detail for adaptive laboratory evolution! ⚙️🧬

Just a few microliters of LB contain enough trace metals to support several subsequent cultures, so we didn’t see this problem until we started doing serial passaging in minimal media. Then our cultures slowly died off, and came right back when we supplemented with a trace metals mix. Now we’re building a regolith trace metal mix to represent what we think the real metal composition of sMRM is.

⛔ result 8: We found a mammalian expression plasmid (CMV promoter and all) as a contaminant in our bacterial selection experiments 😅🧬 Best guess? It hitchhiked in from the shared incubator-space centrifuge where folks do big DNA preps for mammalian cell culture 🚩🧪🙃

This plasmid didn’t have the priming regions to show up on our Illumina sequencing, but it kept showing up when we did nanopore sequencing - in libraries and elsewhere! We don’t use anything like it, so at first we thought it was contamination from our sequencing provider. They disagreed, and we tracked it down to a real contaminant that was all over the common lab spaces in the incubator.

⛔ result 9: We cloned libraries of tagmented gDNA into a suicide plasmid for transposition… only to discover carryover pUC plasmid in our commercial tagmentase 🧬😅 Leftover from protein purification, it totally dodged the suicide selection and snuck into our experiment 🕵️‍♂️🧪🛠️

The big consequence here is that we can’t electroporate these libraries into our final host strain, since plasmids with replicative origins get a massive boost compared to plasmids that need to integrate into the genome during electroporation. This bias doesn’t seem to occur with conjugation, so it doesn’t ruin these libraries; it just means the final integration of the library genome needs to happen via conjugation, not electroporation. Conjugation works with more species, so we were probably going to do this anyway!

⛔result 10: For genomic integration in C. necator, we do repeated rounds of integration & excision🧬🔁That means days of growth without selection, and C. necator is so slow-growing it usually gets contaminated 😭 Other strains outgrow hitchhikers, but not this little guy🐌🧫

Standard bacterial sterile technique isn’t enough here! This protocol involves a week of growth and several passage steps in rich media without antibiotics, and contamination at any point ruins the whole process. You need to follow sterility protocols that are similar to those required for mammalian stem cell culture to entirely avoid contamination.

⛔ result 11: We barcode the genome and use Illumina reads to track fitness of 5e6 strains at once 🧬📊 But at that scale you need to add µg’s of gDNA to avoid bottlenecks. We didn’t, so our variance ended up far above Poisson noise 🤦‍♀️📉🎲

This is a wonderful example of a new QC metric exposing a specific problem. If we’re doing random sampling from an underlying distribution, then the variance should be equal to the man. If the variance is higher, it means we’re bottlenecking the underlying distribution somewhere! Once we knew this we could go through every step to figure out where the issue was - in this case during our primarily illumina PCR step. When you barcode genomes, you have a lot of irrelevant DNA. To get 50x barcode coverage you need about a nanogram of plasmid DNA, but when your barcode genomes that requires more than a microgram!

⛔ result 12: Some of our libraries are already pretty skewed before we start selection 😅📊 Not always bad enough to trash, but it means we need much bigger bottlenecks to avoid chopping off all the low-abundance members 🧬🔍🧵

These libraries go through a lot of processing steps before selection, and it’s crucial to minimize any skew from those steps. Selection is supposed to create skew that carries information we can measure, but if the library is already skewed at the start, we lose dynamic range and have lower sensitivity to real fitness effects.

Honorable Mentions🥈🧫

In brainstorming this calendar, we had nearly 30 ideas, but not all of them made sense without a lot of context. Here are a few of the runner-ups we ended up cutting out.

⛔ result X: Tris is a bad buffer for cell culture 😬 We tried it in our Mars media to avoid phosphate buffer (it was already on the shelf), but Tris can cross membranes at high pH and mess with DNA & enzymes 🧬🧪 After a few experiments, we ditched it and switched to TES ✅

This was a simple lesson to learn. Don’t use Tris in cell culture! Use one of the Good Buffers instead. We’re not entirely sure why C. necator didn’t grow well below pH 7 in TES here. 🤷

⛔ result X+1: We sent our lab DI water to an external lab for ion analysis 🧪📦 and it came back with 2.5 ppm iron 😳🧲That’s ~2.3 µS conductance, still in the “DI” purity range ✅💧 but way more Fe than we expected. Remember, DI water isn’t truly pure 🤦‍♀️✨

Normally “DI” water is 5-1 µS, “MilliQ” water is 1-0.1 µS, and “ultrapure” water is below 0.1 µS. How pure does your water need to be? We re-measured this in our new space, and it’s much lower, with no detectible iron and a total conductivity around 0.12 µS. ✅

⛔ result X+2: We use DAP auxotrophy to kill off the donor strain after library conjugation 🧬🧫🚫…but our full-scale libraries included fragments with dapA, so donor cells grabbed them and survived 😅🦠✅Congrats to us: we accidentally ran an auxotrophy complementation experiment 🤦‍♀️🧪

We turned this accident into its own experiment in our previous post, if you’re curious about more details.

⛔ result X+3: To stop DAP-auxotrophy escape, we made our donor strain double auxotrophic—it needs DAP + 5-ALA to survive 🧬🧫🔒 But 5-ALA has to be really well washed out… and if it isn’t, donor cells carrying dapA fragments can sneak through for several passages 😅🦠➡️🧪

Here, we restreaked our freshly conjugated libraries onto +DAP, +ALA & +ALA & DAP. There should be a lot of growth on the plates with both supplements, but the plates with only a single one should have fairly minimal growth. But excess growth on the +Ala plate tells us that the conjugation carried over a lot of cells that require Ala to survive. 🤔 Washing the conjugation a couple of extra times fixed the issue.

⛔ result X+4: We do tons of high-efficiency electroporations on a GenePulser Xcell ⚡️🧫But if the shockpod lid isn’t fully closed… it will pretend it zapped your cells 😅 It prints a time constant and everything ⏱️✅ but delivers no shock 🙃Mystery “why won’t this transform?” solved 🤦‍♀️🔬

The electroporator is now covered with warnings to close the lid before you use it. ⚠️

⛔ result X+5: Turns out going from a strain name ➡️ a genome accession ID is way trickier than it sounds 😅🧬And if you want to automate it for hundreds/thousands of genomes? It becomes a whole pipeline problem 🤖📚🔍

We have multiple different strains of P. putida, C. necator, and E. coli, so the yellow highlights are notes to check those fields against the strains we actually use! ✅

Share

Closing Thoughts 🧠✨

This 12-day Negative Results Advent Calendar was a cathartic exercise for us and helped us reflect on the past year of experiments. Science isn’t a clean, linear path to discovery; it’s a messy, iterative process often riddled with tiny, often hilarious, technical failures.

Inviting people behind the curtain and getting a positive response showed us there’s a real hunger for this kind of open, behind-the-scenes scientific communication. Being transparent about failure can be just as informative, and definitely more fun, than only celebrating success!

Leave a comment

Comment with your thoughts on our negative results, or share some of your own!

Biology is famously complex, but we’re seeing that data at scale can give us a bird’s eye view of everything from the function of specific genes to the deep biological complexity of how those genes interact. In our previous post, we laid out the basic machinery of our functional genomics platform: fragment genomes, insert into a host strain to simulate horizontal gene transfer (HGT), measure fitness under selection, and use sequencing to build a genome-wide fitness map. As we continue these experiments, we’re beginning to see that the data reveals deep complexity in the underlying biology. We’ve come up with three vignettes about how genetic context, evolutionary distance, and library design shape the outcomes of functional genomics.

In this post, we’re sharing three lessons we’ve learned recently about functional genomics:

1. Neighbors matter! Sometimes beneficial genes can be tricky to spot when they are adjacent to detrimental genes.

2. Data acquisition at scale is possible. A single library can cover tens or hundreds of donor genomes, dramatically increasing throughput.

3. AI may help. Protein similarity is a good predictor of HGT success, and utilizing DNA language model embeddings may even improve prediction success.

By understanding the underlying capabilities and requirements of HGT, we are starting to understand the explanatory power of this kind of functional genomics for decoding biology. Future datasets may cover hundreds of organisms in tens of contexts, and by understanding what kind of data is useful we’ll be better able to ensure that those datasets are as rich as possible.

🔪 Dissecting the Effect of Genomic Context on Gene Fitness

We previously demonstrated that with a high-coverage genomic fragment library we can map fragments to genomic position and get a high resolution fitness readout of individual genes in a donor genome. With this sort of visualization, we found that the top hit in our H. elongata library was the gene galE, when inserted into E. coli and selected in a high-salt environment.

As a next step, we wondered if our libraries had high enough coverage to examine epistatic dependencies. With a library size of almost a million fragments, most genes will be covered by at least several fragments. Thus, could we detect fragments whose fitness strongly depends on the precise composition of present genes?

More precisely, we asked the following question: “Does our assay yield any fragments who register as fitness hits only if they cover a single gene, and not neighboring genes?” To calculate this, we recomputed our genomic fitness map where we defined the fitness of a gene as the mean fitness of all inserts covering that gene. However, this time we only used inserts that fully covered a single gene. So, now the fitness of each gene is defined purely by inserts that only cover that gene and not its neighbors.

In the plot below, we compare the previous genomic fitness map (top) with this single-gene-insert map (bottom). galE shows up as the top hit in both cases. Interestingly, the bottom plot shows a few medium peaks that do not appear in the top plot. This indicates the presence of genes that only contribute to fitness when their neighbors are not present! As an example, we annotate the gene aprD, which has a computed fitness of near zero when considering all inserts but has a mild fitness gain of near 1 when only considering inserts that only cover aprD itself. aprD is an alkaline protease secretion ATP-binding protein¹.

To verify this finding, we directly plotted the fitness distribution of all inserts covering aprD (including neighboring genes) and compared this with the distribution of inserts that only cover aprD itself. We see that in the former case, the fitness is very similar to the null fitness of the empty inserts, whereas in the latter case, the fitness is relatively high around 0.8 (corresponding to a fold-change of nearly 2).

Thus, our libraries have enough coverage that for a single donor genome, we can actually dissect position-dependent effects! In this case, we found an example of a gene that contributes to fitness only when found in isolation. If neighboring genes are included, the fitness gain actually drops out. Moving forwards, this indicates the importance of being sensitive to epistatic dependencies in our fitness hits.

🍲 Generating Pooled Libraries to Better Screen the Microbial World

One of our scientific goals is to leverage the genetic diversity of the microbial tree of life to source uncommon DNA for engineering polyextremophiles suitable for Mars. Traditionally, we built large, high-coverage libraries from single donor genomes. However, we came to realize that with average insert sizes of ~5 kb and bacterial genomes of ~5 Mb, our 10⁶ libraries are large enough to achieve sufficient coverage from multiple genomes within a single library. Our pooled library strategy enables higher-throughput construction, faster screening, and broader sampling of the microbial tree of life, particularly from unculturable microbes from e.g. metagenomic samples.

For an overview of how we originally built our libraries, see our previous Substack, The First Turn of Our Engineering Crank. To test whether pooled libraries could provide good coverage across multiple genomes, we began with eleven donor species representing both gram-negative and gram-positive bacteria. We developed a barcoding system to computationally track the various donor genomes in our pooled library². After normalization and pooling together in equimolar ratios to create a single and complex pooled library representing all donor genomes, we electroporate into a cloning strain, allowing us to capture a diverse array of genomic fragments in a single, multiplexed transformation. Our first pooled library had relatively even and comprehensive coverage of each donor genome – on average, each genome had more than 80% coverage and over 15,000 unique inserts associated with it.

In downstream analysis, we want to characterize each gene with a fitness score from a selection experiment. Crucial to that computation is statistical sufficiency, i.e. does each gene contain enough inserts to calculate a fitness score? Below we plot the average number of inserts that fully contain each gene for each donor genome. We see that while there is variability, most genomes have close to 10 inserts covering each gene on average.

In the future, we plan to scale up to pooling up to 100 genomes in a single library. To get there, we’ll want to improve the evenness and coverage of our pooled libraries to get better representation of each donor genome, for example through normalizing by genome size.

🔍 Identifying Determinants of Gene Transfer with a Complementation Assay

While the primary use of our platform is to identify extremophilic DNA fragments that confer a fitness benefit in harsh selection conditions, we can also repurpose the platform to study aspects of basic biology. Inspired by previous auxotrophy complementation analyses, we leveraged our crank to study aspects of horizontal gene transfer. Specifically, we used a complementation assay where we conjugated our 11-genome pooled library described above into an E. coli auxotroph with the core gene dapA knocked out³. We wanted to know if dapA homologs rescue endogenous dapA function in a way that is correlated with homology.

To tackle this, we first annotated all dapA homologs across the species in our pooled library as well as the inserts that contained them. For each homolog, we calculated its protein identity similarity to the endogenous E. coli gene as well as the average insert frequency at day 0 of our experiment (a proxy for gene transfer success). dapA homologs with high similarity to the E. coli dapA were very fit. For example, K. variicola’s dapA possesses about 90% protein similarity with E. coli and got the second-highest overall frequency. However, we also saw that the dapA homolog from C. necator, a phylogenetically distant relative of E. coli with less than 30% protein identity, still rescued the auxotrophy.

Building on these initial findings, we were curious if we could utilize more contemporary DNA language models to extend the analysis beyond simple protein similarity. Specifically, we used DNABERT-S, a model trained on extensive genomic data, to create high-dimensional embeddings for each homolog gene sequence. These embeddings can capture intricate patterns and relationships within stretches of DNA and may offer a more detailed (yet abstract) comparison than conventional sequence alignment. Our approach compared the sequence embedding of each homolog to the E. coli gene using cosine similarity where higher scores indicate a greater degree of similarity. This led to an even stronger positive correlation: the closer a homolog’s sequence embedding was to that of the E. coli gene, the more effectively it was able to rescue the auxotrophy.

Our dapA complementation experiment proved to be a valuable experiment for studying the factors influencing HGT success and functional rescue. This data allowed us to show a clear relationship between the similarity of a dapA homolog to the endogenous E. coli gene and its ability to rescue auxotrophy. At the same time, we did not have enough data points to try to model HGT quantitatively and there were some outliers (e.g. C. necator). In the future we would like to design larger experiments with 100+ homologs present to start building more sophisticated models of HGT. This provides a powerful opportunity for predicting gene transfer and functionality, even from phylogenetically distant organisms, as well as for applying modern DNA or protein language models.

Conclusion

These three results sharpen how we use HGT as a tool. Epistasis between adjacent genes can show up in our large-fragment libraries, pooled libraries let us multiplex many genomes in one experiment, and ML-derived sequence embeddings give us a better handle on which genes will work in new hosts. Those are exactly the ingredients we need as we push to much larger datasets—turning HGT-based functional genomics into a practical way to mine the microbial world for parts and assemble microbes suited for Mars.

Subscribe now

¹ The actual gene in H. elongata is unannotated, and the aprD annotation comes from a BLAST protein search.

² During cloning, we add organism-specific 8 bp barcodes when amplifying the barcoded backbone, creating uniquely indexed vectors that preserve the genomic DNA (gDNA) source. After characterizing our libraries with long-read PacBio sequencing, instead of aligning millions of inserts against all reference donor genomes at once (a very computationally demanding task) we first use the 8 bp barcode to identify the source organism, then align only to that genome. This sharply reduces the search space, improving speed, accuracy, and computational efficiency. While the benefit is modest with eleven genomes and requires additional front end work, this strategy will be essential as we scale to future massive data collection experiments consisting of pooled libraries containing up to 100 donor genomes.

To maintain clear separation between genomic sources during library construction, we work with uniquely 8 bp barcoded vectors and gDNA independently. Each organism’s gDNA is fragmented and cloned into its corresponding vector containing the matching barcode. Only after ligation are the barcoded libraries pooled, ensuring that all inserts are correctly linked to their source organism. Before electroporation into our cloning strain, we normalize each barcoded library by concentration to achieve balanced representation and prevent any single donor from dominating downstream cloning or sequencing.

³ This complementation assay was actually an unintended experiment! In our crank, we conjugate our libraries into recipient strains to genomically integrate library fragments. Our conjugation host, E. coli MFDpir, is an engineered diaminopimelic acid (DAP) auxotroph, meaning it is genetically dependent on an external supply of DAP for survival. We intended to use DAP withdrawal to impose a strong negative selection to eliminate the host cells following conjugation. However, our multi-species genomic library contained genes like native dapA or functional homologs that were capable of DAP biosynthesis. When these fragments were horizontally transferred into MFDpir, they successfully rescued the auxotrophy by restoring the missing pathway and allowed these host cells to survive and grow even in the absence of DAP supplementation. This unintended functional complementation event, while contaminating our primary screen, provided a unique opportunity to study complementation. It created a pseudo-auxotrophy experiment that allowed us to study the genomic factors that underlie HGT success.

To accelerate evolution to the human timescale, we developed new techniques to access new kinds of genetic diversity that we hoped would perform better than natural mutation. The issue with evolution in the wild is that genetic diversity arises over time from natural mutation, which is slow, limited to the DNA already present in the organism, and gives an unfair advantage to mutations that happen to occur first. We designed new techniques to address these issues, aiming to achieve better outcomes than Adaptive Evolution.

To our surprise, plain old Adaptive Evolution gave us the best results after a single cycle, picking the “low-hanging fruit” more effectively than our new engineering strategies. This suggests that we should let cells self-optimize first, achieving easy wins before applying our more advanced techniques.

🐎🏆 We did a side-by-side competition of three evolution techniques

Our three techniques are:

1. Adaptive Evolution, where diversity slowly arises from natural mutation. This most closely mimics the process in the wild and is the only technique where you diversify and select at the same time. It tends to produce multiple mutations per cell, not all of which actually matter.

2. BioBloom is a technique that aims to perform barcoded, saturating mutagenesis genome-wide: creating every possible single mutation first, then performing selection afterward. This tests more mutations faster than Adaptive Evolution and produces richer data about which ones work. But each cell only gets one mutation.

3. Additive Engineering instead adds DNA rather than changing it. We add a library of DNA from other organisms, giving our organism access to new genes or pathways it didn’t have before. Instead of evolving existing genes one mutation at a time, this method gives you access to completely new genes.

To test how best to evolve microbes for Mars, we ran a competition between these three evolution techniques to benchmark them against each other. We ran each method through a single cycle1 focused on improving E. coli’s salt tolerance, which is a major component of Mars stress. The goal wasn’t to create a winning bacterial strain. The goal was to quantify what each approach delivers on a realistic timescale and to look under the hood at how and why those gains are achieved. Here’s what we found:

Salt tolerance measured with an IC50 assay. Each of the three evolution techniques generated new strains that exhibit higher salt tolerance than the original wildtype “Parent” strain.

The headline result is simple and a little surprising: across this window, all three methods delivered similar absolute improvements, within about 1.5x of each other in terms of ∆IC₅₀. In the figure above, you can see Adaptive Evolution sitting at the top by a modest margin, BioBloom close behind, and Additive Engineering improved but behind the other two on the first pass. This tells us that all three techniques are effective, and we should consider all of them as we make a Mars Microbe. The question of which one is ‘best’ is more nuanced, as they pull different levers, in different ways, with big differences in laboratory implementation. Let’s dive in!

🍎🍊Comparing these techniques is nuanced because they’re apples-and-oranges

Each method results in strains with different types of genetic variation. Taking a peek under the hood, Adaptive Evolution will result in multiple single-nucleotide polymorphism (SNP) mutations, only some of which matter. BioBloom will make a single SNP that’s much more likely to be optimal. Additive Engineering will add new, horizontally transferred DNA that confers some benefit2.

The evolution techniques differ in the kinds of genetic variation they introduce.

The methods are also asymmetric in terms of difficulty, pace, and the type of human effort required. Adaptive Evolution requires a short amount of hands-on time every day to ‘passage’ bacteria, that is, initiate growth of a new bacterial culture with a little bit of yesterday’s bacteria. Fundamentally, Adaptive Evolution is limited by time, as bacteria divide only so quickly3. In contrast, BioBloom and Additive Engineering both require more sophisticated molecular biology upfront, but then can repeat cycles of diversification and selection faster thereafter4.

To discuss the outcomes of each experiment in a little bit more detail:

Adaptive Evolution gave us the largest bump in salt tolerance simply by letting E. coli grind through a month of daily passaging - including weekends! Random SNPs piled up: some helpful, some hitchhiking, and the net result outperformed the other approaches, but at the price of time and genetic clutter. Because every winning clone carries several changes, we still have to disentangle which alleles actually matter. Some of them are likely to be neutral or even negative hitchhiker mutations, which can make strains brittle and overspecialized to the selection environment. We also expect the benefits of Adaptive Evolution to plateau over longer time periods; diminishing returns are predicted by evolutionary theory and have been seen in other projects.

BioBloom doesn’t wait for mutations. We built a barcoded library containing virtually every possible point mutation, ran the whole set through selection in a few days, and read out the tags to spot the strongest single-site winners. Those clones are less fit than the multi-mutant strains from Adaptive Evolution, but they reveal a broader set of important genes than we saw in Adaptive Evolution, and can quantify the actual fitness of each mutation. The catch is that BioBloom currently only works in E. coli and requires substantial upfront effort before it can be used in a new organism.

Additive Engineering tries a different angle - dropping in foreign DNA rather than tweaking what’s there. We introduced libraries of extremophile DNA into cells via plasmids or transposons and observed modest gains for both approaches. But most of the transposon-integration winning strains turned out to have random pieces of DNA that inserted into the very same loci already flagged by Adaptive Evolution and BioBloom. True “new pathway” wins were present, but not in the majority. They’ll probably shine later, once native mutations hit their ceiling. But in this experiment, horizontal gene transfer proved less valuable than simple SNPs.

In summary, it’s tough to come up with a completely fair way to compare these three techniques because they’re apples-and-oranges in terms of speed, amount of effort, and the nuance of how widely each method can be applied in diverse organisms5. But doing a side-by-side bakeoff gives us a benchmark for comparison and deeper insight into how each method works, along with lessons about how to apply them.

🔎🗺️ Takeaways on how to evolve Mars microbes

Our goal is to accelerate evolution to adapt life to a new planet on a human timescale, not a geologic one. From the above experiments, we’ve learned that you should start with Adaptive Evolution because it’s quick to launch, easy to scale, and reliably harvests the “low-hanging” mutations that give the biggest early pay-offs. Any strain that comes out of this phase may need some cleanup and validation, potentially with BioBloom to help isolate which SNPs matter and which are hitchhikers. Then, with the native genome polished, we can use Additive Engineering to introduce new pathways and surpass the limits of the original organism.

Past that, we’re back in speculative territory. The best overall strategy is probably to switch between Additive Engineering and Adaptive Evolution, integrating a batch of novel genes before running a short cycle of Adaptive Evolution to domesticate them. Inserts that merely tweak existing regulation will fade, leaving pathways that genuinely extend the toolkit to help a Mars microbe thrive under poly-extreme stress.

💽💻Could AI do even better? Stay tuned!

All of the work above explores how well different experimental techniques can generate improved strains. One major missing piece is iteration. Each time you evolve an organism, you gain data that can help you next time. Closed-loop experimentation can take many forms, but we’re especially interested in whether AI methods can utilize the vast datasets generated by BioBloom and Additive Engineering to speed up the process.

For BioBloom, we’re writing a more detailed technical report that will be coming out soon. There, we’ll elaborate on the technique, characterize the coverage and statistical power of the library, and dive into some of the specifics of our results for this salt selection. Stay tuned and subscribe to our substack to get notified when this technical report is released!

Subscribe now

For Additive Engineering, we’re currently working on a proposal for a large-scale dataset that will create a “microbial phenomics atlas” by characterizing the fitness of thousands of genes from hundreds of organisms across multiple hosts and conditions. Based on these results, we’re planning to do that screening with plasmids. We’re still seeking technical commentary on the proposal, so please reach out if you would like to get involved.

Leave a comment

And as always, we’d love to discuss these results and their interpretation. Drop us a comment below!

The 2024 workshop subtitle was, “Is terraforming Mars feasible? How could it be done, and what might change our minds?” Much of the workshop was dedicated to working sessions where attendees did order-of-magnitude calculations, identified key technical unknowns, and began assembling the first integrated picture of how terraforming might be achieved using current technologies. By the end of the workshop, a new concept for how to terraform Mars had emerged — one that might actually work.

Subscribe now

How to terraform Mars

So how do you do it? We recently hosted the 2025 Green Mars workshop, and to quickly bring attendees up to speed on where we left off last time, we created the Introduction to Mars Terraforming, 2025 Workshop Summary. This document presents one option for how Mars could be terraformed. It tells the story of how it could be done backwards, beginning with a possible feasible end state and working backwards to how it could be accomplished.

What’s cool about the story told here is that it is consistent with everything we know today about Mars. And the news is good! We think that Mars could be green in our lifetime, but you’d still need an oxygen mask. Yet, many questions remain, and the above document calls out the key scientific unknowns, research priorities, and proposed alternatives.

Rethinking terraforming in the Starship era

Earlier visions of terraforming imagined an Earth-like endstate warmed by greenhouse gases, with a thick atmosphere and an artificial magnetic field. Now that we have more data about the realities of Mars and substantially improved spaceflight capabilities, we can make more detailed proposals for terraforming Mars in today's Starship era.

What, specifically, is different?

A Green Mars is likely to have a thin atmosphere and alpine climate. Our summary envisions a breathable, oxygen-rich atmosphere at ~150 mbar O₂, which can be generated entirely from resources already on Mars. This is not an Earth clone, but rather a thin, life-supporting envelope that still exhibits large day-to-night temperature swings but blocks most radiation. Such a state would allow people to live outside on the planet’s surface. Unlike previous terraforming ideas, this endstate might be quite doable with our current spacefaring capabilities and technology: you don’t need to crash asteroids into Mars, nuke the ice cap, or create an artificial magnetic field!

Heating can be relatively fast; oxygenation cannot. Several approaches (aerosol IR-scattering “glitter,” distributed solar reflectors, and solid-state greenhouses based on transparent insulating materials) could raise global or regional temperatures on a timescale of decades. However, building up oxygen in the atmosphere biologically is a slow process, and order-of-magnitude estimates suggest that it would take millennia for planetary oxygenation under conservative productivity assumptions. However, large local domes could be oxygenated within a few years. That means that within our lifetimes, we could have a green planet, covered in a thriving, global biosphere, where ‘homesteading’ of domed human habitats can happen on human timescales.

Leave a comment

Critical unknowns might be showstoppers. A short list of unknowns deserves special emphasis:

• Water sinks. On a warmed Mars, water might become trapped as ice at cold, high-altitude locations or in aquifers. Additional climate modeling is needed to understand this risk.

• Redox capacity of the crust. Oxygenation requires a sink for hydrogen; if Mars’ crust lacks sufficient electron acceptors (inorganic carbon, ferric iron, sulfate) to bind liberated hydrogen, atmospheric O₂ could be limited regardless of biological productivity. This is a geochemical constraint with considerable implications for the feasibility of making Mars green.

• Regolith chemistry and toxicity. Perchlorates and high salinity are real practical constraints on early biological steps. Understanding the spatial variability of nitrates and other bioavailable elements, data currently available from only a few sites, is essential before proposing large-scale biological interventions.

• Human health and engineering risks. The long-term health impacts of high O₂ partial pressures, aerosol toxicity, and radiation effects under the proposed engineered atmosphere require serious study before humans could walk around on the surface for long periods.

A new discipline in formation

The Introduction to Mars Terraforming, 2025 Workshop Summary is explicitly framed as one possible story for how to make Mars Green, not a definitive plan or the only option. It is anchored in the data, but acknowledges many of the potential alternatives and unknowns that might be “showstoppers.” This is an intentional signal that this narrative should be critiqued and revised upon the acquisition of new data.

A striking moment from the workshop occurred when we presented this introduction, sparking a lively debate about carbon dioxide outgassing from Martian regolith. That discussion revealed new and relevant details we hadn’t fully considered, and it’s a great example of the kind of challenge and refinement we’re looking for. This kind of exchange is how a field matures, and we look forward to capturing these insights in the workshop proceedings and translating them into a more detailed research roadmap.

Taken together, these workshops and the accompanying documents transform terraforming from a handful of disparate papers across disciplines into a multidisciplinary research field.

The field grows

In the past year and a half, this area of research has gained momentum. The 2024 Workshop Proceedings were released, followed by an opinion piece in Nature Astronomy that advocates for serious study of the topic. Multiple attendees of the 2024 workshop published significant research in the past year, including new proposals to easily warm Mars with aerosol nanoparticles and create self-scaling organically generated habitats.

The 2025 workshop had twice as many attendees as the 2024 workshop. The number of scientists working on terraforming projects has roughly 5x’d since the last workshop, in large part thanks to funding from the Astera Institute. As our colleague Edwin Kite put it, “the field went from non-existent to tiny.” This year’s workshop included a larger group of planetary scientists, biologists, engineers, and funders to take stock of what we know and sketch a research roadmap that will take terraforming research through the next step change: from tiny to small.

Shared curiosity

Ethics and policy surfaced repeatedly at the workshop. The question “should we?” remains both vital and unresolved. However, many technical goals are shared regardless of one’s philosophical stance. Whether the desired endpoint for humanity’s interaction with Mars is limited (crewed local research outposts), intermediate (paraterraforming via large greenhouses), or global (planet-scale warming & oxygenation), the early science and technology needs are substantially overlapping. In all cases, we need better climate models, robust ISRU for habitats made from local materials, validated biological tolerance limits, and de-risking of essential hardware with flight tests.

Benefits abound closer to home as well. Other participants emphasized the near-term, non-terraforming payoffs: climate modeling, closed-loop life support, and low-resource biomanufacturing have immediate terrestrial benefits. Even if global Mars terraforming is never pursued, the work strengthens Earth science and technologies for remote, resource-constrained environments. This shared upside helps justify collaboration across philosophical lines.

What comes next

We’re happy to share this document now, but more follow-ups from the workshop are pending:

1. Compile 2025 workshop proceedings that document the workshop’s technical findings and assumptions,

2. Create a research roadmap that prioritizes experiments and data collection (regolith surveys, atmospheric-escape modeling, aerosol toxicity tests, ISRU flight tests, cost and time estimates), and

3. Launch a grant program to catalyze work that can answer the high-value questions quickly.

Today, there are ten times more research questions than there are scientists to work on them, and that means there’s a high-leverage opportunity to use strategic thinking and funding to advance the field rapidly. Pioneer Labs is helping to coordinate these efforts and will publish the proceedings when they are ready. Stay tuned!

Do you have opinions?

We want to get feedback on our work faster than traditional scientific publishing will allow. That’s why we’re posting updates here. Please do comment below if you have thoughts on this post, and sign up to get future updates!

In particular, we would love to hear your thoughts on:

• The Introduction to Mars Terraforming. We welcome critiques, suggested revisions, and additional showstoppers! We’d also love to hear if you can provide some more insight into any of those questions.

• Suggestions for research projects to develop or de-risk critical technologies.

Leave a comment

We’re hiring

Pioneer Labs is recruiting! We’re looking for a Lab Manager and an Operations Associate to join the team in late 2025. Check out our website to see open roles, apply by November 7th, sign up for alerts when new jobs are posted, and to learn more about our work. If you know someone who would be excited to join the team, please pass along these opportunities!

Biomanufacturing has long been attractive for Mars missions due to its potential flexibility—biology can produce everything from building materials to therapeutics. However, the sheer number of possible bioproducts has created decision paralysis and made it difficult for the field to focus on a first application. In this post, we describe our diligent process for selecting the first microbial bioprocess to develop for Mars. We focused on options that require minimal infrastructure, use only in situ Martian resources, and yield products useful to the first human mission. Today, our top candidate is a microbe that produces bioplastic entirely from Martian dirt, water, and air, using only infrastructure that fits in a refrigerator-sized box. We now have a first concrete target for engineering, as well as a framework for reasoning about future biological engineering efforts to fill greenhouses and beyond. 🚀

Our top candidate

The ideal first bioprocess for Mars would have these features:

1. 🏗️ Minimal infastructure: A bioprocess run on Mars will need to do more with less. Less complex equipment with a smaller footprint will be easier to launch and more likely to work.

2. 💯 100% Mars: As much as possible, the bioprocess should use raw materials from Mars as the input. Consumables that need to be brought from Earth inherently limit how much can be produced.

3. 💪 Impactful product: Making anything using only Mars materials would be great. Even better if the product made is valuable and impactful for uses like enabling crewed missions and future terraforming efforts.

Using these goals, we imagine a setup wherein a small reactor (🏗️) can be fed with raw Mars materials: dirt; water; and air (💯), and produce bioplastics that can be turned into fibers, films, or hard-plastic building materials (💪)!

More specifically, we have identified two microbes, Cupriavidus necator H16 and Pseudomonas putida K2440, that grow in our recipe for Mars Regolith Media. In addition to the nutrients from regolith, they require a carbon source. We’ve chosen acetate, which can be easily sourced from the Mars atmosphere with electrocatalysis using a small-footprint device. Both organisms can produce PHA, a bioplastic that can be used for 3D printing. This is the first target for a viable bioprocess that can turn abundant Martian materials into useful products at unlimited scale.

Below, read about how we developed a framework to guide our strategy and decisions around building a bioprocess on Mars.

Developing a Mars Bioprocess Framework

🏗️ Infastructure needed for the bioprocess

When you are designing hardware for space, simpler is better. Less complexity means a lower chance of something breaking. Fewer, smaller devices means less upmass. For this reason, we need to be judicious in our choices about which devices we want to rely on to shield biology from the extreme environment of today’s Mars and generate growth conditions that permit reasonable yield.

Will there already be any existing infrastructure for us to ‘plug in’ to? The answer to this question depends on when we might flight test this biomanufacturing system relative to the first human mission. Although there’s no single agreed upon design for the first human mission to Mars, all reasonable mission architectures involve pre-placing infrastructure to generate rocket fuel for the ascent vehicle ahead of time, before the humans even launch from Earth. In particular, the idea is to pre-place equipment that can convert CO₂ atmosphere and water into rocket fuel. Even if we end up flight testing a device before then, ice mining equipment and power generation are becoming mature technologies that could increasingly be obtained off-the-shelf. For this reason, we assume that there will be energy, water, and raw material extraction available as utilities on the surface of Mars.

To map the space of possible hardware, we enumerated devices that are used in terrestrial biomanufacturing and devices that can solve challenges specific to Mars (excerpt below). In many cases, terrestrial biomanufacturing uses devices that improve yield, but that could in principle be substituted for a biological solution instead. For example, Earth often uses the Haber process to generate NH₃ from atmospheric nitrogen because this is a type of nitrogen that is very easy for microbes to use. An alternative is to use NO₃^- directly, the type of fixed nitrogen in Martian regolith. The upside is that you require fewer devices; the downside is that microbes will grow more slowly.

💯 Growth in 100% Mars

All life as we know it requires carbon, nitrogen, phosphorus, and water to grow. A key to making biomanufacturing useful on Mars is unlocking growth that sources these key ingredients from raw Martian materials as much as possible. The space science community calls this in situ resource utilization (ISRU), and it is a key technology to enable scalable space missions.

In an earlier post, we defined a Mars Regolith Media that could be made by mixing 40 grams of regolith with a liter of water. Since then, we’ve learned how to grow two strains, Cupriavidus necator H16 and Pseudomonas putida K2440, using Mars media as their only source of nitrogen and phosphorous! (As expected, they do grow less well when sourcing these nutrients from Mars media due to the additional toxins that are present like salt, perchlorate, heavy metals, etc. Engineering is in progress to improve their growth in Mars media, stay tuned.)

The primary additional requirement is a source of carbon and energy that could be produced on Mars. One option is acetic acid, which can be produced with electrochemistry from compressed CO₂ from the Martian atmosphere using a device the size of a microwave (see top row of the infrastructure table. These units exist, for example, from Lectrolyst. They’re also in use terrestrially.). Other options1 may include using photosynthetic organisms that can source their carbon directly from CO₂ and gain energy from sunlight.

One major wrinkle we spent a long time debating is the challenge of controlling pH on Mars using in situ resources. Most microbes are only capable of optimal growth within a fairly narrow pH range, and as they metabolize food they often change the pH, quickly leaving the optimal range and halting growth. For this reason, if you want to achieve reasonable bioprocess yields, you almost certainly need a way to maintain pH within the optimal window. In the lab, it’s easy to do this with buffering chemicals like Tris, TES, etc., but these chemicals can’t easily be sourced on Mars without pretty extensive chemistry infrastructure. An alternative option is to use the chloralkali process to split Martian brines into a strong acid and base that can be added to your bioreactor as needed. If you are using acetic acid as your feedstock, another option might be to control pH exclusively with acetic acid addition: the metabolism of acetic acid raises the pH, which can then be lowered again by feeding with more acetic acid.

Together, all the essentials - carbon (air), nitrogen and phosphorus (dirt), and water (ice) - can be sourced directly from Mars, as can convenience chemicals to control pH.

💪 Bioproducts for Mars

The benefit of biomanufacturing with fully in situ resources is that it can, in principle, be conducted indefinitely. The longer the manufacturing device is operating on the surface of Mars, the more useful products it can accumulate in anticipation of human arrival. For this reason, the best first bioprocesses are ones for which essentially an infinite amount of the product would be useful. Want to tile the entire planet with greenhouses? Build them with biomaterials.

Another key feature we looked for is simplicity. There are many possible bioprocesses that require a series of organisms or intermediates. For example, production of food likely requires two steps: first, perchlorate remediation with a primary organism - second, conversion of biomass from the first step into edible protein using a secondary organism. The best first bioprocess will be one that keeps it pretty simple.

Putting it all together

Our ideal Mars bioprocess would 🏗️ require minimal infrastructure, 💯 only use Martian materials as input, and 💪 produce an impactful and valuable product. Making a good choice requires making tradeoffs between these conflicting priorities. Each specification interplays with all of the others, and the search space of bioprocesses and products is very large.

Leave a comment

We’ve talked before about our engineering roadmap that envisions engineering microbes that tolerate increasingly challenging Mars-like conditions, starting with ‘indoor’ biomanufacturing of PHAs, proceeds to more complex organisms that can thrive in a minimal greenhouse, to future organisms that will form the base of the ecosystem on a Green Mars, likely beginning with photosynthetic organisms that grow in ice-covered lakes. Over time we can decrease our reliance on infrastructure to create favorable growth environments 🏗️, instead relying on the biology itself to tolerate extremes 🦠. Eventually, you decrease your infrastructure requirements to zero and have organisms that can grow entirely by themselves outdoors. Those are the organisms that can green the planet.

Our first target is a balance between asking too much and asking too little of the biology. We will challenge this microbe to tolerate toxins in Martian dirt and utilize its nutrients directly, but we won’t yet challenge it to tolerate daily freeze thaw or anaerobic growth, both of which would massively slow growth and reduce yield. This balance means our first microbe will de-risk key technologies needed for terraforming, while also still being practically useful for the first human mission.

Our 2025 focus is a single-microbe, single-step PHA production process operating in a controlled indoor environment. In 2026, we’ll begin bioprospecting for microbes that can grow in more ‘outdoor-like’ conditions—such as minimal greenhouses—and start exploring multi-step bioprocesses. For example, a photosynthetic microbe might convert CO₂ into biomass, which can then feed a heterotrophic microbe that produces a useful product. These layered processes expand the range of bioproducts we can make and begin to resemble full biomanufactories: systems where microbial processes are chained together to build, recycle, and sustain each other. In the long term, such multi-layered designs echo natural ecosystems—and may one day help make Mars a friendly place to live.

Learn more

We are not alone in thinking about the grand possibilities of manufacturing with biology on Mars. Many others have taken on this roadmapping challenge and arrived at similar or different conclusions due to different assumptions about constraints. If you want to learn more, a few great reads include:

• The case for biotech on Mars, 2020. A great rundown of how fairly traditional infrastructure-heavy biotech could be applied in a Mars context.

• Towards a Biomanufactory on Mars, 2021. A great discussion of multi-step bioprocesses and how it could all fit together.

• Microbial biomanufacturing for space-exploration, 2023. Surfaces some good information about the material requirements of space missions, including the expected quantities needed of plastics.

• Many other analyses include use of human waste products as an essential component of the microbial feedstock. We’ve neglected this angle because there’s an inherently limited amount of human waste on Mars, making it unscalable. Still, biomanufacturing may be useful for waste management. A couple good reads to check out in this vein include this one and this one.

Also, congrats to Robin Wordsworth on his recent Science Advances paper! It dives into the challenge of scalable manufacturing of greenhouses on Mars using biomaterials. This paper is a great follow-on from Robin’s great 2019 paper about how transparent insulators like aerogels could replace having an atmosphere.

Reach out!

As we continue iterating on this framework and testing new bioprocess ideas, we would love to engage others working on biomanufacturing! We would especially love to connect with folks working in downstream processing of products, which similarly needs to be simplified and accomplished with in situ resources. Please comment below or reach out!

We are continually growing our team and have a current opening for a Senior Data Scientist. We will be opening a search for a laboratory automation engineer soon. If you want to stay updated on future job openings, please fill out this form.

Subscribe now

Pioneer Labs is engineering microbial evolution to optimize performance on tasks such as biomanufacturing on Mars. But what's the BEST way to achieve this? Natural evolution, fully AI-driven design, or something in between? At Pioneer Labs, we believe in data-driven research. We aim to gather real, quantitative data on the fitness changes that result from evolution and engineering, then measure which approach is best. Furthermore, we need to gather this data at scale in order to enter the regime of big data and unlock the path for computationally driven projects using tools such as machine learning.

With an eye to these long-term goals, we are already establishing experimental platforms that will pave the way for big data efforts. This is the first of two articles that describe our bioinformatics approach, which yields quantitative measurements of evolutionary fitness at scale.

Previously we described our prototype engineering crank, which generates extremophile gDNA libraries and leverages evolutionary selection to find gDNA fragments that convey fitness advantages in high pressure environments such as high salt. Purifying out winners of these selection experiments yields high-fitness strains analogous to natural selection, only faster. With this experimental setup, we are already finding hits!

Can we do even better than just finding the top hits? What if we could quantify the fitness of every gDNA fragment in our library and look for not-quite-top hits? Can knowledge of medium fitness almost-winners help us understand WHY things are conferring fitness advantages? Conversely, can we learn anything from fitness losers? Furthermore, can fitness measurements at scale enable data-driven approaches with tools such as machine learning?

With modern DNA sequencing technologies, we can affordably sequence each individual fragment in our pooled library and measure evolutionary fitnesses in our selection experiment at the single-fragment level. Following the methodology used in the BOBA-seq paper, we use long-read sequencing to associate the full-length fragment (~10kb) with short barcodes, and we infer fitness by reading short-read sequencing to count barcodes at scale over many timepoints.

Subscribe now

The punchline: We can measure quantitative genotype → fitness data at the scale of 10⁶-member libraries every time we run an evolution or engineering experiment! This improvement scales up our hit detection rate from highest fitness winners to hits across the entire genome. Scientifically, these datasets extend our ability to detect selection-relevant genes and operons and unlock big data modeling approaches.

Experimental details

We previously used salt tolerance as a Mars-relevant phenotype to prove out reliably quantifying dose-response curves for individual hits. Here, we use salt tolerance again for our proof of concept experiment, replacing the dose-response readout with a high-throughput bioinformatics approach. We generated a large fragment library from the H. elongata genome, a known halophile that thrives in salty conditions. We sequenced the base library with PacBio HiFi sequencing to view the fragments and link them to randomly generated barcodes.

Our full experimental protocol is described in our previous post. Briefly, we transformed E. coli with the fragment library and grew them up overnight in normal LB media. Then we passaged them into 4% salt media for five days, taking a replicate from each day for Illumina sequencing prep

Before looking for scientific insights, we checked our sequencing results’ quality. Below we describe our preliminary characterizations.

Barcoded libraries of extremophile gDNA fragments

We can characterize our libraries in terms of fragment length and genome coverage via PacBio HiFi long-read sequencing.

Our cloning strategy should yield a library with the following properties:

1. gDNA sequence fragments of ~1-10kb

2. Even(ish) coverage of the extremophile genome with at least 100x depth

From the PacBio data, we can examine the distribution of mapped fragment lengths. We see that the vast majority of sequences are between 1kb and 8kb. Eventually we’d like to produce larger sequence fragments and wider size ranges – this is a potential area of improvement!

We can define coverage by calculating how many fragments map to each region in the genome. We see that the coverage is relatively evenly distributed, and that the depth typically ranges from 200x to 800x. It’s important to note that this isn’t sequencing depth, but rather genome coverage in our library – each position in the genome is covered by 100-600 unique gDNA fragments, each of which are then sequenced deeply. With coverage this high, we can make quantitative statements about the fitness importance of precise locations in the genome!

Characterizing fragment effects through sequencing selection experiments

We characterize our selection experiments in terms of barcode diversity and frequency trajectories across time via Illumina short read sequencing.

First, we can examine the strength of selection. Over time, we expect barcode diversity to dramatically decrease as low-fitness fragments are out-competed. This is clearly exhibited in the sequencing data – below we show the (log10) number of uniquely sequenced barcodes in both our fragment library as well as a control library possessing just barcodes with no fragments over time in a selection experiment. Both drop to a few hundred species by Day 3, but the fragment library has slightly more (~500 vs ~100) at later timepoints, especially at Day 5, indicating the presence of winners! Interestingly, there is some variability in the control library at Day 4, likely due to some heterogeneity in the cloning strain that confers slight fitness advantages.

Because we sequence every timepoint, we can construct frequency trajectories of individual barcodes over time. From this data we can already detect winners on the basis of increasing frequency over time by looking for trajectories that increase over time and then examining the mapped genes from our PacBio data!

In the below plot we show trajectories of two winning barcodes (with mapped gene content) along with three randomly chosen barcodes. We replicated our earlier hit of opuE. In addition, the clear winner here was galE, a new hit that was not discovered in our earlier experiments. We believe that’s due to fragment size – here our fragments are longer (~5kb) and the functional salt tolerance of galE fragments could require longer sequence fragments than our older library (~1kb) possessed. For both hits, the frequencies increased substantially over time compared to randomly selected barcodes, which quickly died out by Day 2. Losers will decrease over time due to selection pressure, whereas winners increase over time due to evolutionary fitness advantage.

Developing a statistical framework for detecting high-fitness hits at scale

While the biggest winners will be obvious, additional learnings can be gleaned from smaller-effect hits. How do we construct a statistical framework for doing so?

Due to unavoidable cloning inefficiency, many barcodes will lack an inserted fragment sequence – this is expected and even beneficial since they can then be used as an in-sample negative control. By leveraging this negative control, we can calculate a fitness z-score for each barcode. We calculate a log-fold-change in barcode frequency over time (for details, refer to the Boba-seq paper), and then normalize against the empty barcodes to derive a fitness z-score.

Below we show the distribution of fitness z-scores at Day 1 and Day 5 between the empty barcodes and the gDNA fragments. At Day 1 the distributions are virtually identical, whereas by Day 5 the fragment distribution is slightly left-shifted. This indicates that the average gDNA fragment actually imposes a slight fitness burden, e.g. due to wasteful production of nonfunctional RNA.

If we focus on the high fitness values, we see that at Day 5, there are many high-fitness winners present. However, even at Day 1 there are several barcodes that have substantially high fitnesses. The sequencing reveals the presence of subtle higher-fitness hits at early timepoints that we would not have discovered in the previous experimental scheme that relied on late timepoint winners fixing in the population.

Furthermore, because we have this distribution of empty barcodes, we can actually use that as a null distribution for statistical testing and label actual fragment-containing barcodes as statistically significant. For a conservative cutoff, we can use a z-cutoff of 3 (black dashed line), which corresponds to a roughly 0.15% false positive rate. Barcodes that have a higher fitness z-score thus register as significant.

Reproducibility of results

One of the most important aspects of generating large scale data is validating the technical reliability of said data. Below we highlight some control analyses to verify that the numbers we’re generating are indeed quantitatively reproducible.

✅ Technical replicates correlate

Examining barcode-level fitnesses between technical replicates validates the technical success of our bioinformatics-based fitness quantification.

Our experiment contained three technical replicates, and we expect a high degree of correlation in fitness values between replicates. First, we can examine how the fitnesses correlate per barcode. Shown below is an example replicate-replicate scatterplot at Day 3, where each dot is a single barcode. We see that the vast majority of barcodes are neutral-low fitness, but there are a few winners. The ones on the diagonal are reproducibly winning between replicates, whereas the “flanks” are barcodes that win in one replicate but not the other. This is possibly due to effects like clonal interference– the ultimate winners in any single replicate are somewhat random and when they start to dominate the pool, other potential winners may actually die out.

✅ Correlation is even stronger when you view the data ‘per gene’

One major advantage of having long read sequencing data is that now we can map the fragment sequences of each barcode to the reference genome. Thus, each individual barcode can be thought of as an “experiment” on a portion of the source extremophile genome, and multiple fragments covering the same region can provide independent measurements of the same underlying biological phenomenon. This poses the following question: does examining fitness effects at the “genomic region” level provide cleaner numbers than at the single barcode level?

To carry out this analysis, we take each annotated gene body in the reference H. elongata genome and assign it the average fitness of all barcodes that contain that gene body in the fragment sequence.

Collapsing barcodes onto genes improves the correlation between technical replicates and indicates that our selection experiments yield fitness data that are biologically grounded. Below is the same replicate-replicate scatter plot, except now each point is a gene, rather than a barcode. We see that the correlations are quite high, and the “flanks” have disappeared. Essentially, the stochasticity at the single barcode level is no longer present once we average out fitness effects and examine things at the genome level.

Overall, examining fitness at the gene level by aggregating across individual barcodes dramatically increases reproducibility.

Initial learnings

⏳ Early timepoints resolve moderate-fitness hits; late timepoints resolve mega-fitness hits

Selection is a complex, dynamic process, and our measured fitnesses are time-dependent due to stochasticity and clonal interference. At early timepoints, fitness effects begin to emerge but are weak due to less selection having taken place. At late timepoints, fitness effects are very apparent. Single barcodes begin to dominate the sample, and stochasticity is introduced since many species begin to die out due to clonal interference.

One core strength of this bioinformatics approach is the ability to construct distributions and conduct statistical tests. This lets us resolve statistically significant hits at early timepoints, pre-fixation. We actually observe more significant hits at earlier timepoints because no single/few fragments dominate the pool yet. Without sequencing, these would be very hard to detect.

Below is a plot of the number of barcodes with statistically significant high fitnesses over time. We see that there are more statistically significant barcodes at early timepoints. This is because, even though the super-high fitness winners emerge at late timepoints, this is at the cost of losing resolution of the less-high but still-significant fitness winners dying out. Because we sequence each timepoint, we can actually resolve these “medium” winners at early timepoints. Otherwise, if we only relied on experimental isolation of winners at the last timepoint, these medium hits would be virtually impossible to detect.

🗺️ We can map the fitness across the WHOLE donor genome!

Because of our long read data, we can map fragments back to the genome and generate fitness maps at genome scale with high resolution, unlocking the ability to query genes/operons of unknown function that can contribute to fitness advantages.

Below is a preliminary demonstration of this capacity. For each position in the reference H. elongata genome, we average the fitness (i.e. frequency fold-change value) of all barcode fragments that contain that position. From this we can generate a “peaks” map, along with a significance threshold based on a sampled null distribution constructed from the empty barcodes. Even by Day 1 we see some strong peaks emerge, such as galE, the strong winner from the above single-barcode trajectory analysis. opuE, the intermediate winner, displays a relatively weaker peak on this plot. Several other peaks are annotated with gene IDs and are candidates for further follow-up.

Most importantly, this visualization aggregates across barcodes to the genome level, providing cleaner and much more resolved information about fitness contributions from the H. elongata genome compared to noisy single-barcode analysis. This type of data transformation yields a more biologically grounded plot to inform follow up investigations of hits!

What’s next?

So, where do we stand? We have a bioinformatics platform that vastly upgrades our engineering crank to produce quantitative readouts of evolutionary fitness contributions of inserted extremophile DNA sequences at scale. Our initial characterizations indicate that these measurements are reproducible within an experiment, at the per-barcode and per-gene level.

There are a few important directions we can go from here, which we’ll describe in an upcoming Bioinformatics Part 2 post:

1. Can we trust these numbers in a wider setting? While we are getting quantitative numbers for each library member, we need to investigate the reproducibility across different libraries, as well as reproducibility across experimental paradigms. Furthermore, do these fitness measurements translate to practical takeaways such as improved growth in a non-competitive assay?

2. Can we extend our experimental regimes? This proof of concept was conducted for a single extremophile genome inserted into a single host strain under a single selection condition. Can we scale up our experimental and computational facilities to multiplex along each of these three axes, including other Mars-relevant environmental conditions?

3. Can we investigate applications of AI/ML? Our dataset generation has the potential to enter the big data regime, at which point we could consider using modeling techniques to try to predict fitness effects from underlying DNA sequences. How will we know if these efforts will bear fruit?

4. How do we follow up on hits? Our bioinformatics data give a nuanced lens into the genes and/or operons that confer fitness advantage. How can we dissect our results for basic science (i.e. can we investigate genes of unknown function?) as well as for engineering applications (i.e. can we determine the best sequence for strain engineering?).

Zooming out, we need quantitative data at this scale to achieve our goals of making a Mars-ready biomanufacturing microbe. The engineering problem is vast across multiple scales. Here we built a library from a single extremophile (H. elongata), selected under a single condition (salt tolerance), and inserted into a single chassis microbe (E. coli). In principle, we could source gDNA from the entire tree of life. Furthermore, Mars possesses a multitude of harsh conditions besides high salt that we’ll need to evolve organisms to tolerate. Finally, there are many putative chassis organisms we could consider using in addition to E. coli.

To explore all of these unknowns, we’ll need to build up our experimental and computational capacities to deliver quantitative and actionable results to inform future evolution experiments.

Right now, the next step is to start turning the crank and scaling up our data generation efforts. If you’re interested to read Part II, or to follow our technical updates in general, please subscribe!

Subscribe now

Data and code availability

If you’re interested in working with this dataset, you can access it on the Align Bio data portal here. Note that you’ll be required to make a free login with Align Bio. The dataset is hosted under the tab “Tesseract.” The portal contains the dataset, a link to our public Github, and a readme describing the data and code.

Work with us!

As these efforts continue, we would love to engage more with computational researchers. If you might be interested in this sort of genotype → phenotype data, please reach out and connect! We also welcome comments on these posts!

Leave a comment

We are also growing the team to support these larger efforts. We will begin a search for two roles — a new member of the computational team, and an automation engineer — over the summer. If you want to stay updated, please fill out this form.

Cite as: Pioneer Labs Reports (2025). Laying the groundwork for data-driven evolution. figshare. https://doi.org/10.6084/m9.figshare.28970489.v1

If you put Mars dirt and water in a greenhouse, could anything grow? It’s a tricky question with enormous implications for humanity’s future in space. On one hand, the soil (or regolith) on Mars is known to have toxins like perchlorate that would kill most living organisms, as we’ve written about before. On the other hand, Martian regolith is surprisingly rich in nutrients like nitrates — it comes pre-fertilized! Are there, on the whole, enough nutrients in Mars dirt that living things could overcome the toxins?

Ideally, to answer these questions we would take direct measurements of actual Martian regolith. Unfortunately, those measurements are rare — the only direct one is from 2008, when the Wet Chemistry Laboratory (WCL) on the Phoenix Lander measured soluble Mars dirt chemistry. Until Mars Sample Return is completed in the 2030s (if it doesn’t get delayed again), we won’t be getting more data on what’s soluble in the dirt.

We are interested in this question because we want to engineer microbes for Mars that can generate valuable supplies for the first astronauts like food, clean water and air, and bioplastic building materials. Transporting materials from Earth is expensive, so the more we can make use of local resources the better. In space science, this is called in situ resource utilization (ISRU), which refers to using the materials (i.e. soil, atmosphere, and water) found at the destination instead of paying to rocket all of that matter out of Earth’s atmosphere and on a half-year trip to Mars.

Below, we outline our current recipe for a soluble Mars Regolith Media (sMRM-Phx). With this recipe, we can determine whether or not microbes could rely on Mars dirt for key nutrients like nitrogen and phosphorous. We dive into where the data for this recipe comes from and what that implies for how certain we are of this recipe. Having a defined recipe will improve standardization across the astro-microbiology community and allow us to 🏆 benchmark organisms for Mars-readiness.

Subscribe now

🧪 Soluble Mars Regolith Media Recipe

Past Mars simulants have focused on physical characteristics, like the bulk mineral composition and texture. These are important for structural and robotic applications, such as testing rover tires 🛞, but not as useful for plant or microbial ISRU 🌱.

Our sMRM-Phx is an analog specifically made for biological assays to properly simulate the presence of the essential nutrients — nitrogen and phosphorus — that would be present after mixing 40 grams of regolith with a liter of water, then straining out anything insoluble.

Actually making media with this composition is tricky! Below is our recipe for sMRM-Phx media formulation. This required careful consideration of the acceptable range of ion concentrations and their overall charge contribution at ~pH 7.7. This recipe balances the ions using the salts, acids, and bases listed:

Briefly, we composed sMRM-Phx by taking the average of values from Phoenix lander measurements1 and two additional Mars regolith chemical stimulants2. See comparison of the concentrations of 11 key species below. The confidence intervals on sMRM-Phx represent the minimum and maximum values from the reported values in the papers.

What’s different?

There are four main differences between our sMRM-Phx and previous chemical simulants that make it better for biological ISRU applications.

NH⁴⁺ | While ammonium was measured and reported because it was expected contamination from the lander’s rocket fuel in the original Phoenix WCL studies, it was found to be below the limits of detection at all three of the sampling locations. It is not included in our simulant.

NO^3- | Since there have been no in situ measurements of soluble nitrates, not all previous simulants take this anion into account. Due to research on the presence of nitrate-containing minerals that are thought to be soluble and the measurement of soluble nitrates in synthetic leaches, we have decided to keep these in the mix.

PO₄^3- | Previous Mars simulants notably do not model the presence of soluble phosphates. This is important, since regolith is is the only possible in situ phosphorus source for living organisms. We chose to incorporate this due to promising data on the solubility of apatite minerals that are also found in Mars regolith.

HCO^3- | Due to the difficulty of incorporating soluble bicarbonate into our media formulation atom and charge balance calculations, and the sparse data on soluble carbonates, we decided not to include it in our recipe.

🪨 How do we know what’s in Mars regolith?

We have limited data about what’s in the regolith on Mars. There’s global spectroscopy data taken from orbit, lander and rover spectrometer measurements from the midlatitude Northern Hemisphere, and one soluble chemistry measurement from the Phoenix Lander in the polar region of the Northern Hemisphere. That’s it. In addition, scientists look to terrestrial analogues to learn more about the likely chemical composition of Martian regolith, and also to manufacture simulants. Here are categories of Mars regolith analogs that have been produced for different uses.

Of previously published soluble Mars regolith analogs, the Mars Simulant that was chemically mixed to mimic the Phoenix WCL soluble chemistry and the Mars Analog Regolith Extract (MARE) that is a leachate of a Phoenix Lander site mineral mimic are the most accurate. Our recipe above is a higher fidelity and more consistent version of these that averages across Phoenix WCL reports in the literature and supplements the things Phoenix didn’t measure but should be present based on other modeling and analysis.

Numerically, here is a breakdown of the composition of all of these various types of Mars regolith analogues:

⚛️ How we decided on our recipe

In more detail, here's a breakdown of how we decided on the concentrations of each of the major species in sMRM-Phx.

🍔 Nutrients

The 2008 Phoenix WCL experiment measured pH, conductivity, and various cations and anions. But it didn’t attempt to measure biologically relevant anions like sulfates, nitrates, phosphates, and carbonates, which would be the source for the biologically essential elements of sulfur, nitrogen, phosphorus, and carbon. We chose to include other modeled and terrestrially measured data in our recipe average. Earlier we discussed deviations between our recipe and the Phoenix WCL measurements, which overlaps with the nutrient sources described below.

Sulfates | After the initial preliminary reports of the Phoenix WCL data, later analyses used equilibrium models to explain the observed in situ cation and anion concentrations. In brief, the measured concentration of magnesium can be explained by the presence of magnesium sulfate minerals, and thus also soluble sulfate anions. Leachates of synthetic mineral mixes were also found to contain sulfates, supporting the retrospective analysis of the in situ data.

Nitrates | While the Phoenix WCL didn’t look for nitrates, we know from the Curiosity Rover Sample Analysis at Mars (SAM) data that there are nitrates in the regolith, and these would be highly soluble. In the same reports as soluble sulfates, leachates of synthetic minerals identify soluble nitrates supporting their presence.

Phosphates | We know there are abundant phosphate minerals on Mars from alpha particle X-ray spectrometry data. The question remains of how many of them are locked up in the rocks or are soluble. A 2013 study performed dissolution experiments on representative apatite minerals and measured soluble phosphate. This study is the basis of our sMRM-Phx phosphate concentration.

Carbonates | Similarly to sulfates, bicarbonate shows up in some, but not all equilibrium models of the Phoenix WCL data. This combined with the fact that it is not very soluble and that we do not plan to rely on these as a carbon source resulted in us not including carbonates in our sMRM-Phx recipe.

☠️ Toxins

In addition to Mars regolith providing essential elements for life, it also harbors soluble compounds and conditions that can be toxic.

Perchlorates | The main novel takeaway from the Phoenix WCL experiment was the presence of perchlorates in Mars regolith. Perchlorates are bleach-like ions that, if consumed, cause short-term thyroid problems and long-term fatality for humans, as well as stunted growth for plants, and microbial death.

Metals | While soluble metals have never been measured on Mars, we know from meteorites that Mars rock contains many trace metals, similar to Earth soil. For this first sMRM recipe, we have decided not to add trace metals; however, those will be important to include in future formulations as some metals are essential for microbial growth and others are cytotoxic.

pH | The Phoenix WCL experiment measured the soluble regolith to be slightly alkaline and quite close to physiological pH. However, location-specific variation could result in both alkaline and acidic compositions.

💨 Homogeneity

Soil on Mars will vary across the planet just like it does on Earth, but probably to a lesser extant. There should be more homogeneity in the “fines” — or essentially the topsoil — of the regolith because they are small enough to be weathered and mixed by the wind on a planetary scale. These wind-affected “fines” are what the Phoenix WCL measured. We think our recipe accurately measures the result of mixing 40 grams of these homogenous “fines” with one liter of water, then straining out everything that doesn’t dissolve.

🌱 What’s next?

Now that we have our liquid Mars media, we want to test to see if anything grows in it. To do so, we can measure an IC50 in sMRM-Phx to compare strains. This lets us determine whether wild-type and engineered strains can grow from entirely in situ-derived nitrogen and phosphorus.

You may have noticed that there’s no carbon source in our sMRM-Phx recipe! In the short-term, we will supplement with a carbon source that could be brought from Earth (i.e. glucose). Next, we’ll experiment with carbon sources that could be locally sourced from Mars’ 95% CO₂ atmosphere — either photosynthetically or electrochemically through feedstocks like acetate and formate.

💬 Do you have opinions?

We want to get feedback on our work faster than traditional scientific publishing will allow. That’s why we’re posting our science updates here. Please do comment below if you have thoughts, on this post or the attached lab notebooks, and sign up to get future updates!

In particular, we would love to hear your thoughts on:

• If you’re a biology researcher and interested in using our sMRM-Phx media, is there anything we can do to make it easier for you? Would it be helpful to onboard this recipe at Teknova?

• We would love to discuss with geochemists who have insight into how rock composition may vary across the planet. We just defined an ‘average’ Mars topsoil recipe, along with bounds on each biologically-important component. How do we now define a a ‘grid’ of possible media recipes that span the reasonable bounds?

• Would it be helpful for us to publish this in a traditional journal?

Acknowledgements

We would like to thank Edwin Kite, Alfonso Davila, Mohit Melwani Daswani, Sam Kounaves, Suniti Karunatillake, and Susanne Schwenzer for helpful discussions and emails about the geology and geochemistry of Mars; and the whole Pioneer Labs team for their input on both this substack and media formulation.

Strain engineering, or optimizing a bacterium for a specific industrial process, is the bread and butter of biotech today. However, the majority of strain engineering focuses on deleting genes, rather than what bacteria in nature do—swap DNA to acquire useful traits horizontally. 

We built our ‘engineering crank’ 🛠 to mimic natural evolution by using functional genomics methods (like Boba-seq) to acquire new genetic material. DNA from an extreme organism is chopped up 🔪, reassembled into a plasmid library🪡, transferred into a recipient microbe 🦠, and selected with directed evolution methods (like the Long Term Evolution Experiment) for fine-tuning 🤼. At the end, we’ll measure whether or not the genes that stick actually improve growth in an extreme condition with a secondary measurement 📏. 

To begin, we built and selected a first library of genes from salt-tolerant organisms in E. coli, and we’ll share those results here  🛠🧂. We hope this approach will allow us to engineer strains that are simply not possible with deletion-based techniques alone, including microbes for Mars!

🔪🪡 Making functional genomics libraries

An essential feature of our engineering approach is to mimic horizontal gene transfer by extracting DNA from a source microbe, putting it into a recipient microbe, and checking whether or not it changes the properties of the recipient. This sort of technique is called ‘functional genomics’ — not just looking at the sequence of DNA (genomics), but also what phenotypes it confers (function). Our goal here was to do a first proof of concept that uses functional genomics workflows to transfer properties related to Mars. 

🧂Salt tolerance is one of the five extremophile properties a Mars microbe would need, and is perhaps the most well studied and easy to work with. We chose H. elongata as our DNA source of choice. H. elongata is a very salt-loving microbe: it was reported to grow in a range of high salt conditions up to 32% wt/vol, and we were able to reproduce its growth in salty media up to 16% wt/vol. It is also easy to grow and standard kits can be used to purify its gDNA. 

Once we had extracted gDNA from H. elongata, the next challenge was putting that DNA into a recipient microbe 🦠. During this first proof of concept, we chose to insert DNA into E. coli, perhaps the most well-studied and easy-to-work with chassis. Techniques for inserting small pieces of DNA - not whole genomes - are more numerous and efficient, so the next step is to fragment the gDNA 🔪 and reassemble those fragments into a plasmid library 🪡.

🔪 Of the many ways to cut DNA, which one should we use? One common approach is to fragment gDNA by physical shearing with ultrasonication then clone it into a plasmid with blunt end ligation. This is great for generating libraries with ~10⁴ to 10⁵ members, but we would like to push toward libraries with 10⁷ members for covering metagenomic samples. We found a paper that co-opted Tn5 transposition and tagmentation, more commonly used for preparing shotgun sequencing fragments, to make functional metagenomic libraries by Mosaic Ends Tagmentation (METa) assembly. They show higher transformation and cloning efficiencies of METa compared to blunt end ligation methods, and suggest that METa is compatible with lower amounts of input gDNA, which would be useful for low yield environmental samples.

We used the METa protocol to build a library of H. elongata gDNA using a commercially available tagmentase following the manufacturer's protocol. We tried the tagmentation reaction at several ratios of tagmentase to gDNA and used an Agilent Tapestation to check the products. This is followed by a gap filling reaction, assembly reaction, and transformation, with various DNA clean-up steps along the way. The end result is a protocol for robustly cutting gDNA into fragments of our desired size range 🔪.

Once we had fragmented gDNA, we needed to clone those fragments into a plasmid library🪡. We had previously constructed a barcoded empty plasmid with 10⁷ variants, and used that as the backbone for our library. For our first time trying the METa protocol we used AarI Golden Gate Assembly and were able to generate two 10⁶ barcoded H. elongata gDNA fragment plasmid libraries in NEB DH10β electrocompetent cells 🦠. 

We submitted our assembled library for Whole Plasmid Sequencing by Plasmidsaurus, and the QC results are shown below. While we were happy with the library size, we were surprised to see that the inserts were relatively short given our input tagmented gDNA profile (mean >10kb), and thus very few of the inserts contained entire gene sequences. 

How not to do it:

🔪Tagmentation and 🪡 Golden Gate Assembly were not the only techniques we tried! We actually tried four different assembly methods.

🫠 Golden Gate Assembly (AarI) | This was our first choice because of its high efficiency and low error rate. You may remember that to effectively use Golden Gate Assembly, you need to be sure the enzyme recognition site doesn’t exist in your sequence. That’s why we used AarI instead of BsaI - there was a BsaI site in our barcoded plasmid backbone. Unfortunately, after being confused by our sequencing results (shown above) for a while, we realized that when we digested our tagmented fragments to generate the AarI sites, we were likely also digesting the 1180 AarI sites (only 7bp long) present on the H. elongata genome and therefore removing larger fragments from our library 😭! Ironically, despite the fact that our first AarI-generated libraries are non-ideal, we still moved ahead with selecting it because it had enough inserts long enough to contain both promoters and fully intact genes.

Electronic lab notebook: First time: fragment genomes (public copy) 

🤔 Gibson Assembly | Now we know why the published METa protocol used Gibson assembly for their libraries. We repeated our tagmentation protocol, but this time with Gibson homology sites appended onto the Tn5 ME sites as well as on our barcoded plasmid, and were able to get a 10⁷ library. But when we got our sequencing data back, most of the transformants didn’t have an insert. Looking closer, somehow during the assembly process, we both 1) did not get fragment insertions and 2) part of the plasmid backbone near the insertion site got deleted in a way that let the backbone ligate to itself. While we could not find a clear reason for why this happened, it could potentially be fixed by shifting the Gibson assembly site and trying again. In the meantime, we decided to try two other options in parallel.

Electronic lab notebook: Add barcodes during PCR (public copy)

👎 Homing endonuclease (I-CeuI) | To avoid cutting up the gDNA, we looked into using a restriction enzyme with a larger recognition sequence. We decided on I-CeuI, which has a 27bp recognition site that does not exist in the H. elongata genome. We ordered new primers for transposome assembly with the I-CeuI recognition site in it, repeated the whole tagmentation protocol, and got a 10⁵ library. Unfortunately, these also appeared to be mostly empty based on our sequencing results. Our hypothesis is that this older homing endonuclease might just have lower activity compared with other more optimized restriction enzymes. 

Electronic lab notebook: Rare restriction enzyme (I-CeuI) for Tagmentation (public copy)

What ended up working:

🥳 Uracil-Specific Excision Reagent (USER) |  In parallel, we also tried USER cloning, where we tagmented with primers that contained uracil. Then we gap-filled with a polymerase that can tolerate uracil and digested with a uracil-excision enzyme to make custom 10-base sticky ends. This yielded a library with 10⁵ members and with larger insert sizes, and we may rely more on this technique moving forward.

Electronic lab notebook: USER for Tagmentation (public copy)

Future improvements: 

While we are able to generate extremophile gDNA-sourced functional libraries, there are several areas in need of optimization that are ongoing and queued up for the future. We would appreciate any suggestions you may have!

🧬🧀 Extracting gDNA from other extremophiles and metagenomic (!!) samples

⚡🧫 Optimizing tagmentation → transformation protocol to generate larger libraries, compatible with metagenomes

🚀🧪 Mars-ifying our media beyond just salt

🔎🏔️ Bioprospecting for our next metagenomes

Leave a comment

🦠🤼 Selecting libraries of horizontally transferred genes

Once we’ve prepared libraries of extreme DNA, we need to add them to our recipient microbe🦠, and challenge them to tolerate an extreme condition and see what survives 🤼. To that end, we took two replicate libraries we had made (with the AarI Golden Gate cloning method), and selected them in high salt to see if we could pull out any genes that enhanced salt tolerance in E. coli.

These libraries are not ideal. The maximum insert size is about 2.5 kB, and the mean is only about 1 kb. Also, the selection happened in the DH10β cells we built the libraries in instead of a more relevant strain. But our purpose here is to demonstrate that the crank works, not to build a super-salty organism in one try. We also only had two independent replicates of the library, so to ensure replicates we split each library in four, then selected for 7 days in salt. 

It’s important to carry at least 100 million CFU through every single step, because otherwise we would ‘bottleneck’ our library and lose members because we didn’t take enough cells. We did some quick tests to ensure that the 500 uL of a glycerol stock and 100 uL of culture was sufficient for that, then went ahead with the experiment described in the figure below. 

How not to do it:

Our first attempt failed because the salt concentration was too high. We chose LB + 5% salt because DH10β seemed to grow reasonably well in 5% salt on the first day. But alas, the day 2 cultures grew significantly worse, and the day 3 cultures didn’t grow at all. If you select a population to stress that is too great, it will go extinct, not evolve! 🦖 Lesson learned.

What ended up working:

We tried a few other stringencies with one of our library replicates and saw that LB + 4% salt retained growth over several days. So we restarted the experiment from the intermediate glycerol stocks and selected 4% salt, which went smoothly. 

Future improvements: 

In the future we plan to move to plate-based selection methods, potentially using 24 or 48-well plates to increase volumes. This would make experimental setup easier and would let us try multiple selection stringencies at the same time. This would both ameliorate the risk of killing everything with too harsh a selection and allow us to compare how the results differ among different selection pressures.  

📏 Did it work?

Once we had our selected libraries, it was time to see if we’d managed to select for salt tolerance. First, we spot-checked some cultures from the last day of selection by miniprepping them and submitting to Plasmidsaurus for Whole Plasmid Sequencing. Then we checked the raw reads and saw that diversity was pretty low - with only one or two variants dominant in each culture.

With confirmation we’d selected for something, we turned to our handy IC50 assay to measure salt tolerance 📏🧂.

How not to do it:

First we ran the IC50 experiment directly on the cultures that emerged from the selection, comparing them against the plasmid expressing an mStayGold (a negative control that should not improve salt tolerance) and the Dr-IrrE (a positive control gene known to improve salt tolerance) in the same strain. While the data looks promising, we can’t conclude much. There could be SNPs in the genomes of these strains that are independently increasing salt tolerance, and we aren’t even certain if they’re clonal or still populations of multiple library members (hint, they’re populations). You can see the lab notebook for this experiment here: IC50 of 1st selection experiment winners (public copy).

What ended up working:

The right thing to do here is to re-transform our winners into a clean background so we can tell how much of the effect came from our plasmid. To do this we miniprepped all of the ‘winning’ cultures - which may still be mixed populations - and transformed them into the K-12 wild-type E. coli, then sequenced colonies. We found 7 unique genes, and tested a few replicates of each in our IC50 assay. 

Not only do these genes successfully transfer salt tolerance from H. elongata to E. coli, when we look them up, biologically their mechanisms of action largely1 make sense!

• Dr-IrrE - a regulatory metalloprotease. This is our positive control.

• fpR/trapT - a ferredoxin-NADP(+) reductase. Indeed, fpR mutants are salt sensitive.

• HELO_4410A - Ankyrin repeats. Perhaps important in plant drought resistance?

• DUF1486 - a 3-beta hydroxysteroid dehydrogenase. Indeed, steroids are associated with microbial salt balance regulation.

• hupB/lon - a histone-like DNA-stabilizer. These are important stress-response protein in mycobacteria.

• BetG - a betaine/carnitine transporter. These are osmoprotectants

• HELO_2306 - Serine protease? Not sure, mostly unannotated.

• opuE - a Sodium/proline symporter which works well in high salt to acquire nutrients.

In summary, we were able to successfully isolate some fragments of the H. elongata genome that increased salt tolerance in this assay. 

Future Improvements: 

Next time we probably won’t bother with checking the salt tolerance of the selection cultures, and will go straight to cloning into a clean background. We’re also continually improving the throughput, ease and accuracy of the IC50 assay, and I’m hoping we’ll be able to test even more variants soon. 

What’s next? 

This first experiment has laid the groundwork for showing that we can simulate horizontal gene transfer between microbes and use it to successfully ‘transfer’ a Mars-relevant property from one microbe to another. Making a Mars microbe that can survive many extreme conditions at once will require applying this technique many times to grab extreme properties from many different organisms 🦠 🚀.

Here are a few next steps that we’re excited about. 

1. In addition to selecting on salt we could select in high perchlorate. Can we make a salt-and-perchlorate tolerant microbe 😵 🧂? This would be valuable for biomanufacturing on human missions to Mars.

2. Which organisms should we source DNA from? We now have a wide variety of interesting libraries now with larger insert sizes and from more exotic extremophiles and even microbial community samples 🧫. We’re trying to decide which ones to select next! 

3. Eventually we want to iterate this library selection, to build up tolerance by combining genetic material from many different sources into one organism. We’ve got a couple (broad host-range) methods for this under development! 🛠

4. All of the library members were barcoded 🏷️ and we’ve done a sequencing run on them! It’s about a billion paired-end reads. Analysis of fitness effects 🏋️ in progress! If successful, this gives us an opportunity to gather big data on the properties of millions of genes in a single assay.

💬 Do you have opinions? 

We want to get feedback on our work faster than traditional scientific publishing will allow. That’s why we’re posting our science updates here. Please do comment below if you have thoughts, on this post or the attached lab notebooks, and sign up to get future updates!

In our experience there’s often a before and after to having good measurements. It’s like switching on the light in a darkened room, where you start to be able to see things and make progress that would have been impossible before. We call this traction, and it’s an idea both Erika and Devon have written about before. Traction is when you have questions to answer and the tools to answer them. It’s the place where good science happens.

Today we’re glad to share that we’ve achieved traction! We can precisely and reliably 📏 measure the 🧂salt tolerance of bacteria in liquid culture. If you want to see the other approaches we tried, we’ve included data on failed experiments, as well as some lessons learned. Finally, we’re teasing how we use this assay to measure improved salt tolerance in strains we have genetically engineered with functional genomics.

The key experiment: dose-response

We measure the relative inhibitory concentration of salt with a dose-response experiment. Our experiments show that a strain expressing the previously published salt tolerance gene Dr-IrrE has 10% higher salt tolerance than strains expressing a series of negative controls like the fluorescent protein mStayGold and the freeze-thaw tolerance gene dWhy.

This means we’re replicating the literature, and assigning quantifiable numbers to salt tolerance increases.

The general protocol for the salt tolerance measurement is fairly simple.

1. From glycerol stocks, start 4 biological replicates of each strain in 96-well plates.

2. The next day, OD normalize your overnights and seed them at OD 0.01 in a broad range of salt concentrations. Grow overnight in 96-well plates.

3. The next day, measure the final OD of your cultures and plot it, then fit a dose-response curve and extract the IC50 value & 95% confidence intervals

4. Normalize the IC50 values by dividing them by the negative control (mStaygold) value from the same experiment.

Past Literature & Failed Approaches

We didn’t start with the idea of doing dose-response curves for salt concentration. At first we took inspiration from the literature, where it is common to show growth curves for a salt tolerant strain & a control strain in salt as evidence of salt tolerance.

However, we quickly discovered that growth curves in salt are extraordinarily variable, and it’s nearly impossible to make this into a reproducible measurement. We tried to directly replicate the above graph on two different occasions and saw different results each time.

We stuck with growth curves because it’s the most common method in the literature. Other papers showed larger differences in LB (here, and figure 2 here), so we tried to replicate their results too. In the graphs below you can see that the Dr-IrrE probably improves salt tolerance, but it’s hard to say conclusively and one experiment shows the opposite result!

We tried a half-dozen things to make these experiments more reliable and reduce the noise, including fitting curves to extract summary statistics like lag time and growth rate, but nothing worked. Past approaches using growth curves to demonstrate salt tolerance aren’t incorrect, but we needed something more repeatable and quantitative for the experiments we’re doing.

The Details Matter!

There’s a lot of ways for things to go wrong in any experiment. Here’s a short list of things that we identified that are responsible for variation from day to day in our IC50 dose-response experiment.

1. Plate shaking conditions. This experiment is not especially repeatable if you shake the deep-well growth plates at 250 pm, the shake speed of a normal incubator. We purchased shaking incubators designed for 96-well cell culture plates that shake at 1000 rpm. This achieves more consistent growth and aeration across the plate, which is critical for this experiment.

2. Media prep details. We see variation in this experiment with different bottles of media, even if those bottles came from the same original bottle of powder. Normalizing to a control is a good way to reduce this variation, but carefully maintaining media standardization is important.

3. Starter culture conditions. We saw variation in the experiment if we grew the starter cultures differently, even if they were OD normalized before seeding into the experimental cultures.

4. Copy number variation. We’re using p15A for these experiments, a low-copy plasmid with low variation in copy number. However, it still does vary in copy number and thus expression & cell burden. We did these experiments with 4 independent transformants of each construct to to try to capture the average performance across high and low copy number variants. We do see that some of our Dr-IrrE clones do consistently better and others do consistently worse.

5. Bubbles in the plates. We used a mix of by-hand and automation to mix and seed our experimental plates, and then again to dilute them down for OD600 reading. During the transition to automation we found that the exact details of how you program the mixing & transfer steps matter, since otherwise you’ll sometimes introduce bubbles that will increase variability in the OD600 measure.

Experimental Development:

Want to learn more? Access our experiments directly! We’re including direct links to our lab notebooks for key experiments. This offers you access to the ‘primary source’ rather than us spending the time to produce a publication-quality version of the same data. Our hope is by doing this we can iterate our findings and feedback faster than is possible with traditional scientific publication.

1. The lab notebook that generated the data for the first experiment above is available here.

2. The full write-up for our normalization, including the summary of many more experiments, coefficient of variation analysis & plans for further improvement can be found here.

What’s next?

Now we can expand this framework to other conditions. The first and most obvious one is perchlorate, a toxic component of Martian soil. Based on early exploratory data, perchlorate looks very amenable to the same kind of dose-response analysis that works for salt 😵.

As part of our roadmap, we described the basic pipeline where we build libraries by taking DNA from intractable extremophiles, fragment it and insert it into more tractable hosts. We have started to get strains out of that pipeline, and our assay shows that they have significantly improved salt tolerance over our expression control mStayGold!

We’re still working out the kinks in our functional genetics pipeline, but now that we have a good assay to measure outcomes we can give it our full attention, and hopefully start iterating to continually improve salt tolerance! Expect a writeup soon about generating large functional genomics libraries from metagenomic samples 🛠!

Follow along

We want to get feedback on our work faster than traditional scientific publishing will allow. That’s why we’re posting our science updates here. Please do comment below if you have thoughts, on this post or the attached lab notebooks, and sign up to get future updates!

In other news, we’re looking for new Research Associate to join our team in December 2024 or January 2025. If you’re just finished your undergrad or masters and are excited about making microbes for Mars, you can apply here!

Thanks to Rachel Dutton, Jamie Bacher & Christina Agapakis for giving comments on this post, as well as the entirety of the Pioneer Team!

Cite as: Pioneer Labs Reports (2025). We can reliably measure salt tolerance. figshare. Online resource. https://doi.org/10.6084/m9.figshare.29042432.v1

Our mission is to create hardy critters that can make biomanufacturing robust, scalable and green — on Earth and beyond 🌎 🚀

A microbe that could grow on Mars would need to tolerate many extreme conditions that would kill most microbes we’re familiar with on Earth. In particular:

• 🧂salt - The dirt (or regolith) on Mars is very salty.

• 😵 perchlorate - Martian regolith contains 0.5-4% perchlorate (ClO₄⁻), a chemical similar to bleach.

• 🥶 cold - Organisms on Mars would need to survive low temperatures and daily freeze-thaw cycles.

• ☢️ radiation - The surface radiation dose of UV-C on Mars is very high, at 3 W/m².

A microbe growing outside on Mars must tolerate all of these extremes, which is quite the challenge! However, organisms can be partially shielded from the elements by growing them in indoor bioreactors or greenhouses. Doing so creates stepping stones toward a fully-outdoor organism, and is also independently useful for biomanufacturing and farming.

As a first step toward an outdoor Mars microbe, Pioneer Labs is engineering microbes optimized for indoor biomanufacturing on Mars. To do so, we are engineering commonly used chassis organisms E. coli and B. subtilis for in situ resource utilization (ISRU) on feedstocks containing Martian regolith. Regolith is a plentiful raw material on Mars that is a good source of nitrogen, but also contains high levels of salt and perchlorates are toxic to today’s chassis organisms. Our Marsified chassis microbes will unlock robust, scalable production of products like food, pharmaceuticals, and building materials for the first astronauts on the red planet.

How we’re doing it: Benchmarks 🏆, Measurements📏, and  Cranks🛠

We’re breaking this task down into three steps:

• Define a 🏆 Benchmark for Mars-readiness. Benchmarks are created in consultation with the broader space science and biotech communities. They synthesize key technical challenges into a measurable statistic. 

• Create assays to 📏 Measure performance of strains. We need robust wet lab techniques to quantify how different microbes perform in stressful conditions.

• Engineer microbes that beat the benchmark! To do this, we need to create engineering methods where you can ‘turn the 🛠 Crank’ to generate new strains with improved properties.

We expect to repeat the 🏆 → 📏 → 🛠 process many times before we get to a green Mars!

Our first project

1. 🏆 Develop the Indoor 0.0 Benchmark

We are currently roadmapping for Benchmark Indoor 0.0, designed to be an easy-to-understand and simple-to-execute test of suitability for indoor biomanufacturing on Mars. Like many biotech companies that want to incorporate a new low-cost feedstock into a manufacturing pipeline, we will approach this problem by quantifying the ratio of rich-media to Mars-regolith-containing-media that an organism can tolerate.

Get involved: we are looking for space scientists to provide feedback on our benchmark design, please reach out to indoor.0.0@pioneer-labs.org

2. 📏 Create assays to measure salt and perchlorate tolerance

We are optimizing assays for quantifying the salt-tolerance and/or perchlorate-tolerance of a strain. This measurement technique needs to be feasible (reasonably fast + inexpensive to conduct) and also quantitative (gives you the same number, within error, when you measure it twice). 

3. 🛠 Engineer strains using functional genomics and directed evolution cranks

We are creating engineering pipelines for transferring extremophile properties between organisms. We are experimenting with a functional genomics approach in which we extract DNA from intractable extremophiles, insert that DNA into E. coli and B. subtilis, and select for microbes that exhibit improved extremophile properties as a result of the new DNA. We will compare the utility of this approach to good, old-fashioned serial passaging in the presence of a stressor. Meanwhile, we are developing bioinformatics approaches to make data-informed decisions about which organisms to work with and how best to evolve them.

We aim to focus on this first project between opening doors in April 2024 and mid-2025. Expect posts in the next couple of months elaborating on our initial results.

Subscribe now

What’s next?

After the first project, we will expand in two directions.

Target more challenging environments on Mars, like growth in minimal greenhouses that experience daily freeze-thaw. These projects are currently $-constrained because they require funding for specialized equipment to simulate Mars-like extremes (low temperature, low pressure, radiation, etc) and funding to expand the team to work in non-model microbes.

• Target biomass farming in Martian minimal greenhouses 🌽

• Make a microbe capable of surviving and growing on the surface of Mars🌲

Target more specific use cases of extremophiles used ‘indoors’, both in space and on Earth. These projects are currently time-constrained because they require identifying, defining, and engineering toward a specific use case. 

• Simulate Martian ISRU conditions to a higher level of fidelity 🧑‍🚀

• Target the specific application of manufacturing bioplastic on Mars from atmospheric carbon 🧑‍🚀

• Expand our ISRU conditions to include the Moon 🌔

• Develop solvent-tolerant microbes for industrial biotechnology on Earth 🌎

Follow along

We want to get feedback on our work faster than traditional scientific publishing will allow. That’s why we’re launching this substack, where we’ll post science updates. We’ll aim to keep the blog posts as easy reading, and may also link out to longer-form technical reports built from our lab notebooks. Follow along!

Read more about Pioneer Labs on our website, https://www.pioneer-labs.org/.

💬 Comments welcome

Place your bets now for which extreme will be hardest 🧂😵🥶☢️ to crack! What specific use cases for extremophile engineering would you want to see us tackle next? Comments welcome - discuss below:

Cite as: Pioneer Labs Reports (2025). How to make a microbe for Mars, one step at a time. figshare. Online resource. https://doi.org/10.6084/m9.figshare.29042450.v1

Subscribe now