Machine learning meets molecular farming

By Brian Finrow and Jim Roberts

Today Lumen announced the first results of our long-running collaboration with Google’s world-beating AI/ML (machine learning) team. The gist: advanced algorithms helped us navigate a dizzying amount of data to double the productivity of cGMP biologics manufacturing platform. The manuscript, titled “Machine Learning Optimization of Photosynthetic Microbe Cultivation and Recombinant Protein Production” has been posted on bioRxiv pending peer review.

To their credit, the Google team only announces such collaborations if/when they prove to be interesting enough to publish. So this is the first time we’ve spoken publicly about the joint effort, and we’re proud to have made that cut with this initial project.

But this exercise wasn’t purely academic. Lumen aims to harness our unique drug development and manufacturing platform to treat and prevent diseases that afflict millions worldwide, from Covid-19 to traveler’s diarrhea. No other biologics platform has anywhere near the scalability to put these disease targets in reach. Google’s AI experts will accelerate our efforts to deliver on this vision.

Grappling with complexity

To our knowledge, this is the first time ML techniques have been used comprehensively to improve the productivity of a photosynthetic production system. There are a million potential extensions of this work, not just for Lumen, but for agriculture and beyond.

Spirulina is uniquely positioned to serve as a starting point for this effort precisely because its growth conditions are so extraordinarily simple. Most importantly, the growth environment is vastly simpler than plants, comprising just water and a few mineral salts. No sugar feedstock is required since the organism makes everything it needs through photosynthesis. Compare this with plants, which have complex multicellular structures and multistage life cycles, and, when grown scalably, must grapple complex soil variability and unpredictable weather.

Yet even this simple system is far beyond the ability of traditional 20th century optimization techniques to quickly navigate. Consider for a moment that — even after narrowing the inputs to just temperature, pH, mix rate, and the relative abundance of 15 or so mineral salts in the media, as we can with spirulina — the experimenter confronts complex 18-dimensional space to navigate if you want to search for the simultaneous settings that optimize productivity. Add on top light fluency (brightness) and spectrum (color), both of which are effectively infinitely variable — and then tack on the fact that all of these factors might vary over time during a single production run (due in part to increasing culture density) — and you get a flavor for why we’re so excited to be working with Google to get a handle on things.

If Lumen were blessed with unlimited time and resources, we could have run the more than 130,000 or so experiments required to find the same optimal growth conditions. But we’re not. And more importantly, the infants and children who are at risk of dying from diarrheal diseases cannot wait. The millions who have yet to receive a Covid-19 vaccine cannot wait. Working with this all-star team of collaborators is the best way to accelerate things in this life-and-death effort.

Doubling down on AI/ML

And we expect there’s even better stuff to come. This particular pilot project was funded in part by the Bill & Melinda Gates Foundation over two years ago, and we haven’t been standing still since then. This pilot project’s scope was quite narrow compared with what’s possible when Lumen’s unique technology platform is combined with Google’s cutting-edge capabilities. And of course AI/ML techniques are useful in many subfields of drug development, and over time we’ve expanded the relationship to include others. Expect to see more from us on discovery-side applications later this year.

But even just in the narrow productivity topic covered by this new paper there is immense potential. This is how today’s other news — significant new funding support from the US Department of Energy — is relevant.

The new DoE-funded work in collaboration with the National Renewable Energy Laboratory builds directly on the research described in the paper. As described, the first phase of this research focused on optimizing environmental factors that are relevant to the production efficiency of spirulina in our photobioreactors. Encouraged by those early results, the new $2 million DoE funding allows us to expand the collaboration in two ways. First and foremost, the new work will further expand the list of input factors we know are relevant, namely genotypic factors. As impressive as they are, spirulina’s environmental growth factors are a trifle compared to the design space represented by even a simple genome like spirulina’s. And of course we expect many genotypic changes to interact in complex ways with the environmental factors.

It’s awe-inspiring, but certainly not impossible. We know from the history of other leading biomanufacturing platforms (yeast, E. coli, and CHO) that tremendous improvements in productivity are possible. With CHO, for example, modern production systems are around 40X more productive than the early 1990s (and thousands of times more productive than the earlier “roller bottle” systems).

Some of these improvements come from optimizing environmental factors, but many of the big leaps are driven by the use of modern production strains that are optimized for growth in human bio reactors rather than out in the wilds of nature. Portland, Oregon-based AbSci just priced a $2 billion IPO for exactly this.

And this is after 40 years of prior enhancement. With spirulina, by comparison, we are still at the very beginning of the learning curve so big jumps should be possible in short order. And indeed this is exactly what we found in the research published today.

This has whetted our appetites, and we are optimistic that Google’s ML team will allow us to do this much faster than the decades of industry effort that were required to make similar strides in E. coli, yeast, and CHO.

Helping more people

But why bother? Remarkably, the Lumen system already meets the productivity needs for all of our commercial drug development programs. Dried spirulina powder can be purchased wholesale for less than $15 per kilogram today, and dose sizes in our clinical trials are typically in just the 100s of milligrams. So even allowing for the somewhat higher cost of making it indoors under pharma-grade GMP controls, our costs are already easily within range for our commercially oriented drug programs.

But the Gates Foundation is an ambitious group, to put it mildly. Their work with us contemplates daily dosing these products for developing world infants for 90 days at a total cost of less than $10. That will take more work.

Part of the solution will come from making more potent, broader-acting biologics cocktails. Here it’s handy that — unlike with injected and other systemic drugs — we are free to use all of the amazing recent breakthroughs in synthetic biology. And we’re using ML there too — expect to hear more about that soon.

But still, pulling this off will also require improved manufacturing productivity. And the severe trauma of losing an infant to these diseases is happening every day of every week, so we can’t afford to wait.

Solving more problems

As aggressive as those cost targets are, even this isn’t the limit of our ambition. Consider for a moment the fact that spirulina is already the only photosynthetic microbe farmed at ultra-large scale globally. And by large-scale, we mean tens of thousands of tons annually — orders of magnitude larger than global cell culture production for traditional biotherapeutics.

Beyond pharmaceuticals and food, spirulina-based biomanufacturing therefore has huge potential for use in high value chemicals and even potentially low carbon fuels and John Cumbers, as the team at SynBioBeta have relentlessly shown. Rapidly growing demand for alternative meats, and it would be nice to see these produced with low-carbon methods.

As the US economy continues to decarbonize, opportunities are created to replace a wide range of petroleum-derived products with bio-based alternatives. Many of these rely on inputs that may be amenable to more scalable, less carbon-intensive approaches than possible with the sterile fermentation processes used today. And doing it in a manner that causes a net positive sequestration of atmospheric carbon — a major goal of DoE’s ACCESS CARBON grant to Lumen — would be even better.

But this transition will only happen if the productivities can be improved enough — the implied cost targets for such applications are even more extreme than the Gate’s Foundation’s. The higher we reach, though, the more of these applications will fall within our grasp.

The first-generation large large photobioreactor array used to generate the complex datasets required for Google’s AI/ML team. (Photo credit: Arianna Lewellyn).

Brian Finrow is cofounder and CEO of Lumen Bioscience; Jim Roberts is Lumen’s cofounder and Chief Scientific Officer.




Jim and I started Lumen Bioscience in 2017 to develop and commercialize ultra-low-cost biologics.

Love podcasts or audiobooks? Learn on the go with our new app.

Recommended from Medium

The 5 Kinds of Problems where Deep Learning is Applicable

AI in Healthcare : Drug Discovery

Beating Alpha — Don’t Think Just Play Or Else… AI AlphaGo Wins!

On Humane Tech: Facial Recognition, the Russian-Ukraine War, and the Future of Surveillance with…

Man standing in dark room with red target lines lighting up on his face. Symbolizes facial recognition technology.

Artificial intelligence and Machine Learning: An introduction

AI.Westminster #1 (31st March, 2019)

What is Coding and Art, the Connection Between the Two?

How Machine Learning is controlling you and what you can do about it

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Brian Finrow

Brian Finrow

Jim and I started Lumen Bioscience in 2017 to develop and commercialize ultra-low-cost biologics.

More from Medium

CS371p Spring 2022: Raju Kakarlapudi

CS373 Spring 2022: John Powers

CS371p Spring 2022: Relena Lai

Cross Platform App in Python — Beeware