Programming E. coli to outsmart disease
Jun 27, 2023
Synthetic biology pioneer Tim Lu on machine learning in biology, the superpower behind Absci’s E. coli platform, and this moment in pharma.
When Tim Lu got his computing degree from MIT, his next question was: Where can I make the biggest impact?
The answer wasn’t very far. In the Boston area, many researchers were working in a new field called synthetic biology. Tim was among the early pioneers in the field to see the potential in programming biological circuits to fight disease, for which he received the US President Early Career Award for Scientists and Engineers and several other awards.
“The reason I got into the field was to understand how sequence translates to function — not just because I like to tinker with stuff,” he says.
Earlier this year, Tim was appointed to Absci’s Scientific Advisory Board, where his expertise will help further optimize Absci’s synthetic biology platform for producing antibodies at speed and scale.
In addition to being a professor at MIT, Tim is also co-founder and CEO of Senti Bio, which uses its sophisticated synthetic biology platform to program next-generation cell and gene therapies to “outsmart” complex diseases. As a computationalist, synthetic biologist, and founder, Tim has a unique perspective on this moment in biotech.
We caught up with him during SynBioBeta 2023, where pharma and AI were two topics everyone seemed to be talking about.
Machine learning’s impact on synthetic biology
Before AI, Tim says that the first 10-20 years of synthetic biology relied on using bottom-up models to simulate molecular and cellular processes.
“That proved to be pretty challenging because there’s too much compute power required to accurately simulate all of the intricate details involved in biomolecular interactions,” he says. “In the last five years, machine learning has come around, which can provide a roadmap for how to design biological systems without having to simulate every single detail,” he says, “and that’s been super exciting.“
“It’s actually a little bit intellectually unsatisfying,” Tim continues, “because it doesn’t necessarily teach you why biology does things a certain way.”
Tim is referring to the “black box” problem in the field of artificial intelligence and machine learning: deep learning models are not good at providing a clear, understandable explanation of how they got from input to output. Tim points out that the situation isn’t much different from how things are now: Human designers create a screen, select the best candidates, and characterize them, but they can’t really explain every piece about how or why it works.
“Machine learning elevates the scale and speed by which you can sort through potential designs before you commit to a lead candidate and everything downstream of that,” he says. “The kind of experimental work you need to do to get a product from discovery to clinic is much the same.”
Making E. coli great again
So how does Tim think biotech can best take advantage of this new scale to create better biologics for patients faster? For starters, by leveraging E. coli for its engineering strengths (rapid growth, ease of editing, high production, established protocols).
“Absci’s E. coli platform is pretty unique,” says Tim. “When people think about producing antibodies in E. coli, the question immediately comes up is how it compares to making them in CHO cells.”
That’s because antibodies are complex proteins that require special cellular machinery in the host cell to correctly fold and assemble them. Normally, E. coli lacks the post-translational apparatus to produce functional antibodies, forcing manufacturers to make antibodies in CHO cells, which are slower, more finicky, and lower throughput, in comparison.
Absci’s roots are in cell line development, and a decade of work produced SoluPro, its E. coli platform with a semi-oxidized cytoplasm that enables the production of complex, disulfide-bonded proteins while maintaining a healthy cell population at commercial scales. SoluPro offers advantages over CHO mammalian systems by achieving dramatically shorter production cycles (1-2 days versus 20-24 days) and higher titers, making it suitable for producing complex mammalian proteins such as IgG antibodies.
“Ten years ago, if you said you’d be making antibodies in E. coli and using them in clinical applications, most people would think you were crazy. The fact that Absci is tackling this problem is pretty interesting to me.”
It’s all about the data
The ability to generate massive quantities of antibodies in E. coli isn’t so they can go directly into humans — as least not yet. Rather, those antibodies are a rich source of training data for AI drug design. In fact, this was one of the recurring themes of SynBioBeta: data is more important than algorithms.
“For any therapeutic antibody, you want to characterize affinity, you want to characterize specificity. But you’d also like to know about manufacturability, immunogenicity, and how it works in specific functional assays,” Tim says. “I think those are actually going to be the ultimate bottlenecks for the field.”
Tim thinks we can overcome those bottlenecks by combining the data to train, the AI to create, and the wet lab to validate. Today, Absci has the ability to screen billions of cells and millions of AI-designed antibodies every week. In the future, Tim expects we will increase these numbers.
From E. coli to humans
Tim says that Absci has already demonstrated the ability to generate gigantic datasets using affinity selection screens, for example. For him, the next big breakthrough in our space will be in creating human disease models to test large libraries against.
“How about throwing that gigantic library into an animal or humanized disease model and then selecting for a biologic that effectively treats the disease?” he says. “That’s really hard still because we don’t have good models for a lot of illnesses. But that’s the type of data we ideally want.”
Today, our best models are a kind of surrogate for how we think drug candidates will eventually work in the clinic. The ideal experiment would be to create a massive library and put it into humans to make sure you identify the molecules that work.
“Obviously, you can’t do that experiment,” he says, “but if we can design model systems to mimic diseases more accurately and be amenable to large-scale studies, I think that’s really where we’re going to take full advantage of machine learning and massively parallel protein engineering.”
A new era, a sunny outlook
Though biotech and pharma have been dogged by a bearish market the last couple of years, the outlook at SynBioBeta was optimistic. Tim shares that outlook.
“Biotech has been through cycles like this in the past,” he says, with swings between exuberance and austerity. “So unless you’re really pessimistic about the need for better therapies, you have to believe that the field will come back.”
Perhaps more importantly, Tim thinks AI and synbio are at an intrinsically new stage of growth and maturity. “What I’m excited about is combining new machine learning methods to help design drugs that effectively treat disease biology and ultimately allow us to uncover how the body works,” he says. “Integrating machine learning with synthetic biology will enable a fundamental shift in how drug discovery was done in the past.”
“We’ve got to focus on the persistent illnesses ahead of us,” he says. “Diseases aren’t going away. We have to continue to find ways of solving problems and helping patients.”