Breaking the Cognition Barrier: How Machine Learning Can Improve Immunotherapy

My first job was as a bioinformatician. I did a PhD in biomedical informatics, which they shortly thereafter rebranded to biomedical data science. I've worked generally in this antibody–ML space for most of my professional life, having worked at two other biotech startups before joining LabGenius: one on the diagnostic side and one on the therapeutic side. I've always been intrigued by the complexity of the immune system and the ways in which computers and models might help us understand and take advantage of it. Our work developing multispecific antibodies sits right at the cross-section of those interests.

It's automated throughout. We're cloning, expressing, and purifying up to 3,000 TCEs at a time, and automation is essential—we've invested in state-of-the-art automation infrastructure. That automation flows through to both the developability and functional assay pipelines. For a bispecific cycle, those 3,000 TCEs are run through three different cell lines to measure selectivity. That generates a mountain of data. We've built software infrastructure for streamlined ingestion and processing of that data, which feeds back into the ML models. Those models then update and generate the next round of designs, closing the loop between the automated experimental infrastructure and the computational layer.

The next round of designs is ultimately proposed by the ML models. One of the key human decision points comes after the data has been synthesized by our models—we can get a snapshot of how well the project is progressing toward the molecular product profile and decide: are we ready to move to the next stage, or do we need more cycles?

[Another decision is] whether we are still exploring the feasible design space and skewing toward higher-risk designs, or have we accumulated enough data to exploit what we've learned and skew toward high-reward designs? Both are human-informed decisions.

At a high level, the EVA platform is about integrating the wet lab and the dry lab to run cycle-based optimization of molecules. We're going after really challenging problems in biology—TCEs that kill all tumor cells and don't touch healthy cells. On-target, off-tumor toxicity is one of the big challenges in the field, and it's something that's stymying others as they go toward the clinic.

By integrating the wet and the dry lab—coupling sophisticated ML models with a really high-throughput lab system—we've been able to engineer molecules that seem almost impossible in terms of having that profile. We've been able to do that because of the breadth of data we gather from the lab and feed into our ML models to optimize further and identify these “diamond-in-the-rough” molecules we're ultimately after.

The cognition barrier comes down to this: if we were to try to design these molecules as humans, we would struggle [with the sheer number of variables involved]. We need molecules that are potent, selective, and developable—they must be produced at high yields, be pure, thermostable, and able to survive [manufacturing] processes. One current program involves optimizing over a space of two million [permutations of antibodies]. That's simply more than the human mind can feasibly grasp.

We're talking about the combination of all the different components that go into the TCE. We have a panel of [TCE] domains varying in [binding target] and affinity. We have a panel of [tumor recognition sites], with affinities ranging from single-digit picomolar to triple-digit nanomolar, covering all available epitopes. We vary the valency of the [tumor binding] arm—monovalent, bivalent, trivalent, and we vary linker structures in terms of lengths and rigidities. All of these are levers we're pulling simultaneously, and we see really surprising results. Very small changes that you would not think have a significant impact can completely change the [efficacy and safety] of a molecule.

When we look at other groups using pH-selectivity models, we see windows of selectivity—maybe 100-fold, maybe 1,000-fold—but with some residual [off-tumor] activity at both tested pH values. What we see with our avidity-driven molecules is complete on/off selectivity: single-digit picomolar killing against tumor cells and no detectable killing of healthy cells at any concentration tested, up to several hundred nanomolar. 

We built our initial infrastructure around avidity-driven bispecific TCEs, so we did have to generalize our capabilities to accommodate broader classes of molecules. There were challenges to solve, but we've solved them. Now we can work with any multispecific antibody format, which is really powerful.

What's genuinely interesting is the different therapeutic hypotheses you can test with these formats. With avidity-driven selectivity and a single target, you're focused on addressing the known limitations of that particular antigen. With trispecifics, you start to look at how targets interact with each other. It gets really interesting.