BioAI is moving from prediction to precision engineering

The BioAI conversation is moving from prediction to design. In this Latent Space episode, Chai Discovery co-founder Matt McPartlon and product lead Neil Patil describe a bet that protein models can make drug discovery look less like searching blindly through nature and more like engineering a molecule toward a target. The important caveat is that the model is not the whole system: the hardest remaining loop is still proving that a designed protein works in the lab. 1

Chai describes itself as a software and modeling layer for drug discovery rather than a pharmaceutical company developing its own drug portfolio. Its product is a computer-aided design suite for molecules, built around models that reason about biochemical structure and interaction. The company lists partnerships with Eli Lilly, Pfizer, Novartis, and argenx on its site. 2

That positioning matters. Traditional antibody discovery can resemble an enormous search problem. Researchers may use immunization or large yeast-display libraries to screen billions of possible molecules, then investigate the small number that bind to a target. A hit tells you that something sticks. It does not necessarily tell you where it binds, whether the interaction creates the desired therapeutic effect, or whether the molecule is stable and manufacturable. 3

The Chai team’s claim is that models can make the search more intentional. An antibody has a relatively stable framework and flexible binding regions. The design problem is to choose those regions so the molecule engages a particular site, or epitope, on a target protein. That precision could matter when the goal is more ambitious than blocking a target: activating a cellular switch, avoiding a similar healthy protein, or building a multispecific drug whose two ends bind different targets. 3

This is a different promise from “the model finds drugs faster.” It is a promise to make new classes of molecules possible because their geometry can be specified before the experiment begins.

The technical explanation in the conversation is easier to understand through an engineering analogy. In ordinary structure prediction, the sequence is known and the model estimates the three-dimensional shape. In inverse folding, the desired shape is known and the model searches for a sequence that can fold into it. Chai’s design approach tries to work on both at once.

The model iteratively adjusts the structure and the sequence until they become self-consistent. The result is not just a protein-shaped picture. It is a candidate sequence intended to produce that shape. Diffusion gives the system room to revise both representations during generation, rather than committing to one and treating the other as an afterthought. 3

The distinction between generation and validation is where the episode becomes more credible. A model can produce a plausible structure that is wrong, overconfident, or identical to every other answer. Chai describes several checks: compare the proposed structure with an independent prediction model, examine confidence, and look at the diversity of generated solutions. None of those is a substitute for an experiment, but together they help decide which candidates deserve scarce lab time. 3

The guests also recount a striking example from their work. In a cryo-electron microscopy validation, they said one predicted structure differed from the measured atomic positions by about 0.33 angstrom. That is a claim made in the conversation, not a reason to assume that every design works. Its significance is narrower: when prediction is close enough to experimental structure, design models have a usable foundation to build on. 3

Chai’s reported test of its approach was deliberately broad. The team set a challenge to design antibodies against 50 targets and said it obtained hits for roughly half. The result is not equivalent to 25 drugs, or even 25 therapeutic candidates. It is evidence that the system can generate experimentally interesting binders across a set of targets rather than succeeding on one carefully chosen demonstration. 3

After a candidate is generated, the work becomes slower and more physical. Lab assays may take weeks. More demanding structural measurements can take longer and require specialized equipment. The researchers describe validation as the field’s central unresolved problem: how can a team learn quickly whether a design works, and then feed that evidence back into the next round?

That is why “AI for science” cannot be judged by model benchmarks alone. In language models, an evaluation can often return in hours. In protein design, the ground truth may require making the molecule, testing its binding, checking its stability, and measuring whether it produces the intended biological effect. A model that produces attractive candidates but cannot close that loop is a generator, not yet an engineering system. 3

The host challenges the idea that designing an antibody is valuable merely because it saves a few million dollars inside a much larger drug-development budget. Chai’s answer is that the platform can unlock targets and modalities that traditional search methods struggle to reach. Precise binding can support more complex designs, including bispecifics and molecules aimed at specific cellular behaviors. The value is the new option, not only the faster version of an old workflow. 3

The partnership model reinforces that point. Chai wants to work across a pharmaceutical company’s portfolio, adapting the platform to different targets and therapeutic goals. Partner data may support specialized models or preferred design constraints, while the general platform remains reusable. That creates a hybrid business: a general software layer with significant integration and scientific collaboration around each customer.

It also explains why the company thinks in portfolios. The episode compares pharmaceutical research to venture investing: many shots on goal, limited capital, and a small number of successes that pay for the rest. Chai’s researchers describe themselves as allocating ideas and compute. The same logic applies to pharma customers, which decide which targets deserve further experimental investment. 3

The BioAI story is often told as if biology is the only hard part. The conversation makes the infrastructure constraint visible. Protein models have different memory and compute patterns from language models, yet the market for new accelerators and software is heavily shaped by LLM workloads. Chai has to buy and optimize scarce compute while building systems that can distribute a molecule-design job across many GPUs. 3

The company’s small size is part of its strategy. The guests describe a team of about 30 people, with roughly 10 focused on research. AI lets a small group allocate more of its attention to experiments and engineering, but it does not remove the need for biological judgment or lab validation. Patil’s answer to the field’s biggest bottleneck is talent: too many capable people flow toward LLMs and software, while biology remains obscure and difficult to enter. 3

That leaves a clear test for the phase shift Chai is describing. BioAI has crossed from speculation into useful engineering when models can propose precise, developable molecules, experiments can validate them fast enough to guide the next generation, and the resulting designs survive the demands of real therapeutics. The episode suggests that the first part is already happening. The next gains will come from the unglamorous middle: better assays, better data, better compute, and more people who can work across all three.