No Priors Ep. 121 | With Chai Discovery Co-Founders Jack Dent and Joshua Meier
Chai Discovery's co-founders discuss Chai 2, their new generative AI platform for antibody design. It achieves a nearly 20% hit rate from just 20 computational attempts, a 100-fold improvement over previous methods, signaling a shift from drug discovery to drug engineering.
The convergence of several key technological breakthroughs created the opportune moment to found Chai Discovery. The first was the maturation of protein structure prediction, moving beyond single proteins (like AlphaFold 2) toward predicting the interactions between multiple molecules—a necessity for drug discovery. The second was the emergence of advanced generative models, like diffusion and language models, which suggested a path to generating a diversity of protein conformations, not just a single static view. This pointed toward the possibility of creating a generalizable, portable AI platform that could treat biology as an engineering discipline, moving beyond the traditional, tightly-integrated lab-in-the-loop model of most AI-bio companies.
Chai 2: From Discovery to Engineering
Chai 2 is a generative model that designs antibodies for specific targets. The platform’s breakthrough is its ability to design novel, functional antibodies with a high success rate. In a benchmark across over 50 diverse targets, Chai 2 achieved a nearly 20% hit rate from just 20 computational designs per target.
This represents a paradigm shift in efficiency:
- Previous Computational Methods: Achieved success rates around 0.1% or lower.
- Traditional Lab Screening: Requires screening millions or billions of compounds in massive yeast or phage libraries to find a single binding molecule.
This leap in performance transforms drug discovery from a high-cost, speculative “search problem” into a more deterministic “engineering problem.” The immediate impact is the ability to accelerate programs, but the more profound implication is unlocking entire classes of targets that were previously inaccessible with traditional methods.
Under the Hood: A Generative Model for Molecules
The model’s success is built upon a state-of-the-art structure prediction engine, which acts as an “atomic-level microscope.” This foundation allows the design model to function as a generative tool, akin to “Midjourney for molecules.”
You can almost think about structure prediction as the ImageNet moment for the field… and then design where you’re trying to design binders that is much more like a generative task.
The model is prompted with a target and then places individual atoms in 3D space to generate a complementary antibody sequence and structure. The precision is often less than the width of a single atom.
A key finding is the model’s powerful generalization capabilities. It performs equally well on targets that are biologically very dissimilar (e.g., down to 25% sequence identity) from its training data. This suggests the model has learned fundamental, underlying principles of molecular interaction that were not previously articulated, moving beyond simple in-domain interpolation.
A New Paradigm for Drug Discovery
The development of Chai 2 signals a platform shift for the biotech industry. The goal is to create a “computer-aided design” (CAD) suite for biology, similar to what SolidWorks is for mechanical engineering.
This will change the role of wet labs. Large-scale screening doesn’t disappear but evolves. Instead of a brute-force search, it becomes a validation and data-generation step. For example, one could test thousands of AI-generated candidates—far more than the 20 used in the paper—to find even better molecules or provide feedback to refine the next generation of models. This marries the best of AI with the best of biology.
The platform is already demonstrating its ability to solve complex, multi-constraint problems that are difficult for traditional methods. In one instance, Chai 2 successfully designed an antibody that binds to both the human and monkey (cynomolgus) versions of a protein—a common bottleneck in pre-clinical development—succeeding with only 14 designs where a partner had struggled for years.
Building an Engineering-Centric Bio Platform
Chai’s strategy emphasizes building a product, not just a model. As capabilities advance, specifying design goals becomes more complex. A robust product and software layer is needed to manage prompts for multi-property optimization, such as:
- Multi-target binding: Designing an antibody to hit two different epitopes on a target (biparatopic).
- Cross-species reactivity: Ensuring a drug works in both humans and animal models.
- Selectivity: Designing a molecule that hits one target while actively avoiding another.
The team’s philosophy is rooted in engineering rigor, which is crucial in deep learning where bugs in training runs can be catastrophic and expensive. By applying principles like unit testing, modularity, and meticulous architectural stewardship, the team aims to build a compounding advantage and maintain high velocity.
The Future for Biologists: The Rise of the Prompt Engineer
For antibody engineers and biologists, the future is not one of obsolescence but of augmentation. The most effective professionals will be those who master these new tools. Their role will shift from painstaking lab work to becoming expert “prompt engineers,” using their deep domain knowledge to creatively specify design challenges for the AI. By using the models to handle the search space, scientists can focus on higher-level problems, dreaming up and executing on novel therapeutic possibilities that were once considered science fiction.