Chai Discovery was founded by experienced leaders in AI and bioengineering, formerly from Meta, OpenAI, and Stripe.
The company was started after founders saw imminent breakthroughs in protein structure prediction and generative modeling, believing the timing was right to commercialize.
Unlike traditional AI-bio companies with tight lab integrations, Chai aimed to build a portable, generalizable AI platform for broad application across various drug discovery projects.
Motivation and Opportunity in Drug Discovery 01:06
Previous computational approaches in drug discovery showed promising research but lacked practical company timelines.
Key inflection point came from advances like AlphaFold 2 in protein folding, which moved from single-protein modeling toward interaction modeling (e.g., antibodies and antigens).
Chai's approach leverages diffusion and language models to increase the diversity and quality of generated molecular designs.
Chai 1 was open-sourced and widely adopted in industry for structure prediction, setting the stage for Chai 2's more advanced capabilities.
Why Focus on Antibody Discovery and Platform Shift 04:03
The potential to engineer molecules with atomic precision presented a massive platform shift affecting not only health but a variety of industries.
Antibodies are among the most therapeutically valuable molecules, comprising about 50% of recent drug approvals and 7 of the top 10 bestselling drugs.
Chai 2 was launched as a generative model platform to design antibodies, achieving much higher hit rates than previous computational or lab methods.
Technical Breakthrough: Chai 2 Results and Methods 06:12
Chai 2 can design effective antibodies against specified targets with a ~20% success rate in just 20 attempts per target, versus previous computational methods (~0.1%) and lab-based approaches (often requiring screening millions of molecules).
The benchmark involved 52 problems to demonstrate generalizability, as opposed to typical studies with only a few targets.
Model assessment was engineered-centric: used vendor catalogues, ensured held-out data, and focused purely on model performance, not therapeutic relevance.
How the Model Works: Intuitive and Technical Details 12:59
At core, structure prediction provides an atomic-resolution “microscope” for 3D protein configurations.
The design model receives prompts specifying targets and generates both sequence and 3D structures for antibodies likely to bind the target.
Analogy: Structure prediction is like classifying images; generative design is like creating new, fitting “keys” for molecular “locks.”
Success rates held even when generalizing to very dissimilar targets, indicating the model learns fundamental interaction principles.
Despite a gloomy biotech market, Chai 2’s breakthroughs are seen as a sign of an impending industry platform shift.
Success rates could rise from 20% toward 50–100% across various molecule classes as models improve.
Envisions a future “CAD suite” for biology (analogous to software for engineering or design), fundamentally altering molecule design and drug discovery.
This shift may lower risk and diversify targets in drug programs, increasing efficiency and effectiveness across the industry.
Expansion Beyond Antibodies and Further Optimization 27:00
Beyond hit rates, other properties like manufacturability and stability will be optimized using the platform.
Next-generation antibody formats (e.g., biparatopic antibodies binding multiple epitopes) become easier and faster to create.
Approach supports more complex engineering needs, from dual-species cross-reactive molecules to optimizing selectivity and multi-target prompts.
Product Direction, Defensibility, and Investment Areas 31:39
Chai 2 is evolving from a research model to a full software product, requiring significant investment in user interface, workflow support, and prompt engineering.
Complexities in prompt specification and supporting multiple drug modalities mean the product must enable sophisticated yet accessible workflows for scientists.
Continuous learning from lab feedback will help models function as co-pilots, enhancing experimental iteration.
Skills for the Next Generation of Biotech Experts 36:23
Antibody engineers and biologists are advised to gain access to Chai 2, master prompt engineering, and explore new experimental possibilities.
The shift will require both technical and creative adaptation, similar to how LLMs changed workflows in natural language tasks.
Collaboration with domain experts will be crucial due to the vastness and complexity of biological knowledge.
Company Building: Engineering Culture and Team Structure 40:04
Chai maintains a highly interdisciplinary, engineering-focused team drawing from biology, chemistry, physics, AI, and computer science.
Strong emphasis on modularity, simplicity, and software best practices to avoid technical debt, which is especially important due to the high cost of deep learning errors.
Despite a small team (~12 members), intense focus and collaboration drive the company’s rapid progress.
Investment in platform engineering from the outset is seen as essential for scaling and speed of innovation.
The company began with scrappy resource use (leveraging free compute credits) and remains disciplined about compute investment, scaling only when justified by results.
The focus is on rapid progress, prioritizing key problems, and being smart about spending amid fast-moving advances in the field.
Chai is hiring across all key functions, including engineering, antibody discovery, business development, and product roles to meet surging industry demand.
The co-founders express optimism about the impact of their work and encourage interested parties to follow their updates and apply for early access or open roles.
Listeners are directed to No Priors’ podcast and website for more information and transcripts.