🔬The BioAI Phase Shift - Matthew McPartlon & Neil Patil, Chai Discovery

What It Covers

Chai Discovery cofounders Matthew McPartlon and Neil Patil explain how their protein design platform uses AI to engineer antibodies and therapeutic molecules. The company has secured partnerships with Eli Lilly, Pfizer, Novartis, and argenx, raised $400M, and achieved sub-angstrom structure prediction accuracy while shifting drug discovery from a waterfall process toward an iterative engineering discipline.

Key Questions Answered

• •Antibody Design Scale: Chai's CHI-2 model demonstrated generalized antibody design by targeting 50 diverse CRO-validated proteins and achieving binders against approximately half, with an average 20% hit rate per target. This moved the field beyond single-target hill-climbing toward genuinely general design models. Pharma partners use this benchmark as proof of viability before committing to full programs, making broad target coverage the key credibility signal for enterprise adoption.
• •Neutral Platform Model: Rather than developing proprietary drug pipelines, Chai operates as a software platform serving pharma partners, which it describes as a "neutral software factory for medicines." This structure allows simultaneous partnerships with competing firms like Pfizer and Novartis. Revenue and model improvement are directly tied to partner success, creating incentive alignment that a pipeline-owning competitor cannot replicate without conflicts of interest across its own portfolio.
• •Epitope-Targeted Design vs. Brute Force Screening: Traditional antibody discovery screens billions of candidates via immunization or yeast display, yielding a handful of binders with unknown binding locations. Chai's design suite lets researchers specify exact epitopes — the precise atomic binding site — before generation. This enables selectivity engineering, cross-reactivity control across species variants like human and monkey, and agonist activity on targets like GPCRs, capabilities that random screening cannot reliably produce.
• •Photoshop-Style Molecular Design Interface: Chai built its product to resemble Autodesk or Figma rather than a chatbot. Scientists load a 3D molecule, use a paint-tool equivalent to highlight target epitopes, and trigger a content-aware-fill equivalent to generate binders. This visual paradigm reflects that molecular engineering requires spatial reasoning and constraint specification, not conversational prompting. The interface also enforces IP isolation through single-tenancy deployment, addressing pharma's core data security concern.
• •Validation Bottleneck and the Weeks-Not-Months Threshold: Cryo-EM structural validation can take months and costs too much to scale. Chai's CHI-2 paper achieved 0.33 angstrom RMSD against cryo-EM ground truth — one-third the width of an atom — validating model accuracy. Wet lab CRO networks now return binding assay results in weeks rather than years, enabling recursive model improvement. This cycle time reduction is what makes the design-test-iterate loop practically viable for commercial drug programs.
• •Compute Architecture Mismatch for Bio Models: Modern GPU infrastructure is optimized for LLM workloads with large KV caches and high-parameter attention. Structure prediction models use pair representations scaling as L-squared sequences, making attention effectively L-cubed in compute cost. Operations like layer normalization become significant bottlenecks. Chai invests in custom inference optimization and uses Temporal for durable execution across distributed GPU fleets, treating infrastructure reliability as a prerequisite for running large-scale molecular design campaigns.
• •Abstraction Ladder for Drug Discovery Products: Chai frames its product roadmap as a series of rising abstraction levels tied to model capability. Current tools resemble Cursor — scientists inspect individual molecular bonds and properties directly. As models improve, the product shifts toward orchestrating hypothesis campaigns across multiple epitope choices, then entire target pathways. Each abstraction level retires lower-level UI components, meaning product teams should build for one-to-two year lifespans rather than designing for permanence.