The optimistic case for protein foundation model companies (www.owlposting.com)

Venture funding for protein foundation model startups ballooned in 2024–25 (e.g., EvolutionaryScale $142M, Chai Discovery $70M, Xaira $1B), provoking skepticism that these models are commoditized and indistinguishable from strong open-source alternatives. The essay pushes back: yes, binder design—backed by big public resources like the PDB—is largely solved by many models, but private teams can still pull ahead by paying to assemble much larger, multi-objective datasets. Early private results (e.g., Chai-2 vs. RFDiffusion) hint at substantial performance bumps in specific tasks, and even modest percentage gains can compound when optimizing downstream workflows. The core optimistic case is that truly valuable models will jointly optimize multiple biochemical properties—not just binding but expression, stability, solubility, immunogenicity, receptor promiscuity, manufacturability and PK/PD—something open datasets currently lack (non-binding datasets often <1,000 antibodies). If a foundation model can reliably solve multi-property design, it could automate large portions of preclinical chemistry, slashing costs and timelines that today run from hundreds of thousands to tens of millions per compound. That potential economic and workflow disruption explains why deep-pocketed investors might rationally bet on differentiated, private protein foundation models: only organizations that can afford to curate the right datasets can plausibly unlock this next tier of capability.