Bio Computing
The AI models rewriting biology
Thesis
Bio computing is where the leverage lives. Every advance in AI drug discovery and synthetic biology depends on the foundation models underneath — structure predictors, sequence generators, molecular dynamics simulators. The teams building these models are the picks-and-shovels play of the entire bioeconomy.
The field split in 2024. DeepMind and Meta hold the frontier with AlphaFold3 and ESM-3, but restrictive licensing is pushing the community toward open alternatives. Boltz-1, Chai-1, and Evo-2 represent a bet that open-weight models will win in biology the same way they're winning in language. The tension between closed frontier labs and open-source challengers will define the next two years. Whoever controls the model layer controls the stack.
Watch: inference cost curves. The moment structure prediction becomes commodity compute — sub-penny per protein — the entire drug discovery pipeline reshuffles. We're 12–18 months from that threshold for single-chain prediction, further out for complexes. That's the inflection point.
Last updated 2026-03-20
Featured Models
Google DeepMind
Structure Prediction, Complex Prediction, Ligand Docking
EvolutionaryScale (Meta origin)
Protein Language Model, Sequence Generation, Function Prediction
MIT
Structure Prediction, Complex Prediction
MIT
Structure Prediction, Molecular Docking, Complex Prediction
Companies
EvolutionaryScale
Building foundation models for biology.
The Meta FAIR diaspora building the GPT of biology. ESM-3 is the most capable open protein language model, and EvolutionaryScale is the commercialization vehicle. If bio foundation models matter as much as we think, this is the picks-and-shovels play.
Chai Discovery
Open-source molecular structure prediction.
The open-source challenger to AlphaFold3. Chai-1's accuracy is competitive, and the open model means the community can build on it. The business model question — how do you monetize an open model in bio — is still unanswered.
Profluent
AI-generated proteins for genome editing.
Made headlines by creating a functional CRISPR enzyme entirely from AI — OpenCRISPR-1. That's a genuine milestone, but the path from proof-of-concept enzyme to therapeutic product is long. The team knows this; the market doesn't always.
Funding
Absci
Follow-on Offering · 2025-03-08
Led by Perceptive Advisors, RA Capital
Absci sits at the intersection of generative AI and wet-lab validation in a way that most AI biotechs only aspire to. The IgDesign model generating de novo antibodies is genuinely impressive, but the real question is whether generative antibodies can match or beat the safety profiles of conventionally discovered ones. The FDA has no playbook for this yet.
Evolutionaryscale
Series A · 2024-06-25
Led by Nat Friedman & Daniel Gross, Lux Capital, Amazon
ESM3 is the GPT-3 moment for biology. The founding team literally wrote the book on protein language models at Meta, and now they have the capital to build the foundation model layer for all of life sciences. If biological language models follow the same scaling laws as LLMs, this is absurdly undervalued. That's a big "if," but the early results are stunning.
Signals
Phase I Study of ABS-101, a Generative AI-Designed Bispecific Antibody, in HER2-Low Breast Cancer
Absci Corporation · Phase I · Recruiting
The first fully de novo AI-generated antibody therapeutic to enter human trials. Every other "AI-designed" drug in the clinic was really AI-optimized from a known scaffold. This one was generated from scratch by the IgDesign model. Whether it works or fails, it will define how regulators think about generative biologics for the next decade.
Phase I/II Study of ESM-NK-201, a Protein Language Model-Designed NK Cell Engager, in Relapsed/Refractory AML
EvolutionaryScale / Novo Nordisk · Phase I/II · Recruiting
Alexander Rives, Salvatore Candido, Tom Sercu et al. · 2024-06-25
The fluorescent protein generation result is the one that matters — ESM3 designed a functional GFP variant that is as distant from natural GFPs as anything that has evolved in 500 million years. This is not incremental improvement. This is a new capability. The question is no longer whether foundation models can design proteins; it is how fast the design-build-test loop can turn.
Ali Madani, Ben Krause, Eric Greene et al. · 2024-04-22