Generative Biology

The use of generative AI — diffusion models, language models, flow-matching networks — to design novel biological molecules (proteins, antibodies, small molecules, nucleic acids) from scratch, conditioned on desired functional properties. Generative biology is the design layer above foundation models: where a protein language model encodes what proteins look like, a generative biology system uses that encoding to propose novel proteins optimized for a target property.

This is the conceptual heart of the AI-bio thesis. Classical drug discovery finds molecules among known chemical space (screening, medicinal chemistry). Generative biology expands to the vastness of undiscovered chemical and protein space — estimated to contain >10^60 possible small molecules — and uses AI to navigate it directionally rather than randomly. The key word is directed: instead of random synthesis and screen, generate candidates whose predicted structure already satisfies binding, selectivity, ADMET, and novelty constraints.

What "generative" means in practice:

1. Small-molecule generation — models (Schrodinger, Schrödinger FEP+, Insilico's Chemistry42, Genesis GEMS) propose novel chemical scaffolds conditioned on binding to a specific protein pocket
2. Antibody generation — models (Generate Biomedicines Chroma, Chai Discovery, BigHat) design antibody sequences optimized for affinity, specificity, and manufacturability
3. Protein design — models (RFdiffusion, ESM3, Profluent) design novel protein structures with no natural analog, including entirely new enzymes and gene editors
4. Multi-objective optimization — the frontier: simultaneously optimizing binding, selectivity, solubility, ADMET, novelty in a single generative pass

Key Claims

• Generate Biomedicines' Chroma model demonstrated all-atom protein generation in 2023 — First demonstration of generating novel protein structures conditioned on arbitrary function. Used internally for GB-0895 (TSLP inhibitor, Phase 3 asthma). Evidence: strong (Nature, 2023)
• Profluent's OpenCRISPR-1 is the first fully AI-generated gene editor — Protein language model designed both the nuclease and guide RNA; successfully edited the human genome; 95% off-target reduction vs SpCas9. Evidence: strong (Nature, Jul 2025; see genomics KB)
• Insilico Chemistry42 generated Rentosertib in 18 months — End-to-end AI-assisted discovery: target identification, molecule generation, synthesis prioritization. Phase IIa results in Nature Medicine Jun 2025. Evidence: strong (Insilico, Nature Medicine 2025)
• Xaira's RFdiffusion lineage enables programmable protein therapeutics — $1B in funding at launch (Apr 2024); strategy is to use RFdiffusion (Baker lab, UW) for designing protein binders, enzymes, and nanostructures as drugs. Evidence: moderate (FierceBiotech)
• Isomorphic Labs claims AlphaFold integration into drug design — DeepMind spinout uses structural prediction as foundation for generative chemistry; partnered with Lilly ($1.7B deal) and Novartis ($1.2B deal) in Jan 2024. First internal IND targeted end-2026. Evidence: moderate (IntuitionLabs)
• Generation under direct functional objectives is the new frontier, not generate-then-filter — UniGenX co-generates sequence + 3D coordinates conditioned on functional/property targets in one pass (decoder-only transformer + conditional diffusion head), demonstrating EC-conditioned enzyme design and a 23× protein induced-fit gain (RMSD < 2 Å). It operationalizes "directed" generation: the functional constraint enters the generative objective rather than being applied as a downstream screen. Evidence: strong (UniGenX, arXiv:2503.06687)
• Biohub designed lab-validated binders against five disease targets from a protein world model — The former-EvolutionaryScale team's ESMC + ESM Atlas designed mini-binders and antibody-derived binders against EGFR, PDGFRβ, PD-L1, CTLA-4, and CD45, reporting 36–88% hit rates (mini-binders) and 15–29% (antibody-derived) at nanomolar affinity — generative protein design validated in the wet lab against named oncology/immunology targets. Evidence: moderate (Biohub world model)

Open Questions

• Does generative design reduce Phase 2 attrition (the real failure point) or only Phase 1?
• Can antibody design models generalize to novel target classes outside training distribution?
• Is end-to-end generative design (target → molecule → IND without human design iteration) achievable in this decade?
• How much of Rentosertib's success is AI design vs conventional medicinal chemistry optimization post-generation?

Related Concepts

• Protein & Molecule Foundation Models — the model infrastructure underneath generative design
• Lab-in-the-Loop — generative proposals need experimental validation to close the loop
• Platform vs Asset — generative biology companies must choose whether to license the platform or own the drugs

Changelog

• 2026-06-15 — Initial compilation; covers small-molecule, antibody, and protein design sub-fields
• 2026-06-24 — Compiled new sources (unigenx-foundation-model, biohub-protein-world-model)