EvolutionaryScale

Type: Private protein foundation model company — backed by a16z, Lux Capital, NVIDIA, NVentures

EvolutionaryScale built ESM3, the most powerful open-weight protein language model as of 2025–2026. ESM3 is a 98-billion parameter multimodal generative model trained on 2.78 billion proteins that natively reasons over:

• Sequence — amino acid sequence (1D)
• Structure — 3D atomic coordinates
• Function — biological annotations and activity

The model is genuinely generative: given functional constraints, it can design novel protein sequences that fold into target structures and exhibit desired activities. The "esmGFP" demonstration — a fluorescent protein with 58% sequence identity to natural analogs but retained fluorescence — showed the model can create biology with no natural ancestor.

Strategic significance: ESM3 is open-weight. This has two effects: (1) it democratizes protein design research, enabling any lab with GPUs to run serious protein experiments; (2) it erodes the moat of commercial platforms that rely on proprietary base models. Companies building applications on top of ESM3 (or fine-tuning it for specific tasks) inherit its capability without needing to reproduce the $50–100M training cost.

Key numbers

• Parameters: 98B (full; smaller variants available)
• Training data: 2.78 billion proteins
• Investors: a16z, Lux Capital, NVIDIA (NVentures)
• Funding: ~$142M

Key Claims

• The ESM lineage now spans ESM-1b → ESM-3 across six major variants (8M–98B params) — A 2026 peer-reviewed survey of ~100 papers maps the family and its downstream uses (structure/function/interaction prediction, variant-effect, directed evolution, de novo + conditional design); ESM-1b was trained on 250M UniParc sequences. Evidence: strong (ESM survey 2026)
• ESM's documented limits bound the open-model thesis — The survey identifies UniProt data bias (de-novo-designed proteins absent from pretraining), an O(n²d) attention/compute bottleneck, weak antibody thermostability + out-of-domain generalization, and black-box interpretability — a sober counterweight to ESM3's capability narrative. Evidence: strong (ESM survey 2026)
• The former EvolutionaryScale team shipped a protein-biology "world model" at Biohub (May 2026) — Led by Alex Rives (ex-EvolutionaryScale chief scientist, now Biohub Head of Science), the team trained ESMC on ~2.8B sequences, built the ESM Atlas (6.8B sequences, 1.1B predicted structures), and designed lab-validated binders against five disease targets (36–88% hit for mini-binders, nanomolar). Note: this is the Biohub release, built by alumni — a talent-and-lineage link, not an EvolutionaryScale product. Evidence: moderate (Biohub world model)

Sources

• EvolutionaryScale ESM3 release blog
• BiopharmatrEnd ESM3 overview
• IntuitionLabs biology foundation models comparison
• Survey of ESM applications (2026) — Yang, Yu & Zheng, ShanghaiTech, Quantitative Biology
• Biohub protein-biology world model (2026-05-27) — GEN, on the former-EvolutionaryScale team

Changelog

• 2026-06-24 — Compiled new sources (esm-protein-lm-survey-2026, biohub-protein-world-model)