Protein & Molecule Foundation Models

The application of large-scale neural networks — originally developed for language modeling — to the sequence-structure-function space of proteins and small molecules. These models are trained on the universe of known protein sequences and structures, then used to predict, annotate, or generate novel proteins and molecules.

The field split into two lineages. Structure prediction models (AlphaFold2, AlphaFold3, RoseTTAFold, Chai-1) treat protein folding as a prediction problem — given sequence, output 3D coordinates. Generative models (ESM3, RFdiffusion, ProteinMPNN) treat protein design as a generation problem — given constraints (function, binding site, stability), output novel sequences that satisfy them. The 2024–2026 frontier is merging these: models that jointly reason over sequence, structure, and function in a single pass.

The most significant 2026 milestone is ESM3 from EvolutionaryScale: a 98-billion parameter multimodal generative model trained on 2.78 billion proteins that natively reasons over sequence (1D), structure (3D coordinates), and function (biological properties) simultaneously. Its key demonstration — "esmGFP," a fluorescent protein with 58% sequence identity to natural analogs but retained function — shows the model can generate genuinely novel biology, not just interpolation. (EvolutionaryScale blog)

AlphaFold3 (DeepMind, 2024) expanded structure prediction from proteins to all-atom prediction of proteins, DNA, RNA, ligands, and their complexes using a diffusion architecture. This made it possible to predict how a drug molecule docks into a protein in silico — previously requiring expensive experimental crystallography.

Key Claims

• ESM3 reasons jointly over sequence, structure, and function — Unlike predecessors, ESM3's multi-track design lets it handle design tasks no purely structure-oriented model can. Evidence: strong (EvolutionaryScale ESM3 release, IntuitionLabs comparison)
• AlphaFold3 enables all-atom structure prediction — Joints protein + ligand + DNA/RNA into a single diffusion model; critical for drug-protein interaction modeling. Evidence: strong (Nature, 2024)
• RFdiffusion enables de novo protein design — David Baker lab (UW); diffusion model conditioned on desired function generates novel protein scaffolds. Xaira Therapeutics is commercializing this lineage with $1B in funding. Evidence: strong (Science, 2023)
• Chai-1 and Chai-2 compete with AlphaFold3 — Chai Discovery's open models; Pfizer has licensed Chai's platform. Evidence: moderate (Chai Discovery press releases, 2025)
• Phase 1 success rates for AI-designed drugs ~90% vs historical ~52% — Early industry data from 2026; mechanism may be better target validation pre-IND rather than superior molecular design per se. Evidence: moderate (2 Minute Medicine)
• The ESM lineage spans ESM-1b → ESM-3 across six major variants (8M–98B params) — A 2026 peer-reviewed survey of ~100 papers maps the family: ESM-1b was trained on 250M UniParc sequences; ESMFold runs much faster than AlphaFold2 with only minor accuracy loss. Evidence: strong (ESM survey 2026)
• Protein language models still have hard, documented limits — The survey flags data bias (UniProt over-represents human/model organisms; de-novo-designed proteins are absent from pretraining), an O(n²d) attention compute bottleneck (≥16GB for 1B params), weak antibody-thermostability and out-of-domain generalization (CNNs still beat ESM on some structure tasks), and "black-box" interpretability. Evidence: strong (ESM survey 2026)
• UniGenX unifies sequence + coordinate generation under direct functional objectives — A decoder-only autoregressive transformer with a conditional diffusion head co-generates sequences and 3D coordinates across proteins, molecules, and materials. On protein induced-fit modeling it reports a 23× improvement (RMSD < 2 Å), plus new SOTA on five chemistry property targets and 436 viable crystal candidates. Evidence: strong (UniGenX, arXiv:2503.06687)
• IsoDDE roughly doubles AlphaFold-3 accuracy on the hardest protein-ligand cases — Isomorphic's drug-design engine reports 50% accuracy vs AlphaFold 3's 23.3% on the hardest protein-ligand structure-prediction cases (technical report 2026-02-10), pushing past the system that underpinned the 2024 Nobel-recognized work. Evidence: moderate (Isomorphic IsoDDE)
• Biohub's ESMC protein-biology "world model" trained on ~2.8B sequences, with lab-validated binders — Built by the former EvolutionaryScale team, ESMC covers 20,000+ human protein types; its ESM Atlas maps 6.8B sequences and 1.1B predicted structures. Designed mini-binders against five disease targets hit 36–88% (15–29% for antibody-derived) at nanomolar affinity, lab-validated. Evidence: moderate (Biohub world model)

Benchmarks & Data

• ESM3 parameters: 98B (inference); training set: 2.78B proteins
• ESM lineage (per 2026 survey): ESM-1b → ESM-3, six major variants, 8M → 98B params; ESM-1b trained on 250M sequences (UniParc)
• AlphaFold3: first model to predict protein-ligand complexes at atomic resolution
• IsoDDE: 50% vs AlphaFold-3's 23.3% accuracy on hardest protein-ligand cases (tech report 2026-02-10)
• UniGenX: 23× protein induced-fit improvement (RMSD < 2 Å); 436 crystal candidates / 11 novel compositions; SOTA on 5 chemistry property targets
• Biohub ESMC: ~2.8B sequences, 20,000+ human protein types; ESM Atlas 6.8B sequences / 1.1B predicted structures; designed binders 36–88% hit (mini), 15–29% (antibody-derived), nanomolar
• RFdiffusion: generated proteins fold within ~1Å RMSD of target in experimental validation
• Chai-1: competitive with AlphaFold3 on PoseBusters benchmark (2024)

Open Questions

• Can generative protein models design functional enzymes that survive in vivo without wet-lab optimization loops?
• What is the right benchmark for "function" — is in vitro binding sufficient or does in vivo validation set a higher bar?
• How much of the Phase 1 improvement is model quality vs better human use of AI tools for target validation?
• Will open-weight protein models (ESM3, RFdiffusion) erode moats of proprietary platforms (Schrodinger FEP+)?

Related Concepts

• Generative Biology — the design application layer above these foundation models
• In-Silico Screening — uses these models as virtual filters before wet-lab synthesis
• The Data Moat Question — who controls the training data for the next generation of models

Changelog

• 2026-06-15 — Initial compilation from web research; ESM3, AlphaFold3, RFdiffusion, Chai covered
• 2026-06-24 — Compiled new sources (unigenx-foundation-model, esm-protein-lm-survey-2026, biohub-protein-world-model, isomorphic-isodde-human-trials)