Biotech scaled

AlphaFold + computational protein design

AlphaFold and computational design have accelerated protein engineering from random selection to intentional computational design, enabling discovery of novel therapeutics, enzymes, and materials while achieving Nobel Prize recognition in 2024.

What to watch next

Multi-domain protein complex prediction, conformational dynamics modeling, and generalized enzyme design across diverse catalytic mechanisms reaching clinical translation within 2-3 years.

Key sub-ideas & techniques

• AlphaFold2 — single-chain folding — AlphaFold2 (2021) solved the 50-year protein-folding problem at near-experimental accuracy on CASP14, and the 200M+ structure database opened it up to every biology lab on earth. [source]
• AlphaFold3 — multimodal complexes — AlphaFold3 (2024) extended prediction beyond proteins to nucleic acids, ligands, ions, and post-translational modifications, making it directly useful for drug discovery and not just structural biology. [source]
• De novo design with RFdiffusion — Baker lab's RFdiffusion (and successors) generate brand-new proteins for a target binding site or function — flipping protein engineering from selection to direct design. [source]
• Enzyme design — AI-designed enzymes for novel reactions (luciferase, plastic degradation, sustainable chemistry) reached single-shot success rates that took years of directed evolution before — now at all 41/41 test cases for RFdiffusion2. [source]
• Protein language models — ESM-2 / ESM-3 (Meta / EvolutionaryScale) treat protein sequences like text, enabling zero-shot mutation effect prediction, sequence generation, and embedding-based search across the proteome. [source]

Current frontier

• AlphaFold 3 (published May 2024 in Nature) predicts structures of proteins, nucleic acids, ligands, ions and modified residues with substantially improved accuracy for protein-ligand and protein-nucleic acid interactions compared to specialized tools. [source]
• RFdiffusion2 (published December 2025 in Nature Methods) generates enzyme active sites from functional group geometries, designing scaffolds for all 41 test cases vs. 16 for previous methods. [source]
• D-I-TASSER (published 2026 in Nature Biotechnology) outperforms AlphaFold2 and AlphaFold3 on both single-domain and multidomain proteins through deep learning potentials combined with iterative threading. [source]
• VibeGen at MIT (announced March 2026) designs proteins by their conformational dynamics and motion, not just static structure, using agentic AI for molecular mechanics. [source]
• David Baker's lab received $7 million from Washington Research Foundation (March 2026) to advance AI-enabled enzyme design with applications in medicine, technology, and sustainability. [source]
• IsoDDE roughly doubles AlphaFold 3 accuracy on protein-ligand structures at <20% sequence identity and predicts novel cryptic binding sites from sequence alone (Isomorphic Labs). [source]
• Isomorphic Labs closed a $2.1B Series B on May 12, 2026 (Thrive Capital lead) to fund IsoDDE deployment and clinical pipeline acceleration — one of the largest private AI-drug-discovery financings on record. [source]
• Core de novo protein design challenges (novel folds, assemblies, binders) are 'close to being solved' per Baker et al. Nature review; next 5-10 years focus shifts to functional nanomachines, switches, and catalysts. [source]
• Baker Lab Nature papers (May 22, 2026) describe de novo quasisymmetric protein nanocages with expanded internal volume, opening a computationally designed delivery vehicle class for genetic medicines. [source]
• NLP-style scaling laws do not reliably transfer to protein language models; data growth on UniRef yielded non-monotonic and sometimes worse variant-effect prediction, implying effective diversity (not token count) drives biology foundation models. [BACKFILL, arXiv Jul 2025, traction: dedicated Align blog + GitHub + active debate vs ProGen3's 'biology has scaling laws' claim] [source]

Key people

• Demis Hassabis CEO of Google DeepMind · Google DeepMind; Isomorphic Labs (co-founder & part-time) [source]
• John Jumper Director of Google DeepMind · Google DeepMind [source]
• David Baker Director of Institute for Protein Design; HHMI Investigator · University of Washington; Howard Hughes Medical Institute [source]
• Tristan Bepler Co-founder and CEO · OpenProtein.AI [source]
• Joshua Meier Founder and CEO · Chai Discovery [source]
• Markus Buehler Department Head of Materials Science and Engineering; Lead on VibeGen · MIT [source]
• Alexander Rives Co-founder & Chief Scientist, EvolutionaryScale · EvolutionaryScale [source]

Startups & labs to watch

• Chai Discovery AI drug discovery startup · STARTUP · Series B: $130M (Oak HC/FT, General Catalyst); Series A: $70M (Menlo Ventures); Total ~$231M — Raised $130M Series B (December 2025) at $1.3B valuation for AI foundation models predicting molecular interactions; announced Eli Lilly collaboration (January 2026) for biologic drug discovery. [source]
• Generate Biomedicines Clinical-stage protein design company · STARTUP · Multiple rounds; pre-IPO valuation undisclosed — Filed for Nasdaq IPO (2026) with 312 employees including 138 M.D./Ph.D.s; first Phase 3 patient dosed January 2026 for GB-0895 (anti-TSLP antibody); received FDA Fast Track designation. [source]
• Iambic Therapeutics AI-driven drug discovery platform · STARTUP · $100M+ oversubscribed financing round (2026) — Announced $1.7B potential value collaboration with Takeda (February 2026) for AI drug design; NeuralPLexer model claims to outperform AlphaFold for protein-ligand prediction. [source]
• OpenProtein.AI AI protein engineering platform · STARTUP · Seed/early stage (not disclosed) — Founded by Tristan Bepler (MIT PhD '20) and Tim Lu (MIT Associate Professor PhD '07); democratizing AI-driven protein design tools for biologists. [source]
• Scribe Therapeutics Scribe Therapeutics · STARTUP · Previously backed by Andreessen Horowitz; Series B disclosed — Couples engineered CRISPR enzymes with AI guide-RNA design (DeepXE) and allosteric epigenetic editors; ASGCT 2026 data show 10-100x off-target reduction. [source]