ML that designs biology — the field, the employers, and where a systems engineer fits · verified June 2026
| Modality | What it designs | Flagship models (2025–26) | Real now vs hype |
|---|---|---|---|
| Protein design | Novel backbones + sequences, binders, mini-proteins | RFdiffusion2/3, AlphaProteo, ESM-3, Chai-2, BindCraft | Binders & mini-proteins real; general function-conditioned design still weak |
| Genome / sequence LMs | DNA, variant effects, regulatory elements | Evo 2 (Arc, 40B params, 1M-bp context, Nature 2025) | Variant-effect prediction real; viable designed genomes still aspirational |
| Virtual cell | Cell-state & perturbation/drug response | Arc State, scGPT, CZI Virtual Cells Platform | Works on seen cell types; causal generalization unsolved |
| Generative chemistry | 3D small molecules fit to a binding pocket | Pocket-conditioned diffusion (DrugDiff, fragment-based) | Diverse candidates; synthesizability & scoring are the bottleneck |
| Antibody / enzyme | CDRs, nanobodies, catalytic active sites | RFdiffusion antibodies, Chai-2, RFdiffusion2 enzymes | Zero-shot binders hit in a 24-well plate; de novo antibodies still hard |
The common shape: a generative model proposes a sequence or structure, a predictor scores it, and a wet-lab round returns the truth. The model is the smallest part.
De novo protein binders. The wet-lab data loop — not bigger models — lifted hit rates from under 1% to tens of percent. BindCraft (Nature 2025) averages ~46% (range 10–100%) at nanomolar affinity; AlphaProteo reports 9–88% across targets.
Antibodies / nanobodies. Chai-2 does zero-shot de novo antibody design at a 16–20% hit rate across 52 antigens with ~20 candidates per target — over 100× prior methods (which were under 0.1%). This is the year's headline jump.
Genome generation. Arc's Evo 2 generates genome-scale DNA and does zero-shot variant-effect prediction across all domains of life.
Still research: reliable novel enzyme catalysis; virtual-cell perturbation prediction on unseen cells/genes (the Virtual Cell Challenge drew 1,200+ teams — the signal that it is not solved).
Where the money goes: binders and antibodies (clear clinical payoff, fast loops) attract capital today; the virtual cell is the next frontier bet (Arc, NVIDIA, 10x).
⚙ = most engineering/compute-bound — these hire software/infra engineers, not only ML scientists. EU/UK flagged for the non-US path.
| Company | Location | Generates | Stage / funding | Eng-bound |
|---|---|---|---|---|
| EvolutionaryScale (ESM) | NYC | Protein LLMs (ESM3) | $142M seed (Amazon/Nvidia) | ⚙ frontier-model infra |
| Chai Discovery | SF | Antibody/structure (Chai-2) | $130M Series B, $1.3B val (OpenAI-backed) | ⚙ |
| Cradle Bio | Amsterdam + Zürich — EU | Protein-design SaaS | ~$102M | ⚙ software-first |
| Latent Labs | London + SF — EU | De novo proteins | $50M (DeepMind/AlphaFold alumni) | ⚙ |
| Isomorphic Labs | London — EU | Generative drug design | $600M (Alphabet) | ⚙ ML research eng |
| Basecamp Research | London — EU | Genomic data + gene-insertion models | ~$85M (Nvidia) | data/infra |
| Arc Institute (Evo) | Palo Alto | Genome/DNA-RNA-protein (Evo 2) | Nonprofit, Nvidia compute | ⚙ compute-bound |
| Recursion (RXRX) | Salt Lake City (public) | Phenomics / cell models | BioHive supercomputer | ⚙ large SWE org |
| Generate Biomedicines | Somerville MA | Proteins / antibody therapeutics | $1B+ raised | SWE/MLOps |
| Profluent | Berkeley CA | Proteins, gene editors (recombinases) | $150M + $2.2B Lilly pact | bioinfra |
| Dyno Therapeutics | Watertown MA | AAV gene-therapy capsids | Series A | mostly ML-science |
| Markov Bio | SF | Virtual cell (aging/longevity) | Early-stage, thin public data | ⚙ "limited by engineering & compute" |
EU/UK without relocating to the US: Cradle (NL/CH), Latent Labs (London), Basecamp (London), Isomorphic (London). See also the Opportunity Catalogue (European hubs folded in) and the Markov Bio profile.
The bottleneck is rarely the model — it is the loop around it. GPU memory and IO bandwidth, not algorithms, cap training and serving; only a handful of orgs control the data and GPU scale to build a foundation model. That is plumbing, scheduling, and throughput — your work, not a biologist's.
FASTQ → align → BAM → count matrix → h5ad/AnnData. Training on the 100M-cell Tahoe-100M dataset already forces custom streaming loaders (e.g. annbatch) — IO, sharding, zero-copy. Reproducible Nextflow-style DAGs are the substrate.The build path is its own page now — ranked, standalone artifacts a single engineer can ship (no agent-orchestration glue). See What To Build To Break In: a Rust QUBO/Ising solver leads, with a float compressor, a sparse direct solver, and bgzf-turbo behind it, plus the noodles contribution as the on-ramp.