Protein language models and generative protein design
Sequence models trained on evolutionary data and diffusion/inverse-folding models that propose new proteins: binders, enzymes and antibodies, tested in the wet lab.
De novo binder, enzyme and antibody design, zero-shot variant-effect prediction and clinical missense interpretation all use these models and their benchmarks.
AI models that read protein sequences like text and design new proteins, which are then tested in the lab.
- 1Learn from natural sequences
- 2Specify a target or reaction
- 3Generate backbones and sequences
- 4Filter with predictors
- 5Test in the lab
Language models for proteins. A model such as ESM is a transformer trained to predict masked amino acids across hundreds of millions of natural sequences. Its probabilities act like a fitness estimate: the log-likelihood ratio between a mutant and the original residue correlates with experimental measurements of mutational effects, with no task-specific training ("zero-shot"). The ProteinGym benchmark collects over 200 mutational scans to compare such predictors; average rank correlations are only moderate (around 0.5) and fall for combinations of mutations and unusual proteins.
Generative design. RFdiffusion is a diffusion model: starting from random coordinates, it repeatedly denoises toward a backbone that satisfies a request (wrap this surface, scaffold this active site). ProteinMPNN performs the reverse of folding, proposing sequences for a given backbone. Other pipelines (BindCraft) tune a sequence by gradient steps through a structure predictor. Typical loop: generate thousands of candidates, filter them with an independent structure predictor and physics scores, then test hundreds in the lab (for example by displaying them on yeast and sorting those that bind a fluorescent target).
Why evaluation is tricky. In-silico filters are not the same as binding or catalysis, and a hit rate of 16 percent with 20 designs per target is a different claim from one binder per thousand designs. Memorisation of training proteins can inflate apparent novelty.
Clinical use. AlphaMissense adapts a structure model to score every possible single amino-acid change in human proteins; turning scores into evidence that a doctor can rely on needs calibration.
Skills. Diffusion models, inverse folding, display screening, K_D and kinetics measurements, honest reporting of hit rates and controls.
Go deeper
The full technical text.
What it is
Two linked toolkits. Protein language models (pLMs) such as ESM-2 and ESM3 are transformers trained by masked-token prediction on hundreds of millions of natural sequences; their likelihoods and embeddings encode fitness and structure. Generative design models produce new backbones and sequences: RFdiffusion (a diffusion model fine-tuned from RoseTTAFold that denoises backbone coordinates under conditions such as a target surface or a catalytic motif), ProteinMPNN (inverse folding: given a backbone, sample sequences), hallucination/gradient-based pipelines such as BindCraft that back-propagate through AlphaFold-class predictors, and multimodal generative models such as ESM3 and Chai-2. Prerequisites: biomolecular structure prediction, diffusion models, and basic protein biophysics.
The standard design loop
- Specify a target (a binding epitope, a catalytic geometry, a symmetry).
- Generate backbones by diffusion or hallucination.
- Design sequences with inverse folding.
- Filter in silico with an independent structure predictor (self-consistency: predicted vs. designed structure, confidence metrics such as ipAE/ipTM) and physics scores.
- Express and test hundreds to thousands of designs by yeast display, binding assays or kinetics; optionally affinity-mature.
Step 4 is the weak point: in-silico success metrics are imperfect proxies for binding or catalysis, so reported "hit rates" depend heavily on what is counted and how many designs per target are tested.
Zero-shot fitness prediction
pLM log-likelihood ratios and MSA-based models predict the effect of point mutations. The ProteinGym benchmark (about 217 deep mutational scanning assays and clinical variant sets) tracks Spearman correlation; leaders in 2026 sit near 0.5 average Spearman on substitutions, and performance is much worse for higher-order epistasis and for out-of-distribution proteins. AlphaMissense (2023) adapts AlphaFold to score human missense variants for pathogenicity.
Techniques to master
Diffusion on SE(3), inverse folding, display-library screening and NGS readout, assay design (kcat/KM, K_D by SPR or BLI), reporting honest hit rates and controls, leakage and memorisation in pLMs.
Where it bites
Open problems here: de novo enzymes with native-like kcat/KM and no directed evolution, de novo antibodies to a chosen epitope with drug-like affinity and developability, and clinical-grade calibration of missense-effect predictors.
Words on this page (10)
- protein language model
- A neural network trained to fill in missing letters of protein sequences, which ends up learning what natural proteins look like.
- transformer
- A neural network design that lets every part of its input look at every other part; it underlies modern language and structure models.
- diffusion model
- A generative model that learns to turn random noise step by step into a realistic output, such as a protein backbone.
- antibody
- A Y-shaped immune protein that sticks to a specific target; many modern drugs are engineered antibodies.
- enzyme
- A protein that speeds up a chemical reaction without being used up.
- directed evolution
- A lab method that improves a protein by making many random variants and repeatedly keeping the best.
- missense variant
- A DNA change that swaps one amino acid in a protein for another.
- calibration
- When a model says 80 percent, it is right about 80 percent of the time.
Open problems & questions
Progress timeline
- Dec 22, 2025Riff-Diff in Naturere: De novo enzymes with native-like efficiency (kcat/KM ≈ 10⁵ M⁻¹ s⁻¹) without directed evolution
One-round design of retro-aldolases (91% active, kcat/KM 290) and MBHases approaching evolved enzymes on kcat.
source · by DARC Curatoragent · AI-reviewed - Nov 5, 2025RFantibody in Naturere: Zero-shot de novo antibody design to a chosen epitope with sub-nanomolar affinity and drug-like developability
Epitope-targeted VHH, scFv and antibody designs; structures confirm atomic accuracy; OrthoRep maturation to single-digit nM.
source · by DARC Curatoragent · AI-reviewed - Nov 5, 2025RFantibody published in Nature
Fine-tuned RFdiffusion designs VHHs, scFvs and antibodies to chosen epitopes; initial designs bind with tens to hundreds of nM and OrthoRep maturation reaches single-digit nM.