Protein AI stops predicting structures and starts designing them

What changed

The four preprints posted this week share a shift in emphasis: from predicting a protein's shape to deciding what it does, which antibody is worth making, and which molecule to design. A language model called PALM annotates protein function and scores variant effects; a method called FoldAbS repurposes frozen folding models to screen antibodies; a framework called PartiDesign designs proteins, nucleic acids and small molecules together; and an experimental map of the cancer protein Ras shows what the design tools are aiming at. m:…6047 m:…3910 m:…2163 m:…2804

PALM, described in a preprint posted this week, was trained on 15 million InterPro-annotated sequences and scored domain and functional-site F1s of 0.941 and 0.877 on 2.3 million held-out proteins. On ClinVar missense variants it reached an AUROC of 0.835, against 0.710 for ESM2, and 0.877 when combined with ESM2 likelihood-ratio signals. Its embeddings also beat AlphaMissense on a benchmark of 128,161 literature-derived missense variants, and the authors present the embeddings as a stand-in for deep mutational scanning, a way to read variant effects without running the experiment for every variant. m:…6047

Screening and design, in one pass

FoldAbS takes a different shortcut: instead of training a new model, it freezes a protein folding model and uses it as an encoder for antibody screening. In a preprint posted this week, the method, built on the Protenix folding model, improved specificity by 13.1%, affinity by 9.2% and developability by 6.2% over baseline methods, and it scores all three in a single model rather than in separate pipeline stages. The point is practical: antibody discovery currently judges candidates on separate criteria in separate steps, and a unified screen changes which candidates survive. m:…3910

PartiDesign goes further toward generation. The framework designs structure and sequence together across proteins, nucleic acids and small molecules, with fine-grained control over secondary structure, chemical bonding and amino-acid preferences. In the preprint, it achieved the highest estimated successful-design throughput among the methods evaluated, and its configurable constraints cover specialised targets such as cyclic peptides and antibodies. The authors describe it as extending the partially latent framework of La-Proteina. m:…2163

The experimental check

The Ras study is the experimental anchor for all of this. Researchers mapped how the small GTPase Ras binds its effector Nore1A by screening about 33,000 Nore1A variants with yeast surface display and deep mutational scanning. They found that GTP hydrolysis barely remodels the binding landscape, while the cancer-linked RasG12D mutation reshapes it substantially, and that some Nore1A variants bind the mutant more strongly while staying selective against the normal protein. The authors trace that to the mutant's greater intrinsic flexibility, and present the map as a starting point for designing inhibitors that hit only the cancer variant. m:…2804

None of this is proven where it counts yet. The Ras paper's engineered inhibitors have not been validated in cells, and the map covers only the Nore1A effector. FoldAbS has no clinical validation, and its performance on targets other than antibodies is uncharacterised. The PALM preprint makes no mention of releasing public code or model weights. These are preprints, and the gains are measured on benchmarks and in screening assays, not in patients or in production pipelines. m:…2804 m:…3910 m:…6047

What it means for builders

The implication for anyone building with protein models is a change in where the work sits. Structure prediction is now a component, frozen, reused and built on, rather than the product. The product is the judgement layered on top: which variant is pathogenic, which antibody is developable, which designed molecule is worth synthesising. The papers this week each claim a piece of that judgement, and each backs the claim with a number. What none of them yet shows is the piece that matters most: a designed or screened candidate surviving contact with a real biological system. That is the test to watch for next. m:…6047 m:…3910 m:…2163 m:…2804