Protein AI's leaderboards stopped agreeing. The builders noticed

What changed

A benchmarking platform called Protify, posted as a preprint this week, delivers an uncomfortable result for anyone who picks protein language models by leaderboard. No single protein language model dominates across 32 diverse functional tasks: 14 different models took the top score on at least one dataset. The platform benchmarks across 35 models and 61 datasets, and its authors show that model rankings are highly sensitive to evaluation design, which means the winner depends on how you set up the test as much as on the model. m:…9759

The sharpest finding concerns homology, the family resemblance between proteins in training and test sets. Under a homology-controlled split, pooled-embedding probes collapse to near-zero R2, while only transformer probes over full residue embeddings recover signal. In plain terms, a chunk of what looked like model skill was the test set resembling the training set. Take that resemblance away and the simple methods fall apart. m:…9759

That does not mean the models are useless. It means the field has been measuring the wrong thing, and Protify's contribution is to make the measurement honest. DPLM3B is currently the top-performing self-supervised encoder on ProteinGym, the platform reports, which is a real result, but the broader lesson is that a single top score no longer settles an argument about which model to use. m:…9759

The new tools are built to be reused

The same week brought three tools that fit the new mood: less leaderboard-chasing, more reuse. PALM, a protein annotation language model, achieves domain- and site-occurrence F1 scores of 0.941 and 0.877 on 2.3 million held-out proteins. Its embedding disruption distinguishes ClinVar missense variants with an AUROC of 0.835, outperforming ESM2 at 0.710, and combining PALM embedding disruption with ESM2 likelihood-ratio signals reaches an AUROC of 0.877. The authors present PALM-derived embeddings as a proxy for deep mutational scanning, which would let teams assess variant effects without running the experiment. m:…6047

FoldAbS takes a different reuse route for antibody screening. It repurposes frozen protein folding models as foundation encoders, capturing interaction dynamics beyond basic structure prediction, and evaluates specificity, affinity and developability inside a single model architecture. FoldAbS-Px achieves performance gains of 13.1 percent in specificity, 9.2 percent in affinity and 6.2 percent in developability compared with baseline methods. The idea is to stop building a separate model for each property and screen candidates once, on all three. m:…3910

RPI-Zero attacks the same reuse problem for RNA-protein interactions. It uses a sequence-only architecture that describes every protein as a mixture of shared motif prototypes, enabling zero-shot prediction for unprofiled molecules, and it achieves the highest score in eight different metrics compared with five state-of-the-art methods on unseen RNAs, unseen proteins and unseen pairs. Its benchmark dataset contains 2.1 million examples derived from 11.6 million CLIP peaks across 209 proteins, which is the kind of scale that makes a zero-shot claim testable rather than aspirational. m:…5262

What to watch

Read together, the four preprints describe a field growing up. The benchmarking paper says the rankings were softer than they looked. The three tools answer by building on frozen or shared foundations instead of training yet another model from scratch, and by testing on held-out proteins and unseen pairs rather than familiar ones. That is a healthier pattern than another round of leaderboard claims, and it is the pattern a builder choosing infrastructure should look for. m:…9759 m:…5262

The caveat is the same one that runs through all four papers: these are preprints, and the headline numbers come from the authors' own benchmarks. PALM outperforms AlphaMissense on a benchmark of 128,161 literature-derived missense variants, the authors report, which is a strong claim, but it is their benchmark and their comparison. The honest way to use this week's results is as a shortlist for your own evaluation, with homology-controlled splits, not as a final ranking. m:…6047 m:…9759