What happened
Protein language models (PLMs) such as ESM-2 encode protein sequences as embeddings for downstream tasks. PLMs are trained on a masked learning objective that leverages evolutionary constraints. While interpretability studies of ESM-2 have focused on folded proteins, their behavior on intrinsically disordered proteins (IDPs), which constitute a substantial fraction of the human proteome and are implicated in numerous diseases, remains understudied. Because IDPs experience different types of evolutionary constraints on their amino acid sequences, we hypothesized that PLMs would behave differently on disordered versus folded regions. Here we show that ESM-2 exhibits reduced attention on disordered regions, yet still encodes meaningful biological signals. The model assigns heightened attention to disease-relevant residues even at high levels of disorder. Moreover, we show that both the radius of gyration and individual dynamic contact maps, key characteristics of IDPs, can be obtained from the model logits and embeddings. These findings suggest PLMs capture valuable information relevant to IDP biology despite their bias toward structured residues. Update (2026-09-18): Virtual experiments bridge sequence and microscopy with generative models . Large-scale screening and mapping efforts have produced vast libraries of perturbation-readout data. Converting these measurements into mechanistic insights requires models that link perturbation and genetic input to phenotypes, i.e., labels from experimental readouts, which are usually task specific. We propose a different, virtual experiment modeling approach: train generative models to recreate readouts conditioned on the experimental context, and then let established downstream models extract phenotypes from the synthetic data. As an illustrative case, we develop a bidirectional sequence-image generative framework, CELL-FM, that maps protein sequence and cellular context to fluorescence microscopy images and back, enabling in silico localization prediction, image-conditioned functional motif analysis and generation, and large-scale virtual mutagenesis revealing the amino acid features controlling condensate formation of intrinsically disordered peptides. This approach decouples representation learning from task-specific annotation, reuses rich experimental modalities across ma Update (2026-09-24): Protein Language Models: Learning From Evolution, Designing Beyond It . Protein language models (PLMs) have transformed our ability to learn from evolutionary sequence space, but protein engineering ultimately asks a different question: not what evolution selected, but what we should build next. Zero-shot likelihoods therefore provide useful, but not universal, measures of fitness and can misalign with engineering objectives. Experimental supervision redirects these priors towards properties of interest, enabling target-specific prediction and closed-loop optimization. PLMs thereby complement structure-based design: structural methods provide geometric control, while PLMs integrate experimental feedback to optimize functional and developability properties. Realizing this potential requires evaluation beyond retrospective predictive accuracy, towards extrapolation, multi-objective optimization, and prospective experimental success. We argue for recurring competitions combining scalable predictive benchmarks with prospective experimental challenges. Ultimately, progress should be measured not by predicting existing experiments, but by enabling successful ne