Off-Manifold Collapse in Guided Protein Language Models

MAXWELL: Calibrating the probabilistic outputs of protein language models to the mutation-induced stability change landscape

Designing mutations that enhance protein stability is a central goal in protein engineering. However, experimentally screening large numbers of candidate mutations is costly and time-consuming, creating a strong need for computational methods that can identify potentially stabilizing mutations. Among these approaches, protein language models are particularly promising because they learn context-dependent amino acid preferences from large-scale sequence and structure datasets. Nevertheless, most existing stability prediction methods use these models primarily as feature extractors and do not fully exploit the amino acid probability distributions they encode. Here, we introduce MAXWELL (Matrix-wise Landscape Learning), a novel post-training method that calibrates the probabilistic outputs learned by protein language models during pretraining to generate mutational landscapes that quantify the effects of individual amino acid substitutions on protein stability. When applied to ProteinMPNN, MAXWELL yields a state-of-the-art predictor of the effects of protein mutations on stability, outperforming ThermoMPNN and other representative methods on a curated benchmark of experimentally measured stability changes. We next applied MAXWELL to the design of ten single-point mutations in the DhaA dehalogenase, seven of which (70%) increased thermal stability. Among them, G171W showed the largest improvement, with a measured ΔTm of 4.91 °C. These experimental results establish MAXWELL as a novel post-training strategy for protein language models and a practical framework for designing stabilizing mutations. Repository https://github.com/ai4protein/Venus-MAXWELL

Property guidance for protein sequence generative models with ProteinGuide.

ProtEnrich: Residual Multimodal Enrichment of Protein Sequence Embeddings

Protein language models effectively capture evolutionary and functional signals from sequence data but lack explicit representation of the biophysical properties that govern protein structure and dynamics. Existing multimodal approaches attempt to integrate such physical information through direct fusion, often requiring multimodal inputs at inference time and distorting the geometry of the sequence embedding space, which can disrupt the semantic organization learned from evolutionary information. Consequently, a fundamental challenge of how to incorporate structural and dynamical knowledge into sequence representations without disrupting their semantic organization, enabling sequence-based models to better capture the biophysical properties governing protein structure and function. We introduce ProtEnrich, a representation learning framework based on a residual multimodal enrichment paradigm. Pro-tEnrich decomposes sequence embeddings into two complementary latent subspaces, an anchor subspace that preserves sequence semantics, and an alignment subspace that encodes biophysical relationships. By converting multimodal information derived from ProstT5 and RocketSHP to a low-energy residual component, our approach injects physical representation while maintaining the original sequence embedding while preserving their original semantic geometry, avoiding the need for multimodal inputs at inference time. Across eight diverse protein foundational models trained on 550,120 SwissProt proteins with AlphaFold structures, enriched embeddings improved zero-shot remote homology retrieval, increasing Precision@10 and MRR by up to 0.13 and 0.11, respectively. Downstream performance also improved on structure-dependent tasks, reducing fluorescence prediction error by up to 16% and increasing metal ion binding AUCROC by up to 2.4 points, while requiring only sequence input at inference. Source code is available at https://github.com/pcdslab/ProtEnrich, pretrained models and datasets are available at https://huggingface.co/collections/SaeedLab/protenrich.

Cold-Start Protein–Protein Interaction Prediction Is Bounded by the Frozen Embedding, Not the Classifier: A Leakage-Controlled Audit and a Calibrated Conformal Baseline

Sequence-based protein–protein interaction (PPI) prediction has reached an empirical plateau on leakage-controlled benchmarks, where the dominant signal under the cold-start regime—predicting interactions among proteins absent from training—comes from evolutionary-scale protein language models (PLMs) rather than from the downstream classifier. A natural hypothesis is that elaborate head-only optimization over a frozen PLM can narrow the in-domain-to-cold-start gap cheaply. We test this hypothesis and report a negative-but-constructive result. We build a many-objective NSGA-III search that attaches a compact head to a frozen ESM-2 t33 backbone and jointly optimizes the head architecture, a physicochemical feature mask, and a contrastive-pretraining curriculum against a fitness vector that explicitly contains the cold-start gap; we then subject it to a controlled audit. On a leakage-controlled yeast DIP benchmark with Park–Marcotte stratification (homology-disjoint at 40% identity), the searched contrastive head, trained to a fair budget, reaches cold-start (C3) average precision (AUPR) 0.683, whereas an off-the-shelf random forest on the same embeddings reaches 0.734; gradient boosting and an RBF-SVM also exceed it. Seven design axes—objective, architecture, pair operator, embedding compression, pooling, training budget, and training-set size—each fail to close the gap, indicating the achievable level is set by the frozen representation, not the classifier. On a second, human benchmark (Bernett), simple classifiers again reach the published state-of-the-art range ( $\approx 0.65$ vs. 0.69 AUPR) at a fraction of the cost. Finally, we add distribution-free calibrated selective prediction via split conformal, turning a frozen-ESM classifier into a deployable cold-start predictor with finite-sample coverage. We release all code and reproduction drivers.

Coevolution-informed Bayesian optimization for sample-efficient protein design

This work introduces ALSEBO (Active Learning Sequence Exploration via Bayesian Optimization), which couples a generative latent sequence landscape to Bayesian optimization and featurizes candidates with direct-coupling-analysis (DCA) coevolutionary statistics.

Shifu: an integrated framework for deep learning of RNA secondary structure

Deep learning has advanced RNA secondary-structure prediction by bypassing explicit energy rules to capture long-range dependencies, yet progress is limited less by model scale than by how structures are measured: single scores hide where and why models fail, and benchmark scores can reflect memorization of one dataset rather than genuine generalization. We address this with Shifu, a framework of three coupled parts. Shifu-Corpus is a leakage-audited dataset of 254123 sequences from six databases, with family-aware splits certified free of exact and near-duplicate leaks. The Shifu Trifecta scores a model on three axes (correctness, breadth across diverse RNAs, and whether its confidence can be trusted) rather than one number. Shifu-LMR, a family of compact RNA language models, serves as controlled experiments: changing the training corpus shifts accuracy by 0.13, and a 65-million-parameter model, Shifu-LMR-Nano, leads on correctness while running on a laptop. We release the dataset, code, and model backbones.