Loading…
Zero-shot mutation effect scoring for designed and viral proteins, mapping frozen ESM2 representations onto MD and normal-mode dynamic properties.
Conservation-based variant effect predictors have a structural blind spot. They infer whether a mutation is tolerated by asking what evolution has already tried, which works well for well-sampled protein families and fails for the two cases where the question is often most urgent: designed proteins, which have no evolutionary history at all, and fast-evolving viral proteins, whose alignments are shallow or misleading.
ESMDance, from Chao Hou, Haiqing Zhao, and Yufeng Shen at Columbia University, attacks that gap by adding a second, non-evolutionary source of signal: biophysical motion. It keeps a frozen ESM2-35M backbone and learns to map its representations onto per-residue and pairwise descriptors of fluctuation and co-movement, derived from molecular dynamics trajectories and normal-mode analysis of tens of thousands of proteins. The resulting model scores mutations zero-shot, and its gains over the backbone it sits on are largest exactly where conservation is weakest.
Its companion SeqDance trains the same architecture on the same dynamics from random initialization instead. The work was posted to bioRxiv in October 2024 and published in PNAS as "Protein language models trained on biophysical dynamics inform mutation effects."
The architecture is a Transformer encoder matching ESM2-35M: 12 layers, 20 attention heads per layer, an embedding dimension of 480, and roughly 35 million parameters. The backbone weights stay frozen; training fits the mapping from those representations to dynamic biophysical descriptors.
Training data are dynamic properties derived from MD trajectories and normal-mode analyses, combining sets spanning on the order of 64,000–65,000 proteins, with MMseqs2 deduplication applied. The full corpus — around 100 GB covering 64,403 proteins — is published alongside the model, together with train-test splits and an independent dynamicPDB test set.
Zero-shot mutation-effect benchmarks show substantial improvement over the frozen backbone alone, concentrated on designed and viral proteins. That concentration is the paper's central evidence: if the dynamics objective were merely re-encoding conservation, the gains would be uniform rather than clustered where evolutionary signal is absent.
The direct use is variant effect prediction where evolutionary information is limited — scoring mutations in de novo designed proteins during an engineering campaign, or prioritizing substitutions in fast-evolving viral proteins where alignment depth is insufficient for conservation-based scoring. Because it requires no finetuning and no labeled variants, it substitutes for a conventional protein language model scorer in an existing pipeline. Researchers interested in the dynamic properties themselves rather than in variant scoring are better served by SeqDance, which is trained on the same data without the evolutionary backbone.
The result reframes what protein language models can be supervised on. Dynamics turns out to be a training signal that adds information beyond conservation rather than duplicating it, and the practical payoff lands precisely on the proteins that defeat conservation-based methods. Peer-reviewed publication in PNAS, GPL-3.0 code, MIT-licensed weights, and a CC-BY-4.0 release of the whole training corpus make the approach reproducible end to end. The scale is modest at 35M parameters, inherited from the frozen backbone, so scaling the dynamics objective to larger encoders is the natural next question and one the authors leave open.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.