The abstract is read at the publisher; the summary is JClub's.
DynamicsPLM, a 35-million-parameter dynamics-enriched protein language model, predicts zero-shot mutation effects from sequence by balancing evolutionary plausibility with structural dynamics disruption.
- Why it matters: Protein variant effect prediction is limited because PLMs capture evolutionary constraints from sequences but do not model explicit folded-protein physical dynamics, while dynamics-only models miss evolutionary sequence statistics.
- What they did: The method uses a second-stage, multi-task continual pretraining on over 60,000 proteins, adding residue-level and pairwise molecular dynamics heads to a foundational sequence encoder.
- The result: On the ProteinGym benchmark and natural variants of the Tsuboyama folding stability dataset, it achieves state-of-the-art accuracy among models under 100M parameters and matches or outperforms several billion-scale sequence-based models of its size.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.