DPLM: Dynamics-aware Protein Language Model via contrastive learning between sequence and molecular dynamics...

Area of Science:

• Computational biology
• Protein bioinformatics
• Machine learning in structural biology

Background:

• Protein dynamics are crucial for function but often overlooked in protein language models (PLMs).
• Existing PLMs primarily rely on static sequence or structure information, missing dynamic insights.
• Accurate modeling of protein dynamics is essential for understanding biological mechanisms.

Purpose of the Study:

• To develop a novel protein language model, DPLM, that incorporates protein dynamics.
• To align sequence embeddings with molecular dynamics (MD) trajectory embeddings using contrastive learning.
• To enhance PLMs' ability to capture biologically relevant dynamic properties for improved functional predictions.

Main Methods:

• Utilized molecular dynamics (MD) features encoded by a pre-trained video model.
• Employed contrastive learning to align sequence embeddings with MD trajectory embeddings.
• Evaluated DPLM on zero-shot mutation-effect prediction, functional clustering, protein stability, and disorder prediction tasks.

Main Results:

• DPLM learns sequence representations correlated with residue-level flexibility.
• DPLM significantly improves protein-level functional clustering compared to static PLMs.
• Achieved superior zero-shot mutation-effect prediction performance against ESM-based representations.
• Demonstrated top-tier results in protein stability and disorder prediction after lightweight adaptation.

Conclusions:

• Contrastive alignment with MD trajectories enables PLMs to capture meaningful dynamic properties.
• DPLM represents a significant advancement in modeling protein dynamics for functional insights.
• The dynamics-aware approach enhances predictive capabilities across various protein-related tasks.