Protein language models trained on biophysical dynamics inform mutation effects.

Area of Science:

• Computational Biology
• Structural Biology
• Protein Dynamics

Background:

• Protein structure dynamics are crucial for function and mutation impact.
• Existing deep learning models often neglect protein dynamics, using static structures or sequences.
• Capturing protein dynamics is essential for accurate functional and mutational effect predictions.

Purpose of the Study:

• Introduce SeqDance and ESMDance, novel protein language models incorporating protein dynamics.
• Enable prediction of dynamic properties for both ordered and disordered proteins.
• Enhance the prediction accuracy of mutation effects by integrating dynamic information.

Main Methods:

• Trained SeqDance and ESMDance on dynamic biophysical properties from molecular dynamics and normal mode analyses of over 64,000 proteins.
• SeqDance utilizes attention mechanisms to model residue comovement and encodes dynamic representations.
• ESMDance builds upon the ESM2 architecture, leveraging its outputs for enhanced predictions.

Main Results:

• SeqDance embeddings capture rich protein dynamics, enabling transfer learning for conformational properties.
• SeqDance predictions accurately reflect mutation effects on protein folding stability.
• ESMDance significantly outperforms ESM2 in zero-shot mutation effect prediction for proteins lacking evolutionary data.

Conclusions:

• SeqDance and ESMDance provide a robust framework for integrating protein dynamics into language models.
• These models offer more generalizable predictions of protein behavior and mutation effects.
• The developed models advance the field of protein deep learning by incorporating crucial dynamic information.