Learning Biophysical Dynamics with Protein Language Models.

Area of Science:

• Computational biology
• Protein dynamics
• Bioinformatics

Background:

• Protein function is intrinsically linked to dynamic structural ensembles, yet modeling these dynamics remains challenging.
• Current methods struggle to integrate complex protein dynamics into deep learning for function and fitness studies.
• Efficient representation of protein dynamics is crucial for advancing biological understanding.

Purpose of the Study:

• To develop a novel protein language model, SeqDance, capable of learning dynamic properties directly from amino acid sequences.
• To address the limitations in representing and utilizing protein dynamics in computational biology and deep learning.
• To provide a sequence-based method for understanding protein behavior and function.

Main Methods:

• SeqDance was pre-trained on extensive biophysical data from molecular dynamics trajectories and normal mode analyses.
• The model learns to represent local dynamic interactions, co-movement patterns, and global conformational features.
• The approach focuses on deriving dynamic properties solely from the protein sequence.

Main Results:

• SeqDance successfully captures essential local and global dynamic features of proteins, even those without known homologs.
• The model demonstrates improved prediction accuracy for protein fitness landscapes and binding regions.
• SeqDance effectively identifies disorder-to-order transition regions and phase-separating proteins.

Conclusions:

• SeqDance offers a powerful new method for learning protein dynamics from sequence, complementing existing structure- and evolution-based approaches.
• This sequence-based dynamic representation provides novel insights into protein behavior and function.
• SeqDance has the potential to significantly advance deep learning applications in protein science.