Multi-Scale Representation Learning for Protein Fitness Prediction

Area of Science:

• Computational Biology
• Protein Engineering
• Bioinformatics

Background:

• Accurate modeling of protein fitness landscapes is essential for designing novel functional proteins.
• Current methods often rely on self-supervised learning from sequence or structure data, with limited success in integrating both modalities effectively.
• Existing models overlook the critical role of detailed surface topology in determining protein function.

Purpose of the Study:

• To introduce a novel multimodal representation learning framework, the Sequence-Structure-Surface Fitness (S3F) model.
• To effectively integrate protein sequence, backbone structure, and surface topology features for fitness prediction.
• To overcome limitations of previous sequence-only and sequence-structure models.

Main Methods:

• Developed the Sequence-Structure-Surface Fitness (S3F) model, a multimodal framework.
• Integrated protein sequence representations from a protein language model.
• Utilized Geometric Vector Perceptron networks to encode protein backbone and detailed surface topology.

Main Results:

• Achieved state-of-the-art protein fitness prediction on the ProteinGym benchmark.
• Demonstrated superior performance across 217 substitution deep mutational scanning assays.
• Provided novel insights into the key determinants of protein function.

Conclusions:

• The S3F model represents a significant advancement in protein fitness prediction.
• Integrating sequence, structure, and surface topology offers a more comprehensive approach to understanding protein function.
• This framework has implications for accelerating the design of novel functional proteins.