LM-GVP: an extensible sequence and structure informed deep learning framework for protein property prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Proteins are crucial for biological systems and therapeutic development.
• Predicting protein properties from sequence and structure is essential for bio-therapeutics.
• Current methods may not fully leverage both sequence and structural data.

Purpose of the Study:

• To develop a generalizable deep learning framework for predicting protein properties.
• To integrate information from 1D amino acid sequences and 3D protein structures.
• To improve upon existing protein language models for property prediction tasks.

Main Methods:

• Developed LM-GVP, a novel framework combining a protein Language Model (LM) and a Graph Neural Network (GNN).
• Leveraged both 1D amino acid sequences and 3D protein structures as input.
• Evaluated performance on property prediction tasks: fluorescence, protease stability, and Gene Ontology (GO) functions.

Main Results:

• LM-GVP outperformed state-of-the-art protein LMs on multiple property prediction tasks.
• Demonstrated the effectiveness of integrating structural information via a GNN.
• Provided insights into fine-tuning LMs using GNN prediction heads for enhanced structural information utilization.

Conclusions:

• The LM-GVP framework offers a generalizable approach for diverse protein property prediction problems.
• This method significantly enhances the prediction accuracy by integrating sequence and structural data.
• The framework has the potential to accelerate protein engineering and drug development pipelines.