GPSFun: geometry-aware protein sequence function predictions with language models.

Area of Science:

• Computational biology
• Bioinformatics
• Structural biology

Background:

• Understanding protein function is crucial for disease research and drug target identification.
• A significant gap exists between the rapid increase in protein sequence data and the availability of functional annotations.
• Previous methods like GraphPPIS, GraphSite, LMetalSite, and SPROF-GO have been developed for protein function prediction.

Purpose of the Study:

• To present GPSFun, a versatile web server for Geometry-aware Protein Sequence Function annotation.
• To integrate language models and geometric deep learning to enhance protein function prediction.
• To provide a user-friendly platform for diverse downstream functional predictions.

Main Methods:

• Utilizing large language models to predict 3D protein conformations and extract sequence embeddings.
• Employing geometric graph neural networks to analyze sequence and structure patterns in protein graphs.
• Developing a web server with intuitive interfaces and visualizations for accessibility.

Main Results:

• GPSFun demonstrates superior performance compared to state-of-the-art methods on various prediction tasks.
• The server accurately predicts protein-ligand binding sites, gene ontologies, subcellular locations, and protein solubility.
• Effective function annotation is achieved without relying on multiple sequence alignments or experimental structures.

Conclusions:

• GPSFun offers a powerful and versatile tool for protein function annotation using sequence-based predictions.
• The integration of advanced AI techniques addresses the challenge of limited functional annotations in large protein sequence datasets.
• GPSFun is freely accessible, supporting broader research in molecular biology, drug discovery, and disease mechanism elucidation.