PANDA-3D: protein function prediction based on AlphaFold models.

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein function prediction traditionally relies on amino acid sequences.
• Limited experimental structures and poor predicted structure quality hindered structure-based methods.
• The AlphaFold protein structure database (AlphaFold DB) offers a rapidly growing resource of predicted protein tertiary structures.

Purpose of the Study:

• To develop a deep-learning tool, PANDA-3D, for predicting Gene Ontology (GO) terms from AlphaFold-predicted protein structures.
• To leverage the expanding AlphaFold DB for enhanced protein function annotation.
• To create a tool specifically trained on AlphaFold models.

Main Methods:

• Developed an advanced deep-learning architecture combining geometric vector perceptron graph neural networks and transformer decoder layers.
• Trained the model using AlphaFold-predicted structures and amino acid sequence embeddings from a large language model.
• Implemented a multi-label classification approach for GO term prediction.

Main Results:

• PANDA-3D significantly outperformed a state-of-the-art deep-learning method trained on experimental structures.
• PANDA-3D achieved comparable or superior performance against other leading language-model-based methods using amino acid sequences.
• The tool is tailored for AlphaFold models, facilitating annotation of the vast AlphaFold DB.

Conclusions:

• PANDA-3D provides accurate protein function annotation using AlphaFold structures.
• The tool is valuable for annotating the large and growing number of proteins in the AlphaFold DB.
• PANDA-3D is accessible via a web server and a GitHub repository.