[AML] Enhancing Protein Function Prediction: Integrating Pretraining and Fine-Tuning within Geometric-Aware Graph Neural Networks

THU 2024 Winter AML Submission8 Authors

Keywords: Bioinformatics, Deep Learning

Abstract: Proteins are essential to biological processes, and accurate function prediction is vital for advancing molecular biology and therapeutic development. Traditional methods often face challenges with low-similarity proteins and novel families, while recent deep learning approaches leveraging sequence and structural data have shown promise. To address limitations in existing methods, we propose a geometric-aware graph neural network (GNN) framework that explicitly models protein structures through node and edge features, incorporating radial basis functions and Fourier encoding to capture spatial relationships. Combined with a self-supervised pretraining task on large-scale, unlabeled structural datasets, our method demonstrates competitive performance across benchmarks, suggesting its potential to enhance protein function prediction and contribute to further progress in the field.

Submission Number: 8