Comparative Analysis of Deep Learning Algorithms for Disease-Related Protein Function Prediction: Performance Optimization and Computational Efficiency Evaluation

Abstract

Protein function prediction remains a fundamental challenge in computational biology, with direct implications for understanding disease and therapeutic development. This study presents a comprehensive comparative analysis of deep learning algorithms for disease-related protein function prediction, evaluating convolutional neural networks, recurrent neural networks, transformer-based models, and GNNs across standardised benchmarks. Our systematic evaluation framework encompasses 6,456 disease-associated proteins from UniProt and PDB databases, employing Gene Ontology annotations across molecular function, biological process, and cellular component categories. Performance metrics demonstrate that GNNs achieve superior Fmax scores of 0.758 for molecular function prediction, while transformer-based models balance accuracy and efficiency, achieving inference times about 1.75× faster than Bidirectional Long Short‑Term Memory (BiLSTM) and 2.6× faster than GNN, though slower than CNN by 2.1×. Multi-task learning frameworks enhanced prediction accuracy by 23% for rare GO terms, with transfer learning from pre-trained protein language models reducing training time by 65%. The analysis reveals critical trade-offs between prediction accuracy and computational resources, providing practical guidelines for algorithm selection in drug discovery pipelines.