Deep learning-based integrated framework for protein structure prediction and functional annotation

Abstract

Deep learning has become a powerful paradigm in computational biology, offering data-driven models for mapping protein sequences to structure and function with high precision. In the context of protein structure prediction and functional annotation, traditional sequence-alignment and template-based methods are often limited by low homology and structural diversity. To overcome these challenges, we propose a unified deep learning framework that integrates sequence modeling, structural representation, and functional inference into a single architecture. The system employs transformer-based encoders to extract contextual features from amino acid sequences and graph neural networks to capture spatial dependencies within predicted structures. A multi-task learning approach is designed to perform C _α backbone reconstruction and enzyme class prediction simultaneously. The framework leverages joint training on sequence features and predicted inter-residue geometry to improve generalization on rare or multifunctional proteins. Experiments on public benchmark datasets such as CASP14 and CAMEO-Hard demonstrate a 12.4% reduction in backbone RMSD compared to AlphaFold2 and a 9.7% improvement in PR-AUC for contact prediction, validating the effectiveness of joint learning and structural integration. Compared to existing state-of-the-art baselines including RoseTTAFold and trRosetta, our system achieves consistently superior accuracy across structure and function tasks. This work provides a modular, end-to-end solution for large-scale protein analysis and shows potential for extension to downstream tasks such as mutation effect prediction or protein–protein interaction modeling.