Abstract
Deep learning for protein function prediction faces significant challenges in identifying disease-specific proteins. We present Extracellular Matrix Protein Predictor (EPOP), an advanced transfer learning framework leveraging protein language models to decode disease mechanisms. Focusing on pelvic organ prolapse (POP), which affects up to 50% of women worldwide, EPOP demonstrates AI's power to reveal novel therapeutic targets. We developed a sophisticated fine-tuning protocol for the ESM-2 model, optimized for ECM protein prediction. Our architecture integrates specialized attention mechanisms with interpretability modules, trained on expertly curated and balanced datasets totaling 80,000 proteins (40,000 ECM and 40,000 non-ECM). The framework employs a novel validation strategy using a 16,000-sample independent test set and clinical proteomics data. EPOP achieved unprecedented performance (99.40% accuracy) in ECM protein classification, significantly surpassing traditional deep learning architectures (10.81% improvement over Transformer models, 21.71% over Long Short-Term Memory). Applied to clinical samples, our model revealed a previously unknown pattern of ECM remodeling, identifying 24 novel disease-associated proteins. Model interpretability analysis uncovered specific sequence motifs and structural features critical for ECM protein function, providing mechanistic insights into disease progression. EPOP demonstrates how advanced AI bridges molecular analysis and clinical applications, uncovering novel therapeutic targets. Its success suggests broader applications across ECM-related disorders, potentially transforming approaches to diseases affecting connective tissue architecture.