ICM-MD: Integrating TM-specific features and MD-derived structures for accurate prediction of inter-chain contacts in...

Characterizing the interactions of alpha-helical transmembrane homodimers at the residue level is crucial for understanding their structure and function. However, most computational tools are designed for globular proteins and fail to translate to transmembrane (TM) proteins, largely due to the unique environment of the membrane and the limited availability of high-resolution structural data. To address this challenge, we present a data-centric machine learning framework that overcomes the scarcity of experimentally resolved TM homodimer structures, which are essential for developing a robust machine learning model. The proposed approach leverages MD-derived structural models as surrogate supervision for training the model. This model integrates sequence-based and structure-based features to enhance inter-chain residue contact prediction in TM homodimers. It also leverages a simple yet effective feed-forward neural network, designed to enhance model's interpretability and scalability. Comparative evaluation against state-of-the-art models, including DeepHomo1, DeepHomo2, Glinter, and DeepTMP, demonstrates that our method achieves improved performance. On a test set of eight alpha-helical TM homodimers, the model outperforms DeepHomo1 and DeepHomo2 with ΔPrecision@L = 42.2% and 43.9% respectively, surpasses Glinter by ΔPrecision@L = 34.6%, and achieves 7.5% higher precision compared to DeepTMP in the mean top L precision ranking metric.