Multi-task learning with knowledge distillation for intra- and inter-protein PTM crosstalk prediction

Post-Translational Modifications (PTMs) in proteins play a critical role in regulating diverse cellular processes. PTM sites often interact to determine a functional outcome, a phenomenon known as PTM crosstalk. This interplay occurs within the same protein (intra-protein) or across different proteins (inter-protein). However, existing computational approaches either address only one type of PTM crosstalk or treat them independently, overlooking their shared regulatory principles. In this paper, we propose a novel Multi-task Learning model with Knowledge Distillation for predicting intra- and inter-protein crosstalk of PTMs, namely MIXTAP. To the best of our knowledge, this is the first work to jointly learn both types of PTM crosstalks via a multi-task learning strategy. The MIXTAP model formulates intra- and inter-crosstalk as two distinct tasks, trains single-task teacher models, and distills their knowledge into a multi-task student model. To capture the interactions between PTM sites, MIXTAP incorporates a task-specific bidirectional cross-conditioned feature reweighting mechanism, enabling the model to account for the unique crosstalk patterns of PTMs within and across proteins. Experimental results on publicly available datasets demonstrate that the proposed MIXTAP model consistently outperforms state-of-the-art methods. Finally, for reproduction purposes, we release our implementation of MIXTAP at https://github.com/egiovanoudi/MIXTAP . Not applicable.