An SE(3)-equivariant and dynamic multi-modal engine advancing from PTM site prediction to network understanding

Protein post-translational modifications (PTMs) orchestrate complex cellular processes, yet current computational methods predominantly tre at them as isolated sequence events. This reductionist approach-coupled with an inability to model 3D steric constraints and long-tail data scarcity-obscures the cooperative nature of PTM networks. Here, we introduce ProteinNexus, a multimodal deep-learning framework designed to bridge the gap between local site prediction and global network comprehension. By dynamically fusing evolutionary sequence data with a physically grounded, conformal equivariant graph neural network, ProteinNexus rigorously captures the 3D microenvironments of modification sites. Coupled with a reinforcement learning-driven specialization strategy, the framework effectively models crosstalk across 10 diverse PTM types, outperforming state-of-the-art methods by 7.5% (Matthews Correlation Coefficient). Beyond superior predictive accuracy, in silico knockouts and evolutionary analyses reveal that ProteinNexus successfully decodes cooperative PTM dynamics, providing a generalizable, structure-aware computational microscope to unravel the regulatory grammar of the proteome.