KAT-Net: a hyperbolic routing framework for joint lysine acylation site prediction with shared sequence-structure representations.

Abstract

Protein lysine acylation regulates protein function through changes in local chemistry, structure, and molecular interactions. Acetylation (Kac), succinylation (Ksucc), and crotonylation (Kcr) share the lysine centre but differ in acyl-group chemistry, motivating joint yet task-aware prediction. We developed KAT-Net, a multi-task framework integrating ESM-2 representations, physicochemical priors, and AlphaFold2-derived structural context. Predicted aligned error-aware graph modelling reduces uncertain structural relationships, while task-driven hyperbolic routing retrieves modification-specific information from shared representations. A masked asymmetric focal loss supports learning with unavailable labels and task-dependent class imbalance. Independent benchmark and external Kcr evaluations showed favourable performance, while ablation and single-task comparisons supported contributions from sequence-structure integration, structural modelling, hyperbolic routing, and multi-task learning. Sequence and structural analyses revealed modification-related positional, residue, and spatial patterns around candidate lysines. External functional analysis using an independent mouse lung acetylome and succinylome provided complementary evidence of functional plausibility, with Kac predictions showing partial concordance with experimentally reported metabolic processes rather than serving as a binary test of general Kac-site discrimination. For Kac prediction, model development and primary benchmarking were performed using the curated non-histone acetylation dataset from TransPTM and should be interpreted within this benchmark scope rather than as a comprehensive evaluation of the complete acetylome. KAT-Net therefore provides a unified framework for lysine-acylation prediction and candidate-site prioritization within the evaluated benchmark settings.