KPTM-MCHR: A multichannel hypergraph regression model for predicting the frequencies of lysine post-translational modification types in proteins.

Lysine post-translational modifications (K-PTMs) are important forms of protein post-translational modification. Identifying K-PTMs in protein sequences is an important topic in bioinformatics. Existing computational models mainly focus on predicting specific K-PTM types at lysine residues in protein sequences. Few studies have investigated the number of K-PTM types that can be annotated to a given protein sequence. In this study, we present a pioneering model for predicting this number, which we term the frequency of protein K-PTM types. Three hypergraphs are constructed from a comprehensive protein similarity matrix. A multichannel hypergraph learning framework then generates high-level protein representations by refining raw features produced by a protein large language model. Node- and channel-attention mechanisms further refine and integrate the protein features, which are then fed into a multilayer perceptron to learn a regression function. Ten-fold cross-validation on a dataset constructed from K-PTM records in the Compendium of Protein Lysine Modifications (CPLM) demonstrates the model's strong performance. KPTM-MCHR outperforms baseline models based on traditional regression methods and models adapted from other protein-related prediction tasks. Ablation experiments support the rationale for the model design. We further examine the relationship between protein similarity and differences in K-PTM-type frequency to investigate why the model can successfully predict this target.