LkaM-PTM: Predicting PTM sites through multimodal protein features from capturing cross-field information

Post-translational modification (PTM) site prediction is significant for a deeper understanding of biological processes and disease mechanisms. Most existing studies on predicting PTM sites focus on specific PTM types and lack universality. Additionally, few methods consider combining the advantages of sequence information, pre-trained information, and structural information. A multimodal adaptive PTM sites prediction method based on 1D Large Kernel Attention (1D-LKA) is thereby proposed, named LkaM-PTM. Specifically, LkaM-PTM fuses sequence representations from the SeqNet, pretrained representations from the PLMNet, and structural representations from the StruNet in the encoder part. Firstly, SeqNet utilizes 1D-LKA adaptively to capture cross-field information, combined with the feature reusability of DenseNet to extract local features of the sequence. Next, PLMNet captures global features from the protein language model ProtBert through a multilayer perceptron (MLP). Subsequently, StruNet compresses and reconstruct structural information of global, amino acid, and atomic granularity through a Stacked Autoencoder (SAE) to extract low-dimensional structural representations. Ultimately, the three feature sets are subsequently concatenated into a unified multimodal representation and fed into a MLP-based decoder for final PTM sites prediction. Experimental results show that LkaM-PTM outperforms the state-of-the-art approaches in the accuracy of predicting PTM sites. The source code and data for LkaM-PTM are available at https://github.com/tony-soochow/LkaM-PTM.