AIRPred: A Deep Learning Model Predictor for Peptide Intensity Ratios in Cross-Linking Mass Spectrometry Improves Cross-Link Spectrum Matching

Cross-linking mass spectrometry (XL-MS) is a powerful tool in structural proteomics, offering insights into protein conformations, interactions and dynamics by linking spatially proximal residues. However, current cross-linked spectrum match (CSM) scoring methods rely heavily on mass-to-charge ratio (m/z) comparisons, often neglecting fragment ion intensity information, which limits their ability to accurately distinguish true CSMs from false positives. To overcome this limitation, we present AIRPred, a deep learning model that predicts intensity ratios between cross-linked peptide pairs to improve CSM identification. AIRPred employs convolutional neural network (CNN) blocks to capture peptide fragmentation patterns and an attention layer to model peptide interactions. Our results show that intensity ratios remain consistent across experiments and can reliably differentiate true CSMs from random mismatches. In external validation, AIRPred outperformed traditional methods, demonstrating high accuracy in predicting intensity ratios. This model significantly enhances XL-MS analysis by leveraging intensity data for more accurate peptide identification.