GeoSeq-EGNN: A geometry-sequence fused equivariant graph neural network for predicting short linear motifs within intrinsically disordered regions.

Short linear motifs (SLiMs) located within intrinsically disordered regions (IDRs) are essential for mediating dynamic protein-protein interactions. However, their short length, sequence degeneracy, and the conformational heterogeneity of IDRs pose significant challenges for accurate computational identification. Traditional methods that rely exclusively on one-dimensional sequences or static three-dimensional structures may not adequately capture the intricate interplay between structure and sequence, while severe class imbalance often results in unacceptably high false-positive rates. To address these challenges, we introduce GeoSeq-EGNN, a Geometry-Sequence Fused Equivariant Graph Neural Network designed for dual-modality SLiMs prediction. This architecture effectively integrates Global-Local Hybrid Equivariant Message Passing Layers (GLH-EMPL) to model the three-dimensional spatial folding dynamics of IDRs, alongside Multi-Scale Geometry-Sequence Fusion Network Layers (MSGS-FNL) that utilize Bidirectional LSTMs to extract critical one-dimensional primary sequence topologies. We also incorporate an adaptive coordinate update mechanism to improve geometric stability in deeper layers and dynamic gate feature fusion to reduce representation over-smoothing. To address class imbalance, we use a confidence-aware focal loss that penalizes overconfident erroneous predictions. On the evaluated SLiMs dataset, GeoSeq-EGNN achieved an accuracy of 92.94% and an AUPRC of 82.17% under the final F6 feature configuration and outperformed the evaluated baseline models in the controlled one-hot comparison, with higher Precision and Matthews Correlation Coefficient (MCC). These results support GeoSeq-EGNN as a computational framework for SLiM prediction and provide a basis for future studies of SLiM-mediated regulatory networks.