A Co-attention Mechanism of Jointly Capturing Sequence and Structure Features for Signal Peptide Prediction

Signal peptides are short amino acid sequences at the N-terminus of proteins that serve as subcellular localization signals, directing proteins to specific compartments such as mitochondria, chloroplasts, or the secretory pathway. Although many recent predictors based on protein language models have achieved strong performance, most of them rely primarily on sequence information and do not explicitly incorporate structural cues. Here, we propose Signal-3L 4.0, a dual-path encoder framework that combines pretrained protein representations with task-specific sequence and structural modeling for signal peptide prediction, and the 2 modalities are fused via a VisualBERT-style co-attention module. To alleviate the impact of the long-tail distribution across organism groups and signal peptide types, we further employ a class-balanced label-distribution-aware margin loss. Benchmark results show that Signal-3L 4.0 achieves improved signal peptide classification performance and competitive or better cleavage-site prediction performance compared with SignalP 6.0, with particularly superior performance in several low-resource or challenging settings. Signal-3L 4.0 is freely available for academic use as a web server at http://www.csbio.sjtu.edu.cn/bioinf/Signal-3L/ and as open-source code at https://github.com/chandlevier/Signal-3L-4.0.