PeptideSGCL, a dual-contrastive learning framework, achieves superior peptide hemolysis and nonfouling prediction.

Peptides are crucial in biological processes and biomedical applications, but their hemolytic (Hemo) and nonfouling (NF) properties directly impact safety and translational potential. Accurate predictive models are essential for rational peptide design. Existing multimodal prediction methods, while using sequence and structural data, often rely on local graph convolution and cross-modal alignment. This limits their ability to model long-range structural dependencies and enhance intramodal discriminability, creating a gap in robust, high-fidelity property prediction.

Researchers developed PeptideSGCL, a multimodal dual-contrastive learning framework for peptide property prediction. The method integrates ProtBERT as a sequence encoder and a hierarchical GNN-Transformer structural encoder to capture both local topological patterns and long-range structural dependencies. A parallel graph spatial channel attention module was incorporated to enhance task-relevant structural features. Within a shared embedding space, an interintra hybrid supervised contrastive learning strategy was designed to optimize sequence-structure alignment and intramodal class discriminability, comparing its performance against existing baseline models on hemolysis and nonfouling prediction tasks.

The proposed PeptideSGCL method demonstrated overall performance superior to all baseline models across both the hemolysis and nonfouling prediction tasks. This superiority indicates that the enhanced structural encoder, which effectively captures local topological patterns and long-range structural dependencies, significantly improved the quality of joint sequence-structure representations. The integration of the parallel graph spatial channel attention module also contributed to enhancing task-relevant structural features.

The interintra hybrid supervised contrastive learning strategy successfully optimized both sequence-structure alignment and intramodal class discriminability, leading to more robust and discriminative peptide property predictions. These results confirm PeptideSGCL as an effective framework for multimodal representation learning in peptide-property prediction, addressing the limitations of previous methods in modeling complex peptide structures and enhancing class separation.