ToxiPep: Peptide toxicity prediction via fusion of context-aware representation and atomic-level graph

Abstract

Peptide-based therapeutics have emerged as a promising avenue in drug development, offering high biocompatibility, specificity, and efficacy. However, the potential toxicity of peptides remains a significant challenge, necessitating the development of robust toxicity prediction methods. In this study, we introduce ToxiPep, a novel dual-model framework for peptide toxicity prediction that integrates sequence-based contextual information with atomic-level structural features. This framework combines BiGRU and Transformer to capture local and global sequence dependencies while leveraging multi-scale CNNs to extract refined structural features from molecular graphs derived from peptide SMILES representations. A cross-attention mechanism aligns and fuses these two feature modalities, enabling the model to capture intricate relationships between sequence and structural information. ToxiPep outperforms several state-of-the-art tools, including ToxinPred2, CSM-Toxin, PepNet, and ToxinPred3, on both internal and independent test sets. Additionally, interpretability analyses reveal that ToxiPep identifies key amino acids along with their structural features, providing insights into the molecular mechanisms of peptide toxicity. To facilitate broader accessibility, we have also developed a web server for convenient user access. Overall, this framework has the potential to accelerate the identification of safer therapeutic peptides, offering new opportunities for peptide-based drug development in precision medicine.

Keywords: Deep learning; Drug discovery; Peptide bioactivity prediction; Sequence modeling.

Graphical abstract

Fig. 1

Framework of the ToxiPep. The model fuses sequence-based contextual representations and atomic-level structural features for enhanced peptide classification. Sequence embeddings are processed by BiGRU and Transformer encoder, while SMILES-derived graph features are refined via multi-scale convolutions. Cross-attention aligns both modalities, and an MLP classifier generates the final prediction.

Fig. 2

Amino acid composition differences between toxic and non-toxic peptides (CD-HIT threshold = 0.9) and their statistical significance. (A) Mean amino acid composition of toxic and non-toxic peptides. (B) Bonferroni-corrected p-values based on Wilcoxon rank-sum tests of amino acid composition differences. (C) Log-transformed corrected p-values representing the level of statistical significance for each amino acid.

Fig. 3

Performance comparison with machine learning and deep learning baseline models. Panels A–C present radar plots comparing the proposed model ToxiPep with three categories of baseline methods: (A) machine learning models using handcrafted features, (B) deep learning architectures trained on one-hot encoded sequences, and (C) pretrained protein language models.

Fig. 4

Performance comparison in ablation analysis. (A) and (B) Performance of ToxiPep compared with different ablation models, including Transformer, GRU + Transformer, Atomic-level graph with multi-scale CNN, Concatenation fusion, and Late fusion.

Fig. 5

UMAP visualizations of feature representations derived from different model modules under a single random seed. (A) Sequence embedding (silhouette = 0.00). (B) Structural encoding (silhouette = 0.00). (C) Context-aware sequence processing (silhouette = 0.33). (D) Atomic-level graph processing (silhouette = 0.34). (E) Without cross-attention fusion (silhouette = 0.41). (F) ToxiPep (silhouette = 0.44).

Fig. 6

Interpretability analysis of the model in predicting toxic peptides. (A) Attention weights of amino acids in a representative toxic peptide sequence. (B–E) Heatmaps generated by CAM analysis for CNN. (F–I) Molecular structure analysis of key amino acids.

Fig. 7

Demonstration of the ToxiPep web interface for peptide toxicity prediction. (1) Starting the Prediction: Users initiate the process by clicking the “Start Prediction” button, which opens a dedicated page for sequence input. (2) Input protein sequence: On the new page, users can enter peptide sequences directly into the text box or upload a file containing multiple sequences using the “Upload” button. This flexible input option supports both single and batch predictions. (3) Running the Prediction: After inputting the sequences, users click the “Start Prediction” button to submit them for analysis. (4) Results Panel: The results page presents the prediction outcomes, including the toxicity classification for each peptide and confidence scores from the model, displayed in an intuitive format.