PepTriX: A Framework for Explainable Peptide Analysis through Protein Language Models

TL;DR

PepTriX tackles accuracy and interpretability in peptide classification by uniting 1D sequence context from ESM-2 with 3D structural context from ESMFold in a lightweight GAT framework. The model uses contrastive learning and cross-modal co-attention to align modalities and provide motif-level explanations without fine-tuning large PLMs. It achieves competitive performance across diverse datasets and delivers interpretable insights linking sequence features to structural determinants. The approach reduces computational barriers while offering domain experts actionable structural hypotheses for peptide design and discovery.

Abstract

Peptide classification tasks, such as predicting toxicity and HIV inhibition, are fundamental to bioinformatics and drug discovery. Traditional approaches rely heavily on handcrafted encodings of one-dimensional (1D) peptide sequences, which can limit generalizability across tasks and datasets. Recently, protein language models (PLMs), such as ESM-2 and ESMFold, have demonstrated strong predictive performance. However, they face two critical challenges. First, fine-tuning is computationally costly. Second, their complex latent representations hinder interpretability for domain experts. Additionally, many frameworks have been developed for specific types of peptide classification, lacking generalization. These limitations restrict the ability to connect model predictions to biologically relevant motifs and structural properties. To address these limitations, we present PepTriX, a novel framework that integrates one dimensional (1D) sequence embeddings and three-dimensional (3D) structural features via a graph attention network enhanced with contrastive training and cross-modal co-attention. PepTriX automatically adapts to diverse datasets, producing task-specific peptide vectors while retaining biological plausibility. After evaluation by domain experts, we found that PepTriX performs remarkably well across multiple peptide classification tasks and provides interpretable insights into the structural and biophysical motifs that drive predictions. Thus, PepTriX offers both predictive robustness and interpretable validation, bridging the gap between performance-driven peptide-level models (PLMs) and domain-level understanding in peptide research.

Figures (3)

• Figure 1: The framework predicts peptide function by integrating a one-dimensional sequence, or blueprint vector, and a three-dimensional structure, or local summary vector. It uses a dual-encoder architecture to learn a shared representation of these two modalities. The one-dimensional sequence is processed using a protein language model (ESM-2) to capture contextual and evolutionary features. Meanwhile, the three-dimensional structure is generated using a model like ESMFold and analyzed by a graph attention convolution network (GATConv). Contrastive training ensures alignment between the vectors. A co-attention module generates a co-attention matrix for domain expert validation. GATConv specializes in identifying unique structural motifs relevant to a specific dataset.
• Figure 2: Interpretability analysis for four different sequences from the datasets: aip_antiinflam, amp_modlamp, hiv_v3 and nep_neuropipred dataset. The heatmaps show the attention matrix learned by the classifier, with rows corresponding to sequence residues and columns to structural residue indices. Brighter colors show higher attention and darker colors lower attention.
• Figure 3: Interpretability analysis for four sequences from the pip_pipel dataset. Two sequences were successfully identified by the model as proinflammatory-inducing and two as non-proinflammatory-inducing. The heatmaps show the attention matrix learned by the classifier, with rows corresponding to sequence residues and columns to structural residue indices. Brighter colors show higher attention and darker colors lower attention.