PepTriX: A Framework for Explainable Peptide Analysis through Protein Language Models
Vincent Schilling, Akshat Dubey, Georges Hattab
TL;DR
PepTriX is a versatile framework that combines sequence and structural features using graph attention networks to improve peptide classification while offering interpretability of biologically relevant motifs.
Contribution
It introduces PepTriX, a novel method integrating 1D and 3D peptide features with contrastive learning and co-attention for improved, interpretable peptide analysis.
Findings
High performance across multiple peptide classification tasks
Provides interpretable insights into structural motifs
Automatically adapts to diverse datasets
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…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
The main strength of the paper is that the approach is strongly motivated by biologically relevant questions. 1. This paper is strongly motivated by the underlying biological problems which addresses the interpretability crisis in theses models. 2. The method leverages multi modal input integrating 1D and 3D features generated by in silico oracles. 3. Strong empirical evidence with large number of datasets tested.
The main weakness of this model is the novelty and the lack of systemic analysis of the results beyond anecdotal observation. 1. This method mainly compose of existing methods, no material novel insights regarding the rationale for the design choices. This paper will hugely benefit from a more principled study of why and how the design choices are made and logical reasoning connecting them back to motivation. 2. The results are mostly anecdotal, no benchmarking on interpretability metrics such
1. **Important Problem:** The paper tackles a critical and relevant problem: creating generalizable, computationally efficient, and interpretable models for peptide function prediction. 2. **Comprehensive Related Work:** The authors provide a thorough review of existing methods for peptide classification and interpretability, clearly positioning their work within the current landscape. 3. **Logical Framework:** The core idea of aligning a 1D sequence "blueprint vector" with a 3D structural "summ
1. **Limited Novelty:** The main weakness of this paper is its lack of originality. The framework is an assembly of standard parts: pre-trained ESM-2 embeddings, ESMFold for structure generation, a GAT for processing graphs, contrastive loss for modal alignment, and co-attention for heatmaps. The paper fails to prove that this specific combination provides a unique advantage that other, simpler combinations would lack. 2. **Weak and Self-Contradictory Interpretability Claim:** The paper's primar
The paper have clear probelm framing andprincipled architecture coupling 1D PLM embeddings with 3DGAT, aligned by constractive learning. Interpretable co-attention that links residue tokens to structural neighborhoods; examples match biophysical expectations (AMP aromatic/cationic residues; HIV V3 electrostatics).
Table 1 lacks standardized, like-for-like comparisons versus strong PLM or structure-aware baselines trained under identical data splits and budgets; limited ablation of each module’s contribution (GAT vs. 1D only; contrastive vs. none; co-attention vs. none). Qualitative heatmaps are compelling but need systematic human-expert studies (blinded scoring, inter-rater reliability) or quantitative localization metrics (e.g., enrichment near known motifs/contacts).Heavy reliance on ESMFold structures
- Interpretation and explanation of deep neural networks, especially foundation models, in peptide analysis is an important and timely problem of interest to the ICLR audience. - The biological explanations of the derived attention maps are interesting (though unclear whether they are generalizable).
- It is somewhat unclear what the specific contribution of the paper is. As written, the paper primarily serves as a lengthy review of background and related work, making it challenging to understand where the proposed method begins. Most of the content before pages 6-7 could be moved into a background or related work section, allowing the proposed method to be emphasized in a dedicated section. Concepts such as attention, graph attention network, and contrastive learning are familiar enough to
1. A new model is developed that combines 1D and 3D information of peptides for property prediction. 2. Expert analysis is provided for some predictions.
1. The model is tested on multiple datasets, but it is difficult to compare with prior work. Are the same train/test splits used for the baselines? Are hyperparameters tuned on a validation set? What is the dissimilarity between train/test splits? More details are needed on the experimental setup to support state-of-the-art claims, or to estimate the validity of the findings. 2. While expert analysis is a nice first step towards interpretability, a global analysis is needed to claim that the mod
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsMachine Learning in Bioinformatics · Antimicrobial Peptides and Activities · Protein Structure and Dynamics