NoiTox: A Machine Learning Strategy for Peptide Toxicity Prediction Reveals Label Noise in Benchmark Datasets
- 1Universidade Federal de Vicosa
Abstract
Motivation: Peptide toxicity remains a major bottleneck in drug development, as small sequence variations can drastically alter cytotoxic profiles. Experimental assessment is costly and time-consuming, motivating computational predictors. Existing approaches often rely on single black-box architectures, limiting interpretability and obscuring the contribution of individual sequence features. Furthermore, benchmark datasets frequently define the negative class through annotation filters rather than direct experimental evidence, introducing structural label noise that can compromises training and evaluation. Results: We present NoiTox, a multi-model strategy for peptide toxicity prediction integrating three complementary architectures, evaluated as four model variants: NoiToxML (Bo/full), an Extremely Randomized Trees classifier with Boruta-selected or complete physicochemical, compositional, and atomic descriptors; NoiToxDP, a deep learning model leveraging ESM-2 protein language model embeddings; and NoiToxGP, a graph neural network that combines ESM-2 residue embeddings, a fully connected residue graph with attention, and the Boruta descriptors as global features. Averaged over five random seeds, the models approached the state-of-the-art predictor ToxiPep on CD-HIT-controlled test sets, and on an independent dataset the interpretable NoiToxML variants and a probability-averaging consensus of the four models exceeded it, showing that physicochemical descriptors are highly informative for toxicity prediction. Cross-referencing negative-labeled sequences against curated databases of hemolytic, cytotoxic, and neurotoxic peptides flagged a small set annotated as toxic elsewhere, consistent with label noise by omission. However, removing the sequences that all models contradicted did not improve, and for some models reduced, ranking performance, indicating that such sequences carry useful information and are better treated as candidates for curation than discarded automatically. Availability and implementation: https://github.com/goalmeida05/NoiTox.
Related
- When Does Active Learning Improve ADMET Prediction? A Multi-Endpoint Benchmark
- Towards AI-Driven Nanomedicine Discovery: A Benchmark and Multimodal Learning Framework for Nano Self-Assembly Prediction
- MutCleaner: Cleaning and Standardizing Biological Mutation Datasets for Variant Effect Prediction
- GPBench: A benchmark framework for genomic prediction
- NEPTUNE: a neuronal cell death pathway prediction tool used to decipher neurodegeneration using single cell RNA-seq datasets
BibTeX
@misc{almeida2026noitox,
title = {{NoiTox: A Machine Learning Strategy for Peptide Toxicity Prediction Reveals Label Noise in Benchmark Datasets}},
author = {G. O. Almeida and M. C. Albuquerque and B. M. Silva and O. A. B. Silva and R. S. Ferreira and W. Meira and D. V. Pires and W. M. Prata and S. A. Silveira},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.08.757741},
url = {https://www.biorxiv.org/content/10.64898/2026.10.08.757741v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - NoiTox: A Machine Learning Strategy for Peptide Toxicity Prediction Reveals Label Noise in Benchmark Datasets
AU - G. O. Almeida
AU - M. C. Albuquerque
AU - B. M. Silva
AU - O. A. B. Silva
AU - R. S. Ferreira
AU - W. Meira
AU - D. V. Pires
AU - W. M. Prata
AU - S. A. Silveira
PY - 2026
DA - 2026/10/08
PB - bioRxiv
DO - 10.64898/2026.10.08.757741
UR - https://www.biorxiv.org/content/10.64898/2026.10.08.757741v1
AB - Motivation: Peptide toxicity remains a major bottleneck in drug development, as small sequence variations can drastically alter cytotoxic profiles. Experimental assessment is costly and time-consuming, motivating computational predictors. Existing approaches often rely on single black-box architectures, limiting interpretability and obscuring the contribution of individual sequence features. Furthermore, benchmark datasets frequently define the negative class through annotation filters rather than direct experimental evidence, introducing structural label noise that can compromises training and evaluation. Results: We present NoiTox, a multi-model strategy for peptide toxicity prediction integrating three complementary architectures, evaluated as four model variants: NoiToxML (Bo/full), an Extremely Randomized Trees classifier with Boruta-selected or complete physicochemical, compositional, and atomic descriptors; NoiToxDP, a deep learning model leveraging ESM-2 protein language model embeddings; and NoiToxGP, a graph neural network that combines ESM-2 residue embeddings, a fully connected residue graph with attention, and the Boruta descriptors as global features. Averaged over five random seeds, the models approached the state-of-the-art predictor ToxiPep on CD-HIT-controlled test sets, and on an independent dataset the interpretable NoiToxML variants and a probability-averaging consensus of the four models exceeded it, showing that physicochemical descriptors are highly informative for toxicity prediction. Cross-referencing negative-labeled sequences against curated databases of hemolytic, cytotoxic, and neurotoxic peptides flagged a small set annotated as toxic elsewhere, consistent with label noise by omission. However, removing the sequences that all models contradicted did not improve, and for some models reduced, ranking performance, indicating that such sequences carry useful information and are better treated as candidates for curation than discarded automatically. Availability and implementation: https://github.com/goalmeida05/NoiTox.
ER -