Skip to content
Open access

Enhanced Multi-Branch Attention Network for Toxicity Risk Prediction in Late-Stage Drug Development

Oct 2026 · VAWKUM Transactions on Computer Sciences · 0 citations · 22 references

TL;DR

An Enhanced Multi-Branch Attention Network is developed that combines Morgan fingerprints, MACCS keys, and physicochemical descriptors through separate feature-processing branches followed by attention-based fusion and suggests that the proposed multi-representation architecture provides a potentially useful screening aid for late-stage toxicity prediction.

Abstract

Late-stage toxicity can lead to substantial attrition in drug development, motivating computational methods that can prioritize compounds for further safety assessment. Recent work such as Trialblazer has demonstrated the utility of molecular fingerprints and bioactivity-derived features for this task. In this study, we develop an Enhanced Multi-Branch Attention Network that combines Morgan fingerprints, MACCS keys, and physicochemical descriptors through separate feature-processing branches followed by attention-based fusion. The training procedure incorporates focal loss, label smoothing, regularization, weighted sampling, and stratified 10-fold cross-validation. The reported experiments use 1,841 compounds for model development and an independent test set of 339 compounds. The reported ensemble test performance is ROC-AUC = 0.910 and MCC = 0.525, with the remaining performance measures retained exactly as obtained from the experimental implementation. The comparison with Trialblazer is presented as a benchmark comparison, while the lower recall of the proposed model is explicitly discussed as an important trade-off. The results suggest that the proposed multi-representation architecture provides a potentially useful screening aid for late-stage toxicity prediction, while experimental validation and careful interpretation remain necessary.

Read PDF

Similar papers

Review Open access 2026

Machine Learning in Drug Discovery: From QSAR Foundations to Deep Learning, Foundation Models, and Toxicity-Ready Prediction

Drug discovery remains constrained by long development timelines, high capital requirements, and substantial attrition, particularly during clinical development and late-stage safety evaluation. Estimates of research and development (R&D) investment required per approved therapeutic agent vary widely across studies and...

Yassine Benachour, Lina Maloukh, Ammar Natsheh · 0 citations
Open access Nov 2026

Identifying Short Half-Life Compounds from Structure Information Using Multi-Task Graph Neural Networks

Accurate prediction of terminal half-life (t1/2) is paramount for prioritizing drug candidates during the lead optimization stage, as it minimizes the risk of clinical trial failure while optimizing resource allocation. Drugs characterized by suboptimal t1/2 profileseither excessively short or prolongedpresent sign...

Ryoko Terada, Reiko Watanabe, H. Komura et al. · 0 citations
#machine learning Preprint Sep 2026

Benchmarking graph-based models for in-silico toxicity prediction in drug discovery

A unified and standardized benchmarking framework for GDL-based toxicity prediction is introduced, systematically evaluating more than 20 representative approaches under consistent experimental conditions and across multiple datasets and partitioning strategies, enabling a fair and reproducible comparison of model perf...

Noel Suarez-Barro, M. Lama, J. C. Vidal · 0 citations
Open access Oct 2026

HAFI-DTA: an interpretable multimodal framework for drug-target affinity prediction based on hierarchical attention guidance and feature iterative interaction

Accurate prediction of drug-target affinity (DTA) is a core technical step in high-throughput drug screening and targeted drug discovery, which directly affects the efficiency of candidate compound mining and the control of research and development costs. Traditional experimental methods are difficult to meet the requi...

Yi-Qun Wang, Shi-Long Song, Wei Tian et al. · 0 citations
Open access Sep 2026

Machine Learning for Toxicity Prediction in Low-Sample Molecular Classes

A framework that fine-tunes a general baseline model trained on publicly available data on small, class-specific datasets to retain the baseline’s generalization ability while gaining class-specific accuracy and produce probabilistic outputs that capture uncertainty in the training data is proposed.

C. Barajas, Laura L. Dunphy, Luke C. Mullany et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.