전산의학연구실

Why predictive toxicology?

Tens of thousands of industrial chemicals and drug candidates need safety data, but animal and laboratory tests are slow, costly and ethically constrained. Safety problems remain a major reason drug candidates fail, and blocking the heart's hERG potassium channel is a classic cause of drug-induced arrhythmia. Regulators increasingly accept new approach methodologies, including computer models, in place of animal tests.

Two computational tools answer the question from different directions:

QSAR (quantitative structure–activity relationship): a model learns how structural features relate to a toxic effect across thousands of tested chemicals, then predicts untested ones.
Read-across: a new chemical is judged from its most similar tested neighbours. It is intuitive and used in regulatory submissions, but it needs close analogues to exist.

Our models combine the two. Several QSAR models vote as a consensus, and similarity to known chemicals is added as extra evidence (read-across structure–activity relationship, RASAR). The aim is predictions that are accurate, explainable, and grounded in chemicals we already know.

Research Journey

Our QSAR work grew along two lines: from single models to consensus models combined with chemical similarity, and from general hazards to the heart. In 2024 we showed that a consensus of machine-learning classifiers predicts chemical hazards more accurately and more stably than the best single model. In 2026 we added deep learning and several fingerprint types across 14 endpoints, then built ToxCML, which weighs consensus QSAR against read-across for 18 endpoints and 54,601 chemicals. In parallel, Cardiosim-Tox brought multimodal, multitask deep learning to the three cardiac ion channels at the centre of the lab's in silico cardiac safety work. Both run as web servers, and their data and code are public.

Phase 1 (2024): consensus QSAR for chemical hazards

Our first question was whether combining models beats trusting one. We trained several machine-learning classifiers for multiple chemical hazard endpoints and merged their votes into one consensus prediction. The consensus predicted more accurately and more stably than the best individual model, the principle behind all our later platforms.

Paper, ACS Omega 2024

Phase 2 (2026): interpretable multi-modality consensus

Next we asked whether different views of a molecule carry different information. The framework predicts 14 toxicity endpoints (8 of 14 toxicity endpoints showed on UMAP-based chemical space distribution), 4 local and 10 systemic, from four fingerprint types (Morgan, MACCS, atom-pair and RDKit). It combines machine-learning models (random forest, XGBoost, support vector machine) with deep-learning models in one consensus and explains which features drive each prediction.

Paper, Toxicology Mechanisms and Methods 2026

Phase 3 (2026): ToxCML, consensus QSAR meets read-across

ToxCML (hybrid multi-feature consensus quantitative read-across structure–activity relationship, mfCoQ-RASAR) treats QSAR and read-across as two independent experts and learns, for each endpoint, how much to trust each one.

• Data: 54,601 curated chemicals across 18 endpoints: eye irritation, skin irritation, skin sensitization, respiratory toxicity, acute oral, dermal and inhalation toxicity, carcinogenicity, Ames mutagenicity, drug-induced liver injury, hepatotoxicity, nephrotoxicity, neurotoxicity, hematotoxicity, mitochondrial toxicity, developmental toxicity, cardiotoxicity (hERG) and FDA maximum daily dose. Nine endpoints were split by molecular scaffold, so test chemicals come from chemical series the model has not seen.

• Model: machine-learning QSAR models (random forest, XGBoost, support vector machine) on MACCS, Morgan and atom-pair fingerprints plus 49 physicochemical descriptors, and k-nearest-neighbour read-across on MACCS, Morgan, atom-pair and RDKit fingerprints. Endpoint-specific weights give QSAR 55–82% of the final vote and read-across the rest.

• Reliability: a tiered applicability domain marks each prediction as in or out of domain; more than 95% of test chemicals were in domain for most endpoints. SHAP fragment maps show which parts of a molecule push it toward toxic; for cardiotoxicity they highlight lipophilic aromatic and halogenated groups, the classic hERG-blocker pattern.

• Results: AUC about 0.86–0.99 and balanced accuracy 0.73–0.98 on unseen and external sets. The hybrid was the best or near-best model on every endpoint and matched or exceeded the OECD QSAR Toolbox, VEGA, EPA TEST and ADMETlab 3.0.

Paper, J Chem Inf Model 2026

Data (Zenodo) · Code (GitHub)

Cardiac track (2026): Cardiosim-Tox

Cardiotoxicity accounts for about 10–14% of drug withdrawals from the market, and it usually starts with block of the heart's ion channels. The CiPA initiative (Comprehensive in vitro Proarrhythmia Assay), led by the US FDA and partners, looks at hERG, Cav1.2 and Nav1.5 together rather than hERG alone. Yet most QSAR models still predict hERG only, and only as a yes-or-no answer.

Cardiosim-Tox answers both questions for all three channels at once: does the drug block the channel, and how potently (pIC50)?

• Data: 31,816 curated compounds. Three external test sets are stratified by structural similarity to the training set, from close analogues to distant chemistry.

• Model: Morgan and MACCS fingerprints, Mordred 2D/3D descriptors and molecular graphs (graph convolutional network) fused in one multitask network that shares what it learns across channels.

• Results: AUC 0.93–0.98 for Cav1.2, 0.88–0.96 for Nav1.5 and 0.79–0.95 for hERG across the external sets, and potency R² up to 0.94. Multitask learning helped most for Cav1.2 and Nav1.5, which have the least data. For hERG potency it reached R² 0.82, against 0.52 for Pred-hERG 5.0 on the same external set.

• Explanation: SHAP rankings match known pharmacophores, such as lipophilicity and a basic nitrogen for hERG blockers and the dihydropyridine scaffold for Cav1.2 blockers.

Paper, Archives of Toxicology 2026

Data (Zenodo) · Code (GitHub)

From Lab to Platform

ToxCML

Get a prediction for toxicity endpoints

ToxCML

Cardiosim-Tox

Block risk and potency for hERG, Cav1.2 and Nav1.5

Cardiosim-Tox (Login Required)

Page updated

Google Sites

Report abuse