ToxCML: A Hybrid mfCoQ‑RASAR-Based Platform Integrating Consensus QSAR and Read-Across for Comprehensive Multi-Endpoint Toxicity Assessment

Authors
Fauzan Syarif Nursyafi¹, Muhammad Adnan Pramudito², Yunendah Nur Fuadah³, Rahmafatin Nurul Izza², Abdul Latif Fauzan4, and Ki Moo Lim¹˒⁴*

¹ Computational Medicine Lab, Department of Biomedical Engineering, Kumoh National Institute of Technology, Gumi, 39177, Republic of Korea
² Computational Medicine Lab, Department of IT Convergence Engineering, Kumoh National Institute of Technology, Gumi, 39177, Republic of Korea
³ Telecommunication Engineering Study Program, School of Electrical Engineering, Telkom University Main Campus, Bandung, Indonesia
4 Meta Heart Co., Ltd, Gumi, 39253, Republic of Korea

*Corresponding author: kmlim@kumoh.ac.kr

ToxCML is a large-scale hybrid mfCoQ‑RASAR (multi-feature Consensus quantitative Read-Across Structure–Activity Relationship) platform that explicitly integrates consensus QSAR and consensus read-across into a weight-optimized hybrid predictor for multi-endpoint toxicity assessment. The framework is designed to provide chemically contextualized, applicability-domain–aware predictions to support large-scale toxicity screening, hazard prioritization, and reduction of animal testing.

---

Framework overview

ToxCML web platform

Study overview

Conventional animal-based toxicity testing is costly, time-consuming, and ethically constrained, motivating robust in silico alternatives. In this work, ToxCML:

• Curates and standardizes SMILES-based toxicity datasets following QSAR-ready best practices (validity checks, salt/solvent removal, parent structure selection, charge normalization, tautomer handling, stereochemical consistency, duplicate resolution, and label harmonization).

• Computes multiple complementary molecular representations using RDKit and CDK: MACCS keys, Morgan circular fingerprints, APF, RDKit fingerprints, and physicochemical descriptors.

• Trains descriptor-specific QSAR models (Random Forest, XGBoost, SVM) with 10-fold scaffold-based cross-validation, selects the best model per descriptor using the composite score:

  Sm = (AUCm + BACCm) / 2

  and builds an Sm-weighted consensus QSAR.

• Implements multi-fingerprint read-across (k-NN similarity-weighted using Tanimoto on MACCS, Morgan, APF, RDKit fingerprints) and builds an Sf-weighted consensus RA using:

  Sf = (AUCf + BACCf) / 2

• Constructs a hybrid mfCoQ‑RASAR ensemble that combines consensus QSAR and consensus RA probabilities via a linear integration:

  P_mfCoQ-RASAR = w * P_QSAR,cons + (1 - w) * P_RA,cons

  where the global weight w is optimized by 10-fold scaffold-CV on the training set.

• Applies a tiered applicability-domain (AD) framework (QSAR AD, RA AD, and mfCoQ‑RASAR AD as their intersection) to flag in-domain vs out-of-domain predictions.

Across 18 toxicity endpoints and evaluation on unseen test or external validation sets, mfCoQ‑RASAR models achieve AUC ≈ 0.86–0.99 and BACC ≈ 0.73–0.98, consistently outperforming individual consensus QSAR and consensus read-across while preserving interpretability and AD transparency.

---

Toxicity endpoints

ToxCML covers 18 curated toxicity endpoints spanning acute systemic toxicity, organ-specific toxicity, and drug-induced safety liabilities.

No.	Endpoint	Brief definition
1	AMES Mutagenicity	Bacterial mutagenicity (e.g., Salmonella typhimurium), used as a screening assay for genotoxic potential.
2	Acute Dermal Toxicity	Systemic adverse effects or mortality after single/short-term dermal exposure, typically LD₅₀-based.
3	Acute Inhalation Toxicity	Systemic toxicity following short-term inhalation exposure to gases, vapors, or aerosols.
4	Acute Oral Toxicity	Systemic effects or death after a single oral dose, often categorized by LD₅₀ classes.
5	Carcinogenicity	Tumorigenic potential under repeated or chronic exposure.
6	Cardiotoxicity	Structural or functional cardiac injury (contractility, conduction, rhythm).
7	DILI (Drug-Induced Liver Injury)	Drug-related liver damage ranging from enzyme elevations to acute liver failure.
8	Developmental Toxicity	Adverse embryo–fetal outcomes (malformations, growth retardation, prenatal death).
9	Drug-Induced Nephrotoxicity (DIN)	Renal injury affecting glomerular/tubular function or clearance.
10	Eye Irritation	Reversible/irreversible ocular damage following direct exposure.
11	FDA MDD	Systemic toxicity at FDA-oriented maximum daily dose thresholds.
12	Hematotoxicity	Toxicity to blood and the hematopoietic system, including altered cell counts or marrow function.
13	Hepatotoxicity	Structural or functional liver injury (e.g., enzyme elevations, cholestasis, hepatocellular damage).
14	Mitochondrial Toxicity	Impaired mitochondrial function (e.g., oxidative phosphorylation disruption, oxidative stress).
15	Neurotoxicity	Adverse effects on CNS or PNS, impacting neuronal structure or function.
16	Respiratory Toxicity	Toxic effects in the respiratory tract (airway inflammation, bronchoconstriction, lung parenchymal damage).
17	Skin Irritation	Non-immunologic inflammatory skin response (erythema, edema) after topical exposure.
18	Skin Sensitization	Immune-mediated allergic contact dermatitis after repeated exposure.

---

ToxCML: A Hybrid mfCoQ‑RASAR

This repository contains the full implementation of the ToxCML / mfCoQ-RASAR framework, a hybrid in silico toxicity prediction platform that integrates consensus QSAR, similarity-based read-across, and their optimized coupling across multiple toxicity endpoints.

The workflow emphasizes:

• robust scaffold-aware validation,
• broad chemical-space coverage,
• transparent applicability-domain (AD) definition, and
• reliability-aware prediction with confidence intervals.

---

Repository Structure

This repository mirrors the local ToxCML project layout.

README.md

Main documentation describing the study motivation, workflow, and usage.

---

Dataset/

• QSAR-ready curated datasets for 18 toxicity endpoints
  (SMILES + binary Outcome labels).
• Structural and label curation already applied.
• For selected endpoints, scaffold-based train/test splits are provided.

---

Molecular Descriptor Computation_Preprosesing data (1).ipynb

QSAR-ready data curation and descriptor computation:

• SMILES validation and removal of unparsable or chemically implausible structures.
• Molecular standardization:
  • salt and solvent removal,
  • parent structure selection,
  • charge normalization,
  • harmonized cleanup rules,
  • consistent handling of tautomers and stereochemistry.
• Duplicate detection using canonical SMILES or InChIKey.
• Removal of compounds with conflicting labels.

Outputs:

• Fingerprints:
  • MACCS (166-bit)
  • Morgan (1024-bit)
  • APF (1024-bit)
  • RDKit fingerprints (2048-bit)
• 49 RDKit/CDK-based physicochemical descriptors.

---

Training_Consensus QSAR_Fingerprint_10foldCrossvalidation (1).ipynb

• Trains fingerprint-based QSAR models (RF, XGB, SVM) using MACCS, Morgan, and APF fingerprints.
• Uses 10-fold Bemis–Murcko scaffold-based cross-validation.
• For each descriptor–algorithm model m, computes:
  • AUCₘ
  • BACCₘ
  • composite score
    Sₘ = (AUCₘ + BACCₘ) / 2
• Selects the best model per fingerprint type based on the highest Sₘ.
• Saves the selected models together with their Sₘ values.

---

Training_Consensus QSAR_PhysicochemicalProperties_10foldCrossvalidation (1).ipynb

• Trains QSAR models on physicochemical descriptors (RF, XGB, SVM).
• Uses 10-fold scaffold-based cross-validation.
• Computes Sₘ for each descriptor–algorithm combination.
• Selects the best physicochemical QSAR model per endpoint.

---

Model Performance Evaluation 1_Consensus QSAR.ipynb

• Builds the consensus QSAR model by aggregating:
  • MACCS-based QSAR
  • Morgan-based QSAR
  • APF-based QSAR
  • physicochemical QSAR
• Uses Sₘ-based weights, where
  weightₘ = Sₘ / Σ Sₖ
• Consensus QSAR prediction for compound i is the weighted average of individual QSAR predictions.
• Evaluates performance on strictly unseen test or external sets:
  • AUC, BACC, ACC, SEN, SPE
  • 95% bootstrap confidence intervals.

---

Model Performance Evaluation 2_Consensus Read-Across Evaluation.ipynb

• Implements k-nearest-neighbor similarity-weighted read-across.
• Uses MACCS, Morgan, APF, and RDKit fingerprints.
• Computes Tanimoto similarity and selects the k nearest neighbors.
• Derives similarity-weighted toxicity probabilities.
• Reports baseline read-across performance per fingerprint and per endpoint.

---

Weight Optimization_Consensus Read Across.ipynb

Required step for building consensus read-across and mfCoQ-RASAR.

• Uses 10-fold scaffold-based cross-validation on the training set.
• For each fingerprint type f, computes:
  • AUC_f
  • BACC_f
  • composite score
    S_f = (AUC_f + BACC_f) / 2
• Derives S_f-based weights:
  v_f = S_f / Σ S_g
• Builds the final consensus read-across as the weighted average of fingerprint-specific read-across predictions.
• These consensus read-across probabilities are used as the read-across component in mfCoQ-RASAR.

---

mfCoQ-RASAR_Evaluation_Framework.ipynb

• Constructs and evaluates the hybrid mfCoQ-RASAR predictor, which linearly combines:
  • consensus QSAR predictions, and
  • consensus read-across predictions.
• Performs 10-fold Bemis–Murcko scaffold-based cross-validation to optimize the global coupling weight w.
• In each fold:
  • consensus QSAR and consensus read-across are trained on 9 folds,
  • predictions are generated for the validation fold.
• For each candidate weight w, computes:
  • AUC
  • BACC
  • composite score
    S_f,w = (AUC_f,w + BACC_f,w) / 2
• Averages performance across folds:
  • S_w = (1 / K) × Σ S_f,w
• Selects the optimal coupling weight w* that maximizes S_w.
• Retrains final models on the full training set using w* and evaluates on test or external sets with confidence intervals.

---

AD Analysis/ and AD Analysis.ipynb

• Pre-computed applicability-domain (AD) annotations for:
  • QSAR,
  • read-across,
  • mfCoQ-RASAR.
• QSAR AD:
  • maximum Tanimoto similarity (MACCS, Morgan, APF),
  • leverage in physicochemical descriptor space,
  • combined by majority voting.
• Read-across AD:
  • based on mean similarity to the k nearest neighbors,
  • thresholds derived from training-set distributions.
• mfCoQ-RASAR AD:
  • defined as the intersection of QSAR AD and read-across AD.

---

SHAP/ and SHAP Analysis (2).ipynb

• SHAP-based explainable AI analysis for top-performing QSAR models.
• Uses:
  • TreeExplainer for RF and XGB,
  • KernelExplainer for SVM.
• Provides fragment-level contribution maps by projecting important Morgan fingerprint bits onto molecular structures.

---

.gitattributes

Git configuration for consistent line endings and handling of large or binary files.

---

Workflow: reproducing the mfCoQ-RASAR framework

1. Data curation and preprocessing
   Run Molecular Descriptor Computation_Preprosesing data (1).ipynb to generate QSAR-ready datasets and molecular descriptors.

2. Descriptor-specific QSAR model development
   Train fingerprint-based and physicochemical QSAR models using scaffold-based cross-validation and compute $S_m$ scores.

3. Consensus QSAR evaluation
   Build the $S_m$-weighted consensus QSAR model and evaluate it on unseen test or external datasets.

4. Read-across development and weighting
   Perform multi-fingerprint read-across and derive the $S_f$-weighted consensus read-across model.

5. Hybrid mfCoQ-RASAR construction
   Optimize the QSAR–read-across coupling weight $w$, retrain models on the full training set, and evaluate final performance.

6. Applicability-domain analysis and explainability
   Analyze prediction reliability using applicability-domain definitions and interpret QSAR predictions using SHAP.

---

How to cite

If you use this repository, the models, or any derived results in your work, please cite the corresponding ToxCML mfCoQ‑RASAR manuscript, you may reference it as:

ToxCML: A Hybrid mfCoQ-RASAR-Based Platform Integrating Consensus QSAR and Read-Across for Comprehensive Multi-End Point Toxicity Assessment Fauzan Syarif Nursyafi, Muhammad Adnan Pramudito, Yunendah Nur Fuadah, Rahmafatin Nurul Izza, Abdul Latif Fauzan, and Ki Moo Lim Journal of Chemical Information and Modeling, DOI: 10.1021/acs.jcim.6c00357