SiliS-PTOXRA

SiliS-PTOXRA predicts acute oral toxicity (AOT) by applying QSAR models that use QuBiLS-MAS molecular encodings to assess toxicological endpoints and support hazard and risk assessment.

Key Features:

• Target endpoint: Quantitative prediction of acute oral toxicity (AOT) values and toxic/non‑toxic labeling according to the globally harmonized system (GHS).
• Molecular encoding (QuBiLS‑MAS): Uses quadratic, bilinear, and N‑linear maps based on graph‑theoretic electronic‑density matrices and atomic weightings to encode molecular structure.
• Training and validation datasets: Models were trained on EPA (5931 compounds), EPA‑full (7413 compounds), and Zhu (10,152 compounds) sets and validated on EPA test (1482 compounds), ProTox (425 compounds), and T3DB (284 compounds) external sets.
• Machine learning algorithms: Base models were built using k‑nearest neighbor, multilayer perceptron, random forest, and support vector machine algorithms.
• Model selection criteria: Base models with correlation coefficients R ≥ 0.75 and mean absolute error (MAE) ≤ 0.5 on the EPA training set were selected for consensus modeling.
• Consensus models: Three consensus models (M19, M22, M24) were derived, with M19 and M22 using the minimum operator and M24 using a weighted average operator.
• Applicability domain performance: M19 showed MAE = 0.4044 (EPA test), 0.4067 (ProTox), and 0.2586 (T3DB); M22 (EPA‑full) showed MAE = 0.3992 (ProTox) and 0.2286 (T3DB); M24 (Zhu) showed MAE = 0.3773 (ProTox) and 0.2471 (T3DB).
• Statistical benchmarking: Models were compared against 14 QSAR methods from the literature, including admetSAR and ProTox‑II, with M22 showing superior performance across evaluated metrics.
• Retrospective safety assessment: A retrospective analysis on 261 drugs withdrawn for toxicity or side effects was used to evaluate model labeling of toxic compounds.

Scientific Applications:

• AOT prediction: Quantitative and classification predictions of acute oral toxicity for chemical compounds.
• Virtual screening: Prioritization of compounds for toxicological risk based on QSAR predictions and consensus model outputs.
• Regulatory labeling support: Assignment of toxic/non‑toxic labels according to GHS criteria to inform hazard classification.
• Benchmarking and method comparison: Performance comparison against existing QSAR methods such as admetSAR and ProTox‑II.
• Retrospective toxicovigilance analysis: Evaluation of withdrawn drug sets to examine model agreement with observed adverse outcomes.

Methodology:

QuBiLS‑MAS molecular encodings (quadratic, bilinear, N‑linear maps from graph‑theoretic electronic‑density matrices and atomic weightings) were used to build QSAR models trained on EPA (5931), EPA‑full (7413), and Zhu (10,152) datasets, validated on EPA test (1482), ProTox (425), and T3DB (284); base models used k‑nearest neighbor, multilayer perceptron, random forest, and support vector machine, with selection by R ≥ 0.75 and MAE ≤ 0.5, followed by construction of consensus models using minimum and weighted average operators, applicability domain analysis, and statistical comparison to 14 literature QSAR methods including admetSAR and ProTox‑II, plus a retrospective analysis of 261 withdrawn drugs.