Combined Molecular Fingerprint and Descriptor Features Enable Classical Machine Learning to Match Deep Learning Performance Across the Tox21 Toxicity Panel

BACKGROUND AND OBJECTIVE: Molecular fingerprints and physicochemical descriptors encode complementary structural information, yet most Tox21 benchmarks evaluate only one representation at a time. This study examined whether integrating both closes the reported gap between classical and graph-based deep learning. METHODS: Six algorithms (Random Forest, XGBoost, LightGBM, SVM, MLP, Logistic Regression) were trained on a 3,131-dimensional feature vector combining 59 RDKit descriptors with ECFP4, ECFP6, and RDKit topological fingerprints (1,024 bits each), across 8,014 Tox21 compounds and 12 endpoints, evaluated under both stratified 5-fold cross-validation and a Bemis-Murcko scaffold split. Bootstrap resampling (n = 1,000) gave 95% confidence intervals. RESULTS: Random Forest achieved the highest mean AUC-ROC of 0.846 (SD = 0.053) under stratified 5-fold cross-validation, with 10 of 12 endpoints exceeding 0.80. Under a Bemis-Murcko scaffold split matching the deep learning benchmark protocol, RF mean AUC was 0.839, at or slightly above the published aggregate means of AttentiveFP (0.829) and GROVER (0.831); because these benchmarks report only a single aggregate mean, comparison is aggregate-level only. Adding fingerprints to the descriptor-only baseline improved RF on 10 of 12 endpoints (maximum gain +0.028 on SR-ATAD5). A linear-kernel SVM (0.752) outperformed the originally reported RBF-kernel SVM (0.710), a kernel-calibration artifact; SVM remained the weakest baseline. CONCLUSION: An integrated fingerprint-descriptor representation lets classical machine learning match published graph neural networks on Tox21 under matched scaffold-split evaluation, without specialized hardware, offering a reproducible, interpretable alternative for computational toxicology.