Combined Molecular Fingerprint and Descriptor Features Enable Classical Machine Learning to Match Deep Learning...

Area of Science:

• Computational toxicology
• Cheminformatics
• Machine learning

Background:

• Tox21 benchmarks typically evaluate molecular fingerprints or physicochemical descriptors separately.
• A gap exists between classical and graph-based deep learning methods in predicting chemical toxicity.

Purpose of the Study:

• To investigate if integrating molecular fingerprints and physicochemical descriptors can close the performance gap between classical and graph-based deep learning on Tox21 benchmarks.
• To evaluate the effectiveness of combining different molecular representations for toxicity prediction.

Main Methods:

• Trained six algorithms (Random Forest, XGBoost, LightGBM, SVM, MLP, Logistic Regression) on a combined feature vector of RDKit descriptors and ECFP/RDKit topological fingerprints.
• Evaluated models on 8,014 Tox21 compounds across 12 endpoints using stratified 5-fold cross-validation and a Bemis-Murcko scaffold split.
• Utilized bootstrap resampling for confidence intervals.

Main Results:

• Random Forest achieved the highest mean AUC-ROC (0.846) under stratified cross-validation and (0.839) under scaffold split, matching or exceeding graph neural network performance.
• Integrating fingerprints with descriptors improved Random Forest performance on 10 of 12 endpoints.
• A linear-kernel SVM outperformed a previously reported RBF-kernel SVM, highlighting the impact of representation and model choice.

Conclusions:

• An integrated fingerprint-descriptor representation enables classical machine learning to achieve performance comparable to graph neural networks on Tox21.
• This approach provides a reproducible, interpretable, and hardware-efficient alternative for computational toxicology.
• Combining diverse molecular representations is key to enhancing predictive accuracy in toxicity assessment.