RDKit Descriptors Capture Most Predictive Performance Across Twenty-Two Drug Endpoints

A comprehensive benchmarking study evaluating fifty-four distinct molecular representation configurations across twenty-two drug absorption, distribution, metabolism, excretion, and toxicity endpoints has demonstrated that standard physicochemical and structural descriptors recover nearly all the predictive power achieved by far more complex hybrid representations. Conducted under a rigorous, fixed stacked-ensemble machine learning architecture using the standardized Therapeutics Data Commons benchmark, the investigation revealed that a compact set of two hundred and seventeen RDKit molecular descriptors alone achieved an average of ninety-four point three percent of the best observed predictive performance across all modeled pharmacological properties, with a median retention of ninety-six point four percent. This compact baseline surprisingly outperformed combined structural fingerprints encompassing two thousand three hundred and sixty-three features, which retained only ninety-three point two percent of top performance. Incorporating six additional feature modalities, including structural fingerprints, tokenized sequence encodings, and randomly initialized neural projections, lifted average performance retention to only ninety-six point two percent, demonstrating that dramatic escalations in feature dimensionality frequently provide marginal utility in computational drug property estimation.

Predicting the absorption, distribution, metabolism, excretion, and toxicity profiles of candidate molecules remains one of the most critical determinants of whether a prospective therapeutic compound survives preclinical screening or fails due to unacceptable pharmacokinetic liabilities or adverse organ toxicities. While early computational profiling allows researchers to triage vast regions of chemical space without relying immediately on costly low-throughput biological assays, previous modeling efforts have predominantly concentrated on inventing novel neural network architectures while conflating model designs with molecular representations. Furthermore, many widely used web-based prediction platforms rely on proprietary training sets, non-standardized validation splits, or undisclosed data overlap, making it virtually impossible to isolate whether superior performance stems from algorithmic refinement, feature engineering, or data leakage. To overcome these long-standing confounding factors, the researchers evaluated all modeling configurations using the official Therapeutics Data Commons benchmark, which enforces strict scaffold-based partitioning to guarantee that evaluation compounds contain entirely distinct molecular scaffolds from those encountered during model training.

The benchmark encompassed twenty-two standardized tasks spanning five core pharmacokinetic and safety domains, consisting of thirteen binary classification tasks and nine continuous regression endpoints. Absorption properties included Caco-2 cell permeability, human intestinal absorption, P-glycoprotein transport activity, oral bioavailability, lipophilicity, and aqueous solubility. Distribution endpoints examined blood-brain barrier permeability, plasma protein binding rates, and volume of distribution at steady state. Metabolism evaluations encompassed substrate specificity and cytochrome P450 isoform inhibition across CYP2C9, CYP2D6, and CYP3A4. Excretion models analyzed human plasma half-life alongside hepatocyte and microsomal metabolic clearance rates, while toxicity endpoints assessed acute oral toxicity, hERG cardiac potassium channel inhibition, AMES bacterial mutagenicity, and drug-induced liver injury. Each endpoint was monitored strictly through its official benchmark metric, employing the area under the receiver operating characteristic curve or the area under the precision-recall curve for classification problems, and mean absolute error or Spearman rank correlation coefficients for continuous regression challenges.

To isolate representation effects while holding the underlying learner constant, the researchers developed ADMET-Stack, a specialized stacked-generalization framework deployed identically across all feature configurations. For regression endpoints, the identical ensemble of tree-based and gradient-boosted regressors generated out-of-fold cross-validated predictions that were integrated through a Ridge Regression meta-learner. Eight distinct molecular featurization strategies were evaluated, spanning Morgan fingerprints, Avalon fingerprints, Extended Reduced Graph representations, MACCS structural keys, RDKit descriptors, tokenized SELFIES sequence strings, SELFIES term frequency-inverse document frequency vectors, and fixed thirty-two- and sixty-four-dimensional randomly initialized neural projections derived from SELFIES sequences.

Crucially, the two hundred and seventeen RDKit descriptors examined in the study are not purely physicochemical metrics, as eighty-five of these features represent SMARTS-based functional-group and substructure counts, while three represent Morgan fingerprint densities. Consequently, more than forty percent of the descriptor set explicitly captures concrete two-dimensional structural topology alongside surface area, partial charge, and connectivity indices. To evaluate the independent contribution of each feature class beyond a baseline structural core of Morgan, Avalon, and ErG fingerprints, the authors examined forty-eight core factorial combinations using matched pairs. In this matched-pair analysis, RDKit descriptors showed the strongest beneficial marginal effect and MACCS keys produced a smaller apparent improvement, whereas sequence tokenization, term-frequency encodings, and random projections failed to show statistically significant improvements after controlling for multiple testing.

Although the compact two hundred and seventeen feature RDKit representation proved exceptionally competitive on aggregate, the researchers uncovered marked differences in representation demands when analyzing endpoints individually. On eight specific tasks, including Caco-2 permeability, volume of distribution, human intestinal absorption, P-glycoprotein activity, oral bioavailability, blood-brain barrier permeability, AMES mutagenicity, and drug-induced liver injury, the descriptor-only model captured at least ninety-eight point seven percent of the performance achieved by the best observed configuration. In fact, for Caco-2 permeability and steady-state volume of distribution, RDKit descriptors alone yielded the best observed performance among all fifty-four evaluated configurations, indicating that additional representation complexity provided little numerical improvement for these tasks.

The most pronounced example of this endpoint-dependent complexity requirement occurred for human half-life prediction, where the descriptor-only model achieved only fifty-nine point seven percent of the top observed performance, registering a modest Spearman rank correlation of zero point three three one. Expanding the representation to incorporate complementary structural fingerprints and sequence encodings elevated the correlation to zero point five five four, representing an approximate sixty-seven percent relative improvement over the compact baseline. Similarly, descriptor-only models captured only eighty-five point nine percent of peak performance for CYP2C9 substrate classification and ninety point five percent for hepatocyte clearance. Grouped by evaluation metric, mean retention for the descriptor-only configuration reached ninety-eight point two percent across eight receiver-operating characteristic endpoints, ninety-six point zero percent across five mean absolute error regression tasks, ninety-three point two percent across five precision-recall classification tasks, and eighty-five point nine percent across four Spearman rank correlation endpoints.

When benchmarked against four prominent public models anchoring the official leaderboard, including the MiniMol foundation model, the MapLight graph neural network ensemble, the ADMET-AI multitask message-passing framework, and the ADMETrix hybrid, ADMET-Stack secured superior performance in sixty-one out of eighty-seven pairwise evaluations, corresponding to a seventy point one percent win rate. On acute oral toxicity, ADMET-Stack lowered mean absolute error by twenty-nine point seven percent compared to the strongest competitor. In contrast, for classification tasks, no statistically significant performance difference was detected against the leading leaderboard baselines following Holm multiple-comparison adjustments, with the framework exhibiting relative limitations across cytochrome P450 metabolism tasks where it matched or exceeded baseline models in only eleven of twenty-four comparisons.

To make abstract algorithmic outputs directly actionable for medicinal chemists during lead optimization, the investigators mapped raw prediction values into normalized percentile ranks relative to an external reference landscape of one thousand and eighty-eight approved drugs curated from DrugBank. By establishing population-relative rankings, individual liabilities such as predicted cardiac potassium channel blockage or blood-brain barrier penetration can be readily contextualized against the distributions of successful commercial therapeutics. Furthermore, calculating an aggregate risk-favorability score across six critical safety endpoints revealed no significant linear association with quantitative estimate of drug-likeness scores across the reference population, yielding a Pearson correlation of negative zero point zero four two with a ninety-five percent confidence interval spanning negative zero point zero eight six to positive zero point zero zero eight. This weak association suggests that aggregate ADMET risk-favorability scores and general drug-likeness capture partially distinct information at the population level, providing complementary insight alongside conventional drug-likeness assessments.

The authors emphasized several critical methodological limitations that must guide the practical interpretation of their findings. First, the top-performing endpoint configurations were identified using the official test partitions, introducing configuration-selection optimism that was quantitatively confirmed through a split-sample analysis, which revealed a median selection bias of zero point zero three five on smaller datasets. Second, the sequence-derived SELFIES vocabularies and random projection transformations involved transductive feature construction from pooled benchmark partitions, while the random thirty-two- and sixty-four-dimensional neural projections were unpersisted mathematical artifacts rather than trained chemical embeddings. Finally, all reported feature dynamics are strictly conditional on the tree-based stacked ensemble utilized throughout the study, meaning that deep neural networks capable of learning end-to-end representations may exhibit different feature synergies. Together with released per-configuration results and leaderboard records, the study provides an auditable framework demonstrating that compact molecular descriptor sets establish a competitive baseline while clarifying when additional representational complexity is warranted.

Subject of Research: Systematic evaluation of molecular representation complexity in machine learning models for ADMET property prediction

Article Title: ADMET-Stack: quantitative guidance on how much representational complexity an ADMET endpoint actually requires

Article References: Mahmoudi, O., Kumar, S., & Kim, M.-H. (2026). ADMET-Stack: quantitative guidance on how much representational complexity an ADMET endpoint actually requires. Frontiers in Chemistry, 14, 1991510. https://doi.org/10.3389/fchem.2026.1991510

Image Credits: AI Generated

DOI: 10.3389/fchem.2026.1991510

Keywords: ADMET Prediction, Molecular Representations, RDKit Descriptors, Therapeutics Data Commons, Stacked Ensemble, Machine Learning, Pharmacokinetics, Drug Discovery, Structural Fingerprints, SELFIES Encoding, Computational Chemistry, Predictive Modeling, Leaderboard Benchmarking, Feature Ablation, Model Generalization, Drug Toxicity