PredHLM: quantitative and interpretable prediction of metabolic half-life in human liver microsomes

Technical Summary: PredHLM – Quantitative and Interpretable Prediction of Metabolic Half-Life in Human Liver Microsomes

Problem Statement
Metabolic stability in human liver microsomes (HLM) is a critical pharmacokinetic determinant influencing systemic exposure, clearance, and oral bioavailability. While experimental HLM assays are the gold standard for evaluating Phase I metabolism, they are labor-intensive and resource-heavy, creating bottlenecks for large-scale virtual screening in early drug discovery. Existing computational models often suffer from three limitations: (1) they are predominantly classification-based (stable vs. unstable), lacking the granularity required for lead optimization; (2) they struggle with the inherent experimental noise and heterogeneity of public datasets, leading many researchers to avoid quantitative regression; and (3) they often function as "black boxes," failing to provide chemical insights that align with known metabolic soft-spots. Consequently, direct, quantitative, and interpretable prediction of continuous HLM half-life ($T_{1/2}$) remains underexplored.

Methodology
The authors developed PredHLM, a quantitative machine learning framework designed to predict HLM half-life directly.

• Data Curation: The study integrated a dataset of 11,790 unique compounds by combining in-house data from the Korea Chemical Bank (KCB) with curated public data from ChEMBL (v36) and PubChem.
  • Experimental endpoints included measured half-life and percent remaining at 30 minutes.
  • Percent remaining values were converted to half-life using a first-order kinetic equation.
  • Rigorous quality control was applied: non-HLM systems (e.g., hepatocytes, S9 fractions) were excluded, censored data was handled with specific bounds, and duplicate compounds were averaged.
  • Data were log10-transformed to reduce skewness, and outliers were removed using the interquartile range method.
• Featurization: Four molecular representation schemes were evaluated: ECFP4 fingerprints (2,048-bit), RDKit 2D descriptors, Mordred descriptors, and MACCS keys.
• Model Architectures: Nine predictive algorithms were benchmarked, including multilayer perceptron (MLP), k-nearest neighbors (k-NN), kernel ridge regression (KRR), support vector machines (SVM), and tree-based methods (Random Forest, LightGBM, XGBoost, CatBoost). Additionally, a Communicative Message Passing Neural Network (CMPNN) and its variants (with attention mechanisms and feature concatenation) were tested.
• Training and Evaluation:
  • The dataset was split 9:1 into training and independent test sets.
  • Hyperparameters were optimized via 5-fold cross-validation using Optuna.
  • A scaffold-based split (Bemis–Murcko frameworks) was employed to assess generalization to structurally novel chemotypes, ensuring no scaffold overlap between training and test sets.
  • Applicability Domain (AD): Prediction reliability was assessed using a Sum of Distance-weighted Contribution (SDC) score based on Tanimoto similarity to the training set.
  • Interpretability: Shapley Additive Explanations (SHAP) were used to quantify feature contributions and elucidate the model's decision-making process.

Key Results

• Model Performance: The XGBoost model using RDKit 2D descriptors achieved the best predictive performance.
  • On the independent test set, it yielded an RMSE of 0.507 and an $R^2$ of 0.431.
  • In cross-validation, it achieved an RMSE of 0.515 and $R^2$ of 0.410.
  • While graph-based models (CMPNN) performed competitively, they did not outperform the descriptor-based XGBoost model. The authors retained RDKit 2D descriptors for their parsimony (~800 fewer descriptors than Mordred) and interpretability.
• Generalization: Under the stringent scaffold-based split, the model maintained robust performance (independent test RMSE of 0.536), indicating it generalizes to novel chemotypes rather than relying on analog memorization.
• Applicability Domain: Analysis showed a strong correlation between the SDC score and prediction error. Compounds with high SDC scores (well-represented in the training space) exhibited lower RMSE, with the top 8.9% of compounds achieving an $R^2$ of 0.683.
• External Benchmarking: When converted to a binary classification task (stable if $T_{1/2} > 30$ min), PredHLM outperformed reference models (Ryu et al. and Shah et al.) in accuracy (0.675), MCC (0.361), and ROC-AUC (0.736) on an external set of 969 compounds.
• Interpretability (SHAP):
  • Lipophilicity (MolLogP) was the most influential feature, with higher lipophilicity negatively impacting predicted half-life (consistent with increased CYP-mediated oxidation).
  • Metabolic Soft-spots: Descriptors for aryl methyl groups, allylic oxidation sites, and esters showed negative contributions (reducing stability).
  • Stabilizing Features: Halogen counts and carboxylic acid groups showed positive contributions, aligning with strategies to block metabolic sites and the reduced susceptibility of acids to Phase I oxidation in the absence of Phase II cofactors.

Significance and Claims
The paper claims that PredHLM provides a practical framework for defining metabolic stability margins, supporting both rapid Go/No-go decisions and finer compound prioritization during lead optimization. By offering quantitative (continuous half-life) rather than categorical predictions, the model preserves stability information that classification models discard. Furthermore, the integration of SHAP analysis ensures the model is interpretable, aligning its predictions with established medicinal chemistry principles regarding lipophilicity and metabolic soft-spots.

The authors explicitly position PredHLM as a decision-support tool rather than a replacement for experimental assays. They acknowledge limitations, including the model's reliance on Phase I data (excluding Phase II, transporters, and in vivo effects), the uncertainty introduced by converting percent remaining to half-life, and the inherent variability of experimental data. The source code, data, and trained model are made publicly available to support reproducibility and further development.