Enhancing CYP3A4 Inhibition Prediction Using a Hybrid GNN–ML Model with Data Augmentation

Somin Woo, Ju-Hyeok Jeon, Sangil Han, C. Justin Lee, Sang Hyun MinPharmaceuticals 2026 · DOI

TL;DR

This paper introduces a hybrid AI framework combining Graph Neural Networks (GNNs) and traditional Machine Learning (ML) models, enhanced with data augmentation techniques, to accurately predict CYP3A4 inhibition. This approach significantly improves the prediction of drug-drug interaction risks, offering a more robust and interpretable tool for early-stage drug development.

What Problem Does This Paper Address?

Cytochrome P450 3A4 (CYP3A4) is a crucial enzyme responsible for metabolizing a large portion of clinically used drugs. Its inhibition can lead to dangerous drug-drug interactions (DDIs) and toxicity, making accurate prediction of this inhibition a critical step in drug development. Traditional methods like in vitro assays are costly and time-consuming, while existing computational models often struggle with generalization, especially for novel chemical structures, and lack sufficient interpretability to understand why a compound is predicted to be an inhibitor. This paper aims to address these limitations by developing a more accurate, robust, and interpretable predictive model for CYP3A4 inhibition.

Methods

The researchers developed a hybrid ensemble model that integrates two main types of AI approaches:

  • Graph Neural Networks (GNNs): They utilized an O-GNN model, enhanced with contrastive learning and manifold mixup. This approach treats molecules as graphs, allowing the model to learn intricate structural relationships. Contrastive learning helps the model learn robust representations by distinguishing between similar and dissimilar molecular pairs, while manifold mixup generates synthetic data by interpolating between existing molecular representations in the latent space, improving generalization.
  • Machine Learning (ML) Models: They employed traditional vector-based ML models, including LightGBM, XGBoost, and CatBoost, along with a weighted ensemble of these models. These models use molecular descriptors and fingerprints as input.

To further boost performance, two data augmentation strategies were employed:

  • Manifold Mixup: Applied during GNN training to create synthetic training samples by interpolating between similar molecular pairs in the latent space.
  • SMILES Enumeration-based Test-Time Augmentation (TTA): During prediction, a molecule's SMILES string was varied to create multiple representations, and the model's predictions for these variations were averaged to produce a more stable and generalized final prediction.

Finally, the best-performing GNN and ML models were combined using a weighted ensemble strategy to leverage their complementary strengths. Model interpretability was achieved using SHAP analysis for ML models and occlusion sensitivity analysis for the GNN.

Datasets

The study utilized a large, curated dataset for training and evaluation:

  • Training Dataset: A combined dataset of 23,713 compounds with CYP3A4 inhibition data. This was compiled from 1681 pairs from the Korea Chemical Bank (KCB) and an additional 33,612 records from public databases like ChEMBL and PubChem, after rigorous data cleaning and aggregation.
  • External Validation Dataset: A set of 100 newly generated experimental data points, distinct from the training set, to assess the model's generalizability.

Key Insights and Ideas

  • Hybridization is Key: Combining GNNs (which capture local structural details) with ML models (which leverage global descriptors) significantly outperforms either approach alone, reducing prediction errors and improving robustness.
  • Data Augmentation Matters: Both manifold mixup and SMILES TTA were crucial for enhancing the model's ability to generalize to unseen chemical structures and experimental conditions.
  • Interpretability is Achievable: Techniques like SHAP and occlusion sensitivity analysis successfully identified key molecular substructures (e.g., benzodioxole, imidazole) that are strongly associated with CYP3A4 inhibition, providing valuable insights for medicinal chemists.
  • Physicochemical Properties Still Reign: Despite the advanced deep learning architectures, features like lipophilicity (SLogP), molecular weight, and polarizability remained highly influential predictors, aligning with known CYP3A4 active site characteristics.

Key Results

  • The best-performing ML model (weighted ensemble) achieved an RMSE of 19.1031 and a PCC of 0.7566.
  • The best-performing GNN model (O-GNN + CL + Mixup) achieved an RMSE of 20.1002 and a PCC of 0.7265.
  • The hybrid ensemble model demonstrated superior performance with an RMSE of 19.0784 and a PCC of 0.7570 on the internal validation set.
  • On the independent external validation set, the hybrid model achieved a Custom Metric score of 0.8035, confirming its strong generalizability.
  • Substructure analysis identified benzodioxole as a key pharmacophore for CYP3A4 inhibition, showing a significant positive correlation with activity.

Limitations and Future Work

  • Single Enzyme Focus: The current model is specific to CYP3A4. Future work aims to extend the framework to predict inhibition for multiple CYP isoforms.
  • Heterogeneous Data: The model was trained on % inhibition data from varying experimental conditions. Future work will involve prospective validation and potentially incorporating protein structure-aware representations.
  • Generalizability to Other Endpoints: The authors envision leveraging this framework as a backbone for predicting other ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) properties and related applications.

Summary updated

Abstract

<b>Background/Objectives</b>: Cytochrome P450 3A4 (CYP3A4) metabolizes approximately 30-50% of clinically used drugs; thus, accurate prediction of CYP3A4 inhibition is essential for early assessment of drug-drug interaction (DDI) risk and toxicity. This study evaluated an integrated artificial intelligence framework for predicting CYP3A4 inhibition (%) using a large, curated chemical dataset. <b>Methods</b>: A dataset of 23,713 compounds was compiled from the Korea Chemical Bank and multiple commercial and public databases. Vector-based machine learning (ML) models (LightGBM, XGBoost, CatBoost, and a weighted ML ensemble) and graph neural network (GNN) models (O-GNN with contrastive learning and manifold mixup (O-GNN + CL + Mixup), D-MPNN, GINE, and GATv2) were evaluated. Manifold mixup was applied during GNN training, and SMILES enumeration-based test-time augmentation was used at inference. The best-performing ML and GNN models were integrated using a weighted ensemble strategy. Model interpretability was examined using SHAP analysis for ML models and occlusion sensitivity analysis for O-GNN + CL + Mixup. <b>Results</b>: The weighted ML ensemble achieved the best performance among ML models (RMSE = 19.1031, Pearson correlation coefficient (PCC) = 0.7566); the O-GNN + CL + Mixup model performed the best among GNN models (RMSE = 20.1002, PCC = 0.7265). The hybrid model achieved improved predictive accuracy (RMSE = 19.0784, PCC = 0.7570). External validation on 100 newly generated experimental data points confirmed generalizability (Custom Metric = 0.8035). <b>Conclusions</b>: This study demonstrated that integrating ML and GNN models with data augmentation strategies improves the robustness and interpretability of CYP3A4 inhibition prediction and established a practical framework for metabolic screening and DDI risk assessment.
BibTeX
@inproceedings{woo2026enhancing,
  author = {Somin Woo and Ju-Hyeok Jeon and Sangil Han and C. Justin Lee and Sang Hyun Min},
  title = {{Enhancing CYP3A4 Inhibition Prediction Using a Hybrid GNN–ML Model with Data Augmentation}},
  booktitle = {Pharmaceuticals},
  year = {2026},
  doi = {10.3390/ph19020258},
}
Figure