ADMET evaluation in drug discovery: 21. Application and industrial validation of machine learning algorithms for Caco-2 permeability prediction

Dong Wang, Jieyu Jin, Guqin Shi, Jingxiao Bao, Zheng Wang, Shimeng Li, Peichen Pan, Dan Li, Yu Kang, Tingjun HouJ. Cheminformatics 2025 · DOI

  1. Application and industrial validation of machine learning algorithms for Caco-2 permeability prediction

Introduction

In the marathon of drug discovery, oral bioavailability remains a significant hurdle. If a drug candidate cannot effectively cross the intestinal barrier, it will fail to reach the bloodstream in therapeutic concentrations. The "gold standard" for assessing this property in the laboratory is the Caco-2 cell assay. However, this biological assay is a notorious bottleneck, often requiring up to three weeks to yield results.

A recent study published in the Journal of Cheminformatics by researchers at Zhejiang University and the Shanghai Qilu Pharmaceutical R&D Center addresses this challenge. The team developed a suite of machine learning (ML) models designed to predict Caco-2 permeability and, crucially, validated these models against real-world industrial data to ensure their practical utility in pharmaceutical R&D.

Methods and Modeling Strategy

The researchers conducted a comprehensive comparison of several machine learning architectures. They evaluated four traditional ML algorithms—XGBoost, Random Forest (RF), Gradient Boosting Machine (GBM), and Support Vector Machine (SVM)—alongside two sophisticated deep learning models: a Directed Message-Passing Neural Network (DMPNN) and a hybrid "CombinedNet."

To represent the chemical space, the team utilized three distinct molecular representations:

  • Morgan Fingerprints: Capturing local chemical environments.
  • RDKit2D Descriptors: Representing physical properties such as molecular weight and lipophilicity.
  • Molecular Graphs: Allowing deep learning models to process the chemical structure directly.
The study workflow involves training models on public data and validating them against private industrial datasets using various ML architectures and molecular representations.

To ensure the robustness of their findings, the authors employed Y-randomization to rule out spurious correlations and Applicability Domain (AD) analysis to define the chemical boundaries within which the models remain reliable.

Datasets and the Challenge of Imbalance

The models were trained on a curated "augmented" dataset of 5,654 non-redundant compounds compiled from major public studies. For the critical "stress test," the researchers used an external validation set of 67 compounds from Shanghai Qilu Pharmaceutical’s proprietary collection. This industrial dataset represents a significant challenge, as proprietary chemical spaces often differ from those found in public repositories.

One notable characteristic of the training data is its distribution. As shown in the histogram of $P_{app}$ values, the dataset is significantly unbalanced, with a higher density of compounds exhibiting high permeability compared to those with low permeability. This imbalance is a common challenge in ADMET modeling and can influence model bias.

Distribution of Caco-2 permeability (Papp) values in the dataset, categorized into high, medium, and low permeability classes.

Key Insights and Industrial Validation

The study yielded several critical insights for computational drug discovery:

  1. Boosting Algorithms Outperform Deep Learning: Despite the current trend toward complex deep learning architectures, XGBoost consistently provided the most accurate predictions. In the "small-data" regime typical of drug discovery (thousands rather than millions of data points), human-engineered features combined with boosting algorithms often outperform raw graph-based learning.
  2. The Transferability Paradox: When models were applied to the private industrial data, their absolute accuracy ($R^2$) decreased. However, the Pearson correlation ($r$) remained high (~0.7). This suggests that while models may struggle to predict the exact numerical value of permeability in a novel chemical space, they remain highly effective at ranking compounds, which is the primary goal in lead optimization.
  3. Actionable Chemistry Rules: By using Matched Molecular Pair Analysis (MMPA), the authors identified 24 specific transformation rules. For instance, increasing lipophilicity through the addition of hydrocarbon groups generally enhanced permeability, while the introduction of polar groups tended to decrease it.

Results and Performance

The top-performing model, XGB_combined (utilizing both fingerprints and 2D descriptors), achieved an $R^2$ of 0.647 and an RMSE of 0.494 on the public test set. On the private industrial dataset, it maintained a Pearson correlation of 0.702, confirming its value as a prioritization tool in real-world R&D settings. Furthermore, by applying Applicability Domain filtering, the researchers demonstrated that they could further reduce prediction errors by focusing on molecules chemically similar to the training set.

Conclusion and Future Directions

This research provides a rigorous validation of ML tools for Caco-2 permeability, bridging the gap between academic modeling and industrial application. While the models are currently biased toward high-permeability compounds due to dataset imbalances, they offer immediate value for compound prioritization. Future work will focus on "active learning" strategies to specifically target low-permeability chemical spaces and improve the transferability of models across diverse proprietary datasets.

Reference: Wang et al. "ADMET evaluation in drug discovery:

  1. Application and industrial validation of machine learning algorithms for Caco-2 permeability prediction." Journal of Cheminformatics (2025). https://doi.org/10.1186/s13321-025-00947-z

Summary updated

BibTeX
@article{wang2025admet,
  author = {Dong Wang and Jieyu Jin and Guqin Shi and Jingxiao Bao and Zheng Wang and Shimeng Li and Peichen Pan and Dan Li and Yu Kang and Tingjun Hou},
  title = {{ADMET evaluation in drug discovery: 21. Application and industrial validation of machine learning algorithms for Caco-2 permeability prediction}},
  journal = {J. Cheminformatics},
  year = {2025},
  doi = {10.1186/s13321-025-00947-z},
}
Figure