ML in Caco-2 Apparent Permeability Prediction

Oral dosing is the most common way drugs are administered, and understanding ADMET properties (absorption, distribution, metabolism, excretion, and toxicity) is essential to ensure a drug is effective and safe. Poor absorption can cause drug development to fail even if the drug works well on its target. Orally administered drugs are mainly absorbed in the stomach and small intestine, with a small amount absorbed in the large intestine. Since drug absorption cannot easily be measured in living systems during early development, in vitro methods are used to estimate it. One common approach uses Caco-2 cell cultures to model the intestinal barrier and assess drug permeability, which correlates with intestinal absorption. However, this method has a major limitation because it requires 21–24 days for the cells to form a suitable monolayer.

As an alternative, quantitative structure–property relationship (QSPR) models offer a faster and more cost-effective way to predict drug permeability. Advances in machine learning (ML) have enabled researchers to build predictive models using various algorithms such as multiple linear regression, support vector machines, random forests, and neural networks. Some studies have also used deep learning methods, including Long Short-Term Memory (LSTM) and graph convolution networks (GCN), to predict Caco-2 permeability.

Molecules can be modeled as multi-relational graphs, where atoms act as nodes and chemical bonds act as edges. Each bond type—single, double, triple, or aromatic—represents a distinct relationship between atoms. These relationships are encoded using a three-dimensional adjacency tensor, which captures whether a specific bond type exists between pairs of atoms. Alongside this, a degree tensor summarizes connectivity by counting how many bonds of each type are associated with each atom. Together, these tensors provide a structured representation of molecular topology and interactions. In addition, each atom is described by a set of numerical features that capture its chemical properties. A multi-relational graph convolutional network (MRGCN) processes these inputs to generate a compact, 500-dimensional representation of the molecule. This approach allows the model to simultaneously capture both local atomic details and broader structural relationships within the compound.

The initial node features used to describe atoms included a variety of atom-level descriptors derived from computational chemistry tools and literature sources. These descriptors encompassed both numerical values, such as atomic weight, valency, and partial charge, and categorical attributes like hybridization state or aromaticity. Additional features, including atomic contributions to log P and physicochemical properties such as van der Waals radius, electronegativity, and polarizability, were incorporated to enrich the representation. Categorical variables were converted into numerical form using one-hot encoding, ensuring compatibility with machine learning models. Altogether, the feature set aimed to capture diverse chemical characteristics at the atomic level. However, multiple combinations of these descriptors were tested to determine which contributed most effectively to model performance, highlighting the importance of careful feature selection in graph-based molecular modelling.

Interestingly, the optimal feature set contained only 17 numerical descriptors, excluding all categorical features. This result suggested that the structural information encoded in the adjacency and degree tensors already captured patterns typically represented by categorical attributes. The final node feature matrix, therefore, consisted of 17-dimensional vectors for each atom. Using these features, a seven-layer MRGCN was trained to refine atom-level embeddings, producing high-dimensional representations that encode both atomic properties and relational context. The model’s parameters were tuned so that each node embedding had a dimensionality of 500. To generate a single, fixed-size representation for an entire molecule, the model applied average pooling across all atom embeddings. This process aggregated individual atomic information into a unified molecular vector, suitable for downstream tasks such as prediction or classification.