Predicting ADMET Properties from Molecule SMILE: A Bottom-Up Approach Using Attention-Based Graph Neural Networks

Abstract

Understanding the pharmacokinetics, safety and efficacy of candidate drugs is crucial for their success. One key aspect is the characterization of absorption, distribution, metabolism, excretion and toxicity (ADMET) properties, which require early assessment in the drug discovery and development process. This study aims to present an innovative approach for predicting ADMET properties using attention-based graph neural networks (GNNs). The model utilizes a graph-based representation of molecules directly derived from Simplified Molecular Input Line Entry System (SMILE) notation. Information is processed sequentially, from substructures to the whole molecule, employing a bottom-up approach. The developed GNN is tested and compared with existing approaches using six benchmark datasets and by encompassing regression (lipophilicity and aqueous solubility) and classification (CYP2C9, CYP2C19, CYP2D6 and CYP3A4 inhibition) tasks. Results show the effectiveness of our model, which bypasses the computationally expensive retrieval and selection of molecular descriptors. This approach provides a valuable tool for high-throughput screening, facilitating early assessment of ADMET properties and enhancing the likelihood of drug success in the development pipeline.

Keywords: ADMET prediction; attention-based architecture; graph neural network; model-based drug development.

Figure 1

Example of how adjacency matrices are extracted from molecular SMILE. For each type of bond (i.e., single, double, triple or aromatic), a specific adjacency matrix is derived in order to focus on molecular substructures.

Figure 2

Schematic representation of the GNN adopted. The architecture is organized as a stack of three main modules, each with a specific function.

Figure 3

Schematic representation of the implemented five-fold cross validation. At each step, one fold (orange) is used as an external test set; the remaining four are used for training and validation. And $20\%$ of the four folds are used as validation data.

Figure 4

Distributions of regression variables in two benchmark datasets. Histograms of Lipophilicity AZ panel (A) and AqSolDB panel (B) data.

Figure 5

Example of the weighting strategy adopted for both regression tasks. Panel (A) shows the weights introduced for training the GNN on LogD prediction. Panel (B) focuses on LogS. For both tasks, $\alpha$ was set to 0.55.

Figure 6

Models used in the ablation study to benchmark the implemented GNN architecture. Panel (A) illustrates the ‘Whole Molecule’ GNN, which does not consider molecular substructures. Panel (B) represents the ‘Convolutional’ GNN, in which the attention mechanism for the entire molecule is replaced by a graph convolutional (GC) layer.

Figure 7

Results of the ablation study on the regression tasks.

Figure 8

Results of the ablation study on the classification tasks.