AFX-221

This study evaluated the use of machine learning to leverage drug absorption, distribution, metabolism and excretion (ADME) data together with physicochemical and pharmacological data to develop a novel anticholinergic burden scale and compare its performance to previously published scales.

Experimental and in silico ADME, physicochemical and pharmacological data were collected for antimuscarinic activity, blood–brain barrier penetration, bioavailability, chemical structure and P‐glycoprotein (P‐gp) substrate profile. These five drug properties were used to train an unsupervised model to assign anticholinergic burden scores to drugs. The model performance was evaluated through 10‐fold cross‐validation and compared with the clinical Anticholinergic Cognitive Burden (ACB) scale and nonclinical Anticholinergic Toxicity Scores (ATS) scale, which is based primarily on muscarinic binding affinity.

In silico software (ADMET Predictor) used for screening drugs for their blood–brain barrier (BBB) penetration correctly identified some drugs that do not cross the BBB. The mean area under the curve for the unsupervised and ACB scale based on the five selected variables was 0.76 and 0.64, respectively. The unsupervised model agreed with the ACB scale on the classification of more than half of the drugs (49 of 88) agreed on the classification of less than half the drugs in the ATS scale (12 of 25).

Our findings suggest that the commonly used ACB scale may misclassify certain drugs due to their inability to cross the BBB. By contrast, the ATS scale would misclassify drugs solely depending on muscarinic binding affinity without considering other drug properties. Machine learning models can be trained on these features to build classification models that are easy to update and have greater generalizability.

Drug-induced liver injury (DILI) is the main cause of drug failure in clinical trials. The characterization of toxic compounds in terms of chemical structure is important because compounds can be metabolized to toxic substances in the liver. Traditional machine learning approaches have had limited success in predicting DILI, and emerging deep graph neural network (GNN) models are yet powerful enough to predict DILI. In this study, we developed a completely different approach, supervised subgraph mining (SSM), a strategy to mine explicit subgraph features by iteratively updating individual graph transitions to maximize DILI fidelity. Our method outperformed previous methods including state-of-the-art GNN tools in classifying DILI on two different datasets: DILIst and TDC-benchmark. We also combined the subgraph features by using SMARTS-based frequent structural pattern matching and associated them with drugs' ATC code.