Poster L14 - Paired Molecular Machine Learning for Drug Metabolism and Prodrug Activation

Introduction: Metabolism is a critical consideration in drug development, as drugs are broken down into metabolites with different properties, activities, and toxicities than their parent drug. In some cases, drugs with no inherent toxic side effects are converted in the body into metabolites with dose-limiting toxicities. Conversely, a class of drugs called prodrugs are inactive derivatives of drugs that rely on metabolism to generate the active pharmaceutical ingredient (API). Thus, profiling drug metabolites is necessary for comprehensive characterization of drug safety and rational development of improved drug delivery solutions. Experimental approaches for profiling drug metabolism exist but remain time consuming and expensive. Further, the thousands of known human metabolic enzymes interact in complex pathways, which impedes understanding and prediction of metabolic outcomes. Machine learning approaches are optimized to resolve relationships in such complex data and represent a valuable tool to improve metabolism profiling. Several computational metabolite prediction approaches have been developed but are limited in breadth of predictable metabolism pathways and applicability across molecule classes, leaving room for improvement in predictive performance. For example, the current best-in-class metabolism predictors exhibit poor performance in predicting prodrug activation, a critical limitation given the recently proven utility of prodrugs and anticipated broader application. To combat these shortcomings, we created a novel, paired molecule machine learning approach capable of predicting the probability of any two input molecules being a parent compound/metabolite pair. Our novel machine learning approach shows significantly improved performance in correctly predicting drug metabolism and prodrug activation compared to the current state-of-the-art methods.

Materials and

Methods: A dataset of 6,680 human endogenous and xenobiotic metabolic reactions with their metabolizing enzyme annotated was curated from DrugBank, Recon3D, MetaCyc, and BioTransformer. This dataset of validated parent/metabolite pairs was split into a training set consisting of 80% of pairs and a test set containing the remaining 20% of pairs, with data split to ensure that the same parent molecule was not present in both training and test sets. The two datasets were then separately augmented with metabolically reasonable negative compound pairs generated by published metabolic reaction rules to create negative data while preventing data leakage. All generated negative pairs that existed in the positive datasets were discarded, and all negative instances were randomly sampled to create balanced final training and test datasets with equal numbers of positive and negative pairs.

The balanced training dataset was then used to train several traditional (random forest, gradient boosting, XGBoost, logistic regression, and support vector machines) and deep (message passing neural network, feed forward neural networks, and graph neural network) machine learning models to predict the probability of each input compound as a true parent/metabolite pair. Models were evaluated on the test set and additionally on an external validation set of FDA-approved prodrugs and their API metabolites. Performance was quantified using accuracy, ROC-AUC, F1, MCC, and FPR. Test and prodrug set performance was benchmarked against existing metabolism predictors (SyGMa, GLORYx, BioTransformer, and MetaTrans) and statistical analysis was used to evaluate if significant improvement in ability to predict metabolism and prodrug activation was achieved.

Results, Conclusions, and Discussions: We trained all traditional and deep learning models on 3 replicates of the training dataset, where the randomly selected negative data varied between replicates. Each model was then applied to classify molecule pairs in the test set and accuracy, ROC-AUC, F1 score, Matthew’s correlation coefficient, and false positive rate (FPR) were calculated. The random forest, XGBoost, MPNN, and GNN exhibited superior performance across all metrics on the test set (Figure 1A). Each model was then applied to classify FDA-approved prodrug/API pairs and generated negative data for each prodrug, revealing superior performance of the random forest, XGBoost, and GNN models across all metrics. These three models were benchmarked against existing computational metabolism models’ predictions of prodrug activation, which revealed significant improvements in accuracy and FPR (Figure 1B-C). The predictions made by our best model (XGBoost: accuracy = 0.86, FPR = 0.13) were further analyzed to determine whether performance varied between prodrug activation types (Type 1: prodrugs require atom removal for API formation; Type 2: prodrugs require additions for API formation; Type 3: prodrugs require rearrangement/replacement for API formation). Comparison of prodrug predictions made by our model to existing metabolism models revealed significant improvement in ability to predict both Type 2 and Type 3 prodrug activations (Figure 1D).

In summary, we here conceived and developed a novel computational approach for drug metabolism prediction by rigorous evaluation of machine learning models trained to classify molecule pairs as parent/metabolite structures. We demonstrate that our best model significantly improves accuracy and reduces FPR on a test set of endogenous and xenobiotic metabolism pairs in comparison to current state-of-the-art metabolism predictors. We also show improved performance in predicting prodrug activation and identify specific classes of prodrug activations captured by our model that were previously outside the scope of existing metabolism predictors. Ultimately, our model represents a useful tool for computational metabolism interrogation of any compound of interest, with specific utility anticipated for rational prodrug design. The insights generated by this model can contribute to a computationally derived profile of drug safety and efficacy to accelerate early-stage drug development and delivery.

Acknowledgements (Optional):