Large‐Scale Machine Learning to Screen for Small‐Molecule Senolytics

Senescent cells are promising therapeutic targets given their key roles in cancer and age‐related diseases. Virtual screening of vast chemical spaces for compounds selectively killing senescent cells (senolytics) is a promising approach. However, this requires computational models able to identify senolytics among a much larger proportion of highly diverse non‐senolytic compounds. Here, we investigate the application of machine learning to enable such virtual screening. For the first time, we exploited real‐valued training data via regression and analyzed the impact of detrimental factors such as data inconsistencies, training‐test distribution shifts, and non‐early‐recognition performance metrics to identify senolytics. To this aim, we leveraged two senolytic datasets. The results show that many learning models cannot discriminate between senolytic and non‐senolytic compounds. Furthermore, while graph neural networks lose their efficacy with test molecules chemically different from training molecules, a simple classification and regression tree (CART) algorithm led to a 21% hit rate in this realistic case study. Moreover, as CART is explainable, we discussed the extracted insights into how chemical structures relate to senolytic activity. We make this model available at https://github.com/alexisdougha/virtual‐screening‐senolytics .