Application of Machine Learning Methods in Predicting Lysine-specific Histone Demethylase 1 (LSD1) Inhibitors

Background Lysine-specific histone demethylase 1 (LSD1) is a well-known anti-cancer target for drug discovery. Novel reversible inhibitors of LSD1 are desirable to be developed. Objective This study aimed to build reliable predictive models to evaluate the antineoplastic efficacy of compounds and reveal the structural foundation underlying the inhibitory activity of LSD1. Methods Multiple artificial intelligence algorithms were utilized in the development of quantitative structure-activity relationship (QSAR) models. A dataset comprising 915 compounds with well-defined IC50 values against LSD1 was assembled for analysis. The structures of these compounds were described by different descriptor packages. Principal component analysis (PCA) was performed to explore the chemical space distribution of each dataset. Y-randomization test and applicability domain (AD) analysis were deployed to validate the reliability of models. Results For regression models, a consensus model was constructed by integrating the predictions of four top-performing individual models (SVM_ECFP4, RFR_PyDescriptor, RFR_RDKIT, and TRANSNNI), resulting in enhanced predictive performance as compared to the individual models. The consensus model achieved a determination coefficient (r2, for the test set) value of 0.82. For classification models, the consensus model was derived from the amalgamation of three individual models (ASNN_RDKIT, ASNN_ECFP4, and RFR_ECFP4), and the overall prediction accuracy of the model was 0.96 for external validation. Conclusion The reliable models could be used to identify highly active LSD1 inhibitors for the purpose of efficient virtual screening. In addition, the important molecular properties and structural fragments derived from this work can provide guidance for the structural optimization of novel LSD1 inhibitors.