Comparison of transcriptomic and phenomic profiles for the prediction of drug mechanism

Abstract Transcriptomic and phenomic profiling assays analyze drug perturbations to provide unbiased information regarding the mechanisms of action (MOAs) of drugs. However, few studies have compared the bioinformatics contents derived from these assays. This study investigated the transcriptomic and phenomic features in terms of diversities and MOA prediction. From publicly available L1000 and Cell Painting datasets, transcriptomic and phenomic features for 274 compounds annotated with 30 MOAs were prepared for analyses. Feature-extraction analyses with tSNE and Isomap algorithms showed that the compound distribution based on transcriptomic features was more dispersed than that based on phenomic features. Pairwise comparison across compounds showed high correlative clusters in phenomic feature heatmap. To explore the predictive potential for the MOA of compounds, transcriptomic and/or phenomic features were used to train machine learning models. XGBoost and Extra Tree models resulted in overfitting, whereas the KNN and Adaboost models yielded a relatively lower performance. Notably, the glucocorticoid receptor agonist was the class of MOA with the highest predictability based on transcriptomic and/or phenomic features. In conclusion, L1000 features were more diverse than the Cell Painting features. Machine learning analysis suggested new similar pairs of compounds and predicted certain classes among MOAs more accurately than others.