On the robustness of mixture model-based unsupervised learning in single-cell analyses

High-dimensionality of single-cell RNA sequencing (scRNA-Seq) data needs methods or heuristics to reduce the feature space (genes) prior to using as inputs for machine learning methods to analyze the data. Using an unsupervised learning approach, mixture-model based single cell analyses (MiMoSA) were proposed to infer single-cell subpopulations induced after drug treatment. In this method, a threshold was used to select genes among those that exhibited the highest variance with the assumption that the expression variability was introduced by the drug. In this work, through an expanded range of the thresholds, we study the effects of the varying feature space on the original clustering pattern identified by MiMoSA, and then observe if the set genes with the most differential expression were different. In the particular case of using single-cells treated with anti-diabetic drug metformin, our results demonstrate the robustness of MiMoSA wherein neither the default clustering pattern of cells post drug treatment nor the set of originally identified differentially expressed genes changed. This result demonstrates the ability of MiMoSA to reliably identify the most differentially expressed genes which indeed have shown to establish anti-cancer mechanisms of metformin in triple-negative breast cancer, for which there are still no targeted therapeutics.