Explainable t-SNE for single-cell RNA-seq data analysis
Single-cell RNA (scRNA-seq) sequencing technologies trigger the study of individual cell gene expression and reveal the diversity within cell populations. To measure cell-to-cell similarity based on their transcription and gene expression, many dimension reduction methods are employed to retrieve the corresponding low-dimensional embeddings of input scRNA-seq data to conduct clustering. However, the methods lack explainability and may not perform well with scRNA-seq data because they are often migrated from other fields and not customized for high-dimensional sparse scRNA-seq data. In this study, we propose an explainable t-SNE: cell-driven t-SNE (c-TSNE) that fuses the cell differences reflected from biologically meaningful distance metrics for input scRNA-seq data. Our study shows that the proposed method not only enhances the interpretation of the original t-SNE visualization for scRNA-seq data but also demonstrates favorable single cell segregation performance on benchmark datasets compared to the state-of-the-art peers. The robustness analysis shows that the proposed cell-driven t-SNE demonstrates robustness to dropout and noise in dimension reduction and clustering. It provides a novel and practical way to investigate the interpretability of t-SNE in scRNA-seq data analysis. Unlike the general assumption that the explainanbility of a machine learning method needs to compromise with the learning efficiency, the proposed explainable t-SNE improves both clustering efficiency and explainanbility in scRNA-seq analysis. More importantly, our work suggests that widely used t-SNE can be easily misused in the existing scRNA-seq analysis, because its default Euclidean distance can bring biases or meaningless results in cell difference evaluation for high-dimensional sparse scRNA-seq data. To the best of our knowledge, it is the first explainable t-SNE proposed in scRNA-seq analysis and will inspire other explainable machine learning method development in the field.