Machine learning analysis of Drosophila testis transcriptomic data reveals potential regulatory sequences

The number of accessible transcriptomic data is increasing rapidly due to advances in affordable sequencing technologies. There is also an improvement in the quality and resolution of gene expression maps, which was driven by the advent of single-cell technologies.

Here we present a method, where we integrate transcriptomic data from five different sources, including segmented, single cyst, and single-cell transcriptomic data from Drosophila testis to investigate the expression and accumulation of transcripts in this tissue. The analysis showed that the testis-specific genes have a characteristic profile in the testis, which is predictable using supervised machine learning algorithms (XGBoost). Moreover, dimension reduction with an unsupervised machine learning algorithm (t-SNE) followed by clustering (DBSCAN) of the genes revealed potential regulatory motifs shared by genes in the same group. This approach links the robustness of tissue expression data with the resolution of the single-cell techniques, while masking the weakness of each technique.

The presented approach can be used to find similarly expressed genes and shared regulatory elements or new cell-specific transcripts that could have potential annotation benefits in further research.