Machine learning approaches infer vitamin D signaling: Critical impact of vitamin D receptor binding within...

The vitamin D-modulated transcriptome of highly responsive human cells, such as THP-1 monocytes, comprises more than 500 genes, half of which are primary targets. Recently, we proposed a chromatin model of vitamin D signaling demonstrating that nearly all vitamin D target genes are located within vitamin D-modulated topologically associated domains (TADs). This model is based on genome-wide binding patterns of the vitamin D receptor (VDR), the pioneer transcription factor PU.1, the chromatin organizer CTCF and histone markers of active promoter regions (H3K4me3) and active chromatin (H3K27ac). In addition, time-dependent data on accessible chromatin and mRNA expression are implemented. For the interrogation and in deep inspection of these high-dimensional datasets unsupervised and supervised machine learning algorithms were applied. Unsupervised methods, such as the vector quantization tool K-means and the dimensionality reduction algorithm self-organizing map, generated descriptions of how attributes, such as VDR binding and chromatin accessibility, affect each other as a function of time and/or co-localization within the same genomic region. Supervised algorithms, such as random forests, allowed the data to be classified into pre-existing categories like persistent (i.e. constant) and time-dependent (i.e. transient) VDR binding sites. The relative amounts of these VDR categories in TADs showed to be the main discriminator for sorting the latter into five classes carrying vitamin D target genes involved in distinct biological processes. In conclusion, via the application of machine learning methods we identified the spatio-temporal VDR binding pattern in TADs as the most critical attribute for specific regulation of vitamin D target genes and the segregation of vitamin D's physiologic function.