Skip to main content
Fig. 1 | Genome Biology

Fig. 1

From: Biologically relevant transfer learning improves transcription factor binding prediction

Fig. 1

Two matrices of TF binding events across accessible genomic regions. The matrices, one more “sparse” for training single-task models and the other “less sparse” for training multi-task models, summarize the binding (from ReMap ChIP-seq peaks and UniBind TFBSs) of 163 TFs to 1,817,918 DHSs across 52 cell and tissue types, with rows representing TFs and columns representing DHSs. For each TF-DHS pair in the matrix, a “1” indicates that, according to both ReMap and UniBind, the TF binds to the DHS, and that the DHS is open (i.e., accessible), a “0,” that the DHS is open but that the TF does not bind to it, and a null sign (“”) indicates that there is insufficient evidence to assign a one or a zero, for instance, if the DHS is closed (i.e., not accessible). Both matrices have the same number of ones; however, they differ in the number of zeros and nulls. In the sparse matrix, TF-DHS pairs wherein the DHS is open and the TF binds to it according to either ReMap or UniBind data (not both) are assigned a null value; instead, in the less sparse matrix, they are assigned a zero value. DHS, DNase I hypersensitive site; TF, transcription factor; TFBS, TF binding site

Back to article page