Figure 1

NMF-based algorithms for epigenomic data. (A) Schematic representation of the transformation of ChIP-seq data. (top) transformation of read counts into elements of the V matrix in the basic "absolute" mode. Each mark - locus combination is a single element in matrix V. Reads at each locus are summed. Columns of V are additionally scaled. (bottom) transformation of read counts from paired samples in "differential" mode. The differential signal is obtained by subtracting sample A coverage from sample B coverage after correction for sequencing-depth. Positive and negative area under curve is summed (integrated) into "gain" and "loss" scores. (B) NMF factorization in the "absolute" algorithm. V matrix is the same as the top of sub-panel (A). (C) NMF factorization in the "discriminatory" algorithm. contains two classes of loci V₁ and V₂, which are used to derive two independent basis pattern matrices H₁ and H₂. All codes are concatenated and used to derive a single matrix W₁₂. (D) NMF factorization in the "differential" algorithm. The V matrix corresponds to the bottom of sub-panel (A) and contains columns for the "gain" and "loss" of each mark.