Fig. 1

The Avocado deep tensor factorization approach. a A collection of epigenomic data can be visualized as a 3D tensor (blue), in which some experiments (white cells) have not yet been performed. Avocado models the tensor along three orthogonal axes, learning latent factors (gray) that represent the cell types (in orange, with 32 factors each), the assay types (in purple, with 256 factors each), and the genomic axis (in red, with 25, 40, and 45 factors at each of the three resolutions). b During the training process, the respective slices from these three axes corresponding to the location of the training sample in the tensor are concatenated together and fed into a neural network comprised of two hidden dense layers each with 2048 neurons to produce the final prediction (in green)