Fig. 1: Schematic representation of the eDICE model.

a For each cell type, we collect all measured signal values from assays performed in that cell type at the target bin, and project this set of values into a shared embedding space, where it is combined with a global embedding representing the cell type (1). We do likewise for assays, projecting the sets of values measured in different cell types from each assay into a distinct embedding space. We then apply self-attention over both sets of embeddings, allowing the network to capture relationships between cell types and assays to produce `contextualized' latent embeddings which are functions of the local signal values in all observed tracks (2). Finally, a feed-forward neural network combines the contextual embeddings for a target cell type-assay combination to generate a prediction for the local signal value (3). b Transfer learning scheme for the imputation of unseen tissues in the EN-TEx dataset.