ConvNet-VAEs integrate single-cell bi-modal epigenomic profiling data from mouse brain. (a) UMAP visualization of cell embeddings from ConvNet-VAEs (left, middle) and FC-VAE (right). (b) For single-Conv1D-layer ConvNet-VAE, more channels in the convolutional layer lead to a larger marginal log-likelihood of the validation set at the cost of longer runtime in training, according to the result from five-fold cross-validation. (c) The number of trainable parameters depends on the number of Conv1D layers and stride. ConvNet-VAEs from Group 1 require fewer parameters than FC-VAEs while those from Group 2 need more parameters. (d) Training time across five-fold cross-validation is reported for each model. (e) Negative marginal log-likelihood of validation set estimated through importance sampling (five-fold cross-validation). A lower value implies a larger marginal log-likelihood. (f) Comparisons between ConvNet-VAEs with single Conv1D layer (Group 1) and FC-VAEs in terms of the quality of cell embeddings (training set: left; validation set: right). The bars show the median number of clusters obtained by the Louvain algorithm from five splits in cross-validation over a range of resolutions. The corresponding average Adjusted Rand Index (ARI) is calculated by comparing the resulting clusters to the published cell type labels, displayed as a line plot. Error bars indicate SD across five-fold cross-validation. (g) Comparisons between ConvNet-VAEs with multiple Conv1D layer (Group 2) and FC-VAEs in terms of cell embeddings’ quality (training set: left; validation set: right), exhibited in the same way as (f).