Extended Data Fig. 1: Comparison of reconstruction loss measures.

The suitability of different reconstruction loss measures was assessed by fitting thirty individual models on the Buenrostro et al. 2018 dataset. The total loss across the dataset was determined for each model and models with poor outlier losses were excluded (for example due to poor local minima; see Methods), leading to 28, 29 and 29 models for binary, multinomial and negative multinomial loss for the visualization, respectively. The x-axis represents different reconstruction losses: binary cross-entropy loss (Binary), negative log-likelihood for the multinomial distribution (Multinomial), and negative log-likelihood of the negative multinomial (Neg. multinomial) distribution. Otherwise, the model architecture remained the same. Latent features were subjected to clustering using k-means, hierarchical clustering and Louvain clustering and clustering performances were computed based on adjusted mutual information (AMI), adjusted Rand index (ARI) and Homogeneity (Hom) against ground truth cell labels. The best score across the clustering algorithms was considered. Boxes represent quartiles Q1 (25% quantile), Q2 (median) and Q3 (75% quantile); whiskers comprise data points that are within 1.5 x IQR (inter-quartile region) of the boxes.