Fig. 1: Model architecture and basic performance metrics.

a Framework architecture main variations: I) AutoTransOP v1: One global space is constructed by mapping omic profiles in a space where the distance between embeddings coming from the same perturbation is minimized. II) AutoTransOP v2: Architecture combined with the CPA approach, where the latent space is separated into two, one global devoid of species/cell effect and a composed latent space. d_i signifies “drug perturbation i” and the illustrated vector corresponds to the vector of drug-induced gene expression values. b Model performance in reconstructing and translating gene expression profiles between the two cell lines with the most common perturbations in the L1000 dataset, A375 and HT29, by using only the 978 measured landmark genes. AutoTransOP v3 is the one with a classifier simultaneously trained in one global latent space. For DCS modified v1–v2, see the corresponding methods sections. It is worth noting that DCS modified v2 has a distance term and a direct translation term in its training loss. c Model performance in reconstructing and translating gene expression profiles between A375 and HT29 by using all 10,086 genes that are either measured or belong to those that are well-inferred computationally. d Performance in inferring transcription factor activity by using the translated/predicted gene expression. e Performance in correctly classifying cell lines in different cases. Reported values are the mean ± standard error (SE). f Performance by using different inputs in the L1000. For all comparisons in this figure, a two-sided Wilcoxon test was used with n = 10 per group. The error bars in the bar plots (b, c) denote 95% Confidence Intervals (CI). In all boxplots, the centerline denotes the median, the bounds of the box denote the 1st and 3rd quantiles, and the whiskers denote points not being further from the median than 1.5 × interquartile range (IQR).