Supplementary Fig. 1: Linearity of the latent space.
(a)Building a linear classifier based on the dot product between the difference vector δ and the latent representation of each cell. (b) Dot product results between latent representation of all cells with their corresponding difference vector δ for each condition show that two conditions are approximately linearly separable using dot product classifier. (c) Cosine similarity of \({\mathrm{\delta }}_{ {{\mathrm{stim}} - {\mathrm{k}}}},{\mathrm{\delta }}_{{\mathrm{celltype}} - {\mathrm{ij}}}\) with δ where\({\mathrm{\delta }}_{{\mathrm{celltype}} - {\mathrm{ij}}} = avg\left( {z_{celltype = i}} \right) - avg\left( {z_{celltype = j}} \right)\) and δstim-k=avg(zstim,cell type=k)–avg(zctrl,cell type=k) for all seven cell types present in the Kang et al.3 (n=18,868) dataset (z denotes the latent representation of all cells with the corresponding label). First and second violin plot have n=21 and n=7, respectively. The third violin plot shows pairwise (n=499,500) cosine similarity for a set of 1000 random samples from 100–dimensional standard normal distribution. Vertical axis: cosine similarity. Horizontal axis: dot product results in different scenarios as described before.
