Extended Data Fig. 6: Benchmarking of the neural network classifier.

a, b, t-SNE visualization of the P70 optic lobe single-cell transcriptomes, using 120 principal components calculated on the log-normalized integrated gene expression. Cells colours indicate the clusters they belonged to according to unsupervised clustering (a), or the adult clusters they were classified as by the neural network (b, same as in Fig. 1e). Black circles indicate high granularity regions, where less frequent cell types were grouped together by unsupervised clustering but could be resolved accurately by the neural network. c, Same as in a, b but cells are named and coloured by the adult cluster they were classified as by Seurat label transfer (Methods). d, t-SNE visualization (same as c) including only the cells that were assigned inconsistent identities by Seurat and the neural network. Highest rates of inconsistencies were observed in the centre (LQ cells), in L1 and L2 clusters (red ellipses), in most glia clusters (green ellipses), the TE neurons and a glia-like cluster (identity 214, Supplementary Table 1) with no adult correspondence (blue ellipses). e, f, t-SNE visualizations of 56,902 cells sequenced from whole fly brains¹⁹, using 120 principal components calculated on the log-normalized gene expression. e, Cells are named and coloured by the clusters they were classified as by our neural network. f, Cells are named by the cluster identities from the original study and coloured by the confidence score they received from our neural network. Black circles mark the following central brain clusters (from left to right): Poxn, OPN, clock neurons and dopaminergic neurons, that all received low scores from the neural network. Kenyon cells (red circles) were assigned with high confidence as our adult dataset was contaminated by them (cluster 112).