Extended Data Fig. 3: Additional unsupervised and fine-tuned autoencoder metrics.

(a) Fully unsupervised autoencoder reconstruction losses for training and test sets across 25 epochs. (b) Examples of manually labelled faint, punctate, cytoplasmic, and peripheral input cell images with accompanying unsupervised autoencoder reconstructions. (c) Fine-tuned autoencoder trained using negative log likelihood loss with balanced validation accuracy also reported across 50 epochs of training. (d) Best model train and test set accuracies for the VP35 protein localization prediction task using SVMs on latent embeddings from the unsupervised autoencoder, predefined features, a Resnet-50 architecture trained on the prediction task, or the fine-tuned autoencoder. Predefined features include intensity, correlation, and texture morphological features similar to those previously described for Cell Painting¹⁸. (e) Confusion matrix of model predictions vs manually labelled classifications on model test set. (f) Proportion of cells in each VP35 localization category for non-targeting controls and the genes with the largest proportion of faint (NPC1), punctate (UQCRB), and peripheral (ITGB1) cells. Error bars indicate SEM across sgRNAs targeting the same gene.

Source data