Fig. 4: Pre-trained MorphDiff potentially promotes drug discovery.

a General generative evaluation of pre-trained MorphDiff and IMPA. Methods requiring reference control images used 10 distinct control image groups from independent plates (n = 10 biological replicates), while other methods used 10 sampling iterations with different random seeds. Statistical comparisons used one-sided Wilcoxon signed-rank tests with Bonferroni correction (p value < 0.05). b F1 scores measure performance in identifying whether more than 200 CellProfiler features undergo significant changes under drug perturbations across various targets (n = 10 biological replicates). The tests are one-sided Wilcoxon signed-rank tests with Bonferroni correction (p value < 0.05). For a, b, “*” indicates significance; “ns” indicates non-significance; hashtag (“#”) indicates the situation that the baseline performs better significantly. Data presented as mean values ± SD. c Wasserstein distance between different pairs of target-level perturbations for the ground-truth and generated cell morphology, computed with DeepProfiler embeddings. For each pair of target-level perturbations, the ground-truth Wasserstein distance is the x-axis, and the corresponding generated Wasserstein distance is the y-axis. d Schematic overview of retrieval workflow. Created in BioRender. Group, A. (2025) https://BioRender.com/492h3v7. e Top 5 MOA retrieval results across modalities. Sixty-nine drugs were randomly divided into reference (47) and query (22) sets across 10 iterations. Boxplots show median, quartiles, and range. The whiskers in the boxplots mark the minima to maxima. f MOA matching performance using generated cell morphology DeepProfiler embeddings: Mean Average Precision (x-axis) versus Folds of Enrichment (y-axis). Points represent means across all query morphologies using 0.90 threshold (top 10%). g UMAP projections by drug MOAs across modalities: drug structure embeddings, gene expression, and DeepProfiler embeddings from ground-truth and generated cell morphology. BML-259, RG-14620, and BRD-A72066420 structures shown in Supplementary Fig. 23. Source data are provided as a Source data file for (a–g).