Fig. 3: MorphDiff captures morphological changes across thousands of drug treatments.

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. Linear normalization converted FID and CMMD reciprocals, plus three other metrics, to 0–1 range. Statistical comparisons used one-sided Wilcoxon signed-rank tests with Bonferroni correction (p value < 0.05). “*” indicates MorphDiff(I2I) superiority; “ns” indicates non-significance; hashtag (“#”) indicates the baseline performs better significantly. Data are presented as mean values ± SD. b The R² score between the ground-truth CellProfiler feature vectors and the generated CellProfiler feature vectors on the CDRP ID set. The x-axis represents the R² scores between IMPA and ground truth for each sample. The y-axis represents the R² scores of MorphDiff(G2I) and MorphDiff(I2I) against ground truth respectively. The p values calculated by one-sided Wilcoxon signed-rank test indicate the significance of the distribution of the y-axis being greater than that of the x-axis. c The difference between control and perturbations for specific CellProfiler features on the CDRP ID set. The x-axis displays the difference between the ground-truth perturbations and the control, while the y-axis displays the difference between the generated perturbations and the control. The higher the point density along the diagonal, the closer the generated difference is to the ground truth. More results can be found in the Supplementary Fig. 14b, c. The R² score measuring the similarity between the predicted difference and ground-truth difference. d The x-axis represents the ground-truth classification accuracy scores between different pairs of perturbations, and the y-axis means the classification accuracy scores of the corresponding samples generated by different methods. Source data are provided as a Source data file for (a–d).