Prediction of cellular morphology change under perturbations with transcriptome-guided diffusion model

Investigating the cell morphology change after perturbations with high-throughput image-based profiling is of growing interest, considering its wide applications in phenotypic drug discovery, including MOA (Mechanism Of Action) prediction, compound bioactivity prediction, and drug repurposing. However, the vast space of chemical and genetic perturbations makes it infeasible to fully explore all the potential perturbations with image-profiling technologies. Consequently, developing a powerful in-silico method to simulate high-fidelity cell morphological response under perturbations can reduce the experiment costs and accelerate drug discovery. Motivated by this, we proposed MorphDiff, a transcriptome-guided latent diffusion model for accurately predicting the cell morphology response to perturbations. We applied MorphDiff to two large-scale datasets, including one drug perturbation and one genetic perturbation cell morphology dataset covering thousands of diverse perturbations. Extensive benchmarking and comparison with baseline methods show the remarkable accuracy and fidelity of MorphDiff in predicting cell morphological changes under unseen perturbations. Furthermore, we explored the utilities of MorphDiff in identifying and retrieving the MOAs of drugs, which is a crucial application in phenotypic drug discovery. With the designed pipeline for MOA retrieval, we demonstrated MorphDiff’s capability to boost the retrieval of the drugs’ MOAs (Mechanism Of Actions) by generating realistic cell morphology profiles. The average MOA retrieval accuracy of MorphDiff-generated morphology is comparable with that of the ground truth cell morphology, and consistently outperforms the baseline method and gene expressionbased retrieval by 29.1% and 9.7% respectively. We also validated that complementary information provided by cell morphology generated by MorphDiff can help discover drugs with dissimilar structures but the same MOAs. In summary, with its strong capabilities in generating high-fidelity cell morphology on unseen perturbations, we envision MorphDiff as a powerful tool in phenotypic drug discovery by accelerating the phenotypic screening of vast perturbation space and improving MOA identification.