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

Xuesong Wang, Yimin Fan, Yucheng Guo, Chenghao Fu, Kinhei Lee, Khachatur Dallakyan, Yaxuan Li, Qijin Yin, Yu Li, Le SongNature Communications 2025 · OpenReview

Abstract

Investigating cell morphology changes after perturbations using high-throughput image-based profiling is increasingly important for phenotypic drug discovery, including predicting mechanisms of action (MOA) and compound bioactivity. The vast space of chemical and genetic perturbations makes it impractical to explore all possibilities using conventional methods. Here we propose MorphDiff, a transcriptome-guided latent diffusion model that simulates high-fidelity cell morphological responses to perturbations. We demonstrate MorphDiff’s effectiveness on three large-scale datasets, including two drug perturbation and one genetic perturbation dataset, covering thousands of perturbations. Extensive benchmarking shows MorphDiff accurately predicts cell morphological changes under unseen perturbations. Additionally, MorphDiff enhances MOA retrieval, achieving an accuracy comparable to ground-truth morphology and outperforming baseline methods by 16.9% and 8.0%, respectively. This work highlights MorphDiff’s potential to accelerate phenotypic screening and improve MOA identification, making it a powerful tool in drug discovery.

Introduction

In drug discovery, phenotypic screening—observing how a cell's physical appearance (morphology) changes after treatment—is a gold standard for understanding drug mechanisms. However, the perturbation space is effectively infinite, encompassing millions of potential chemical compounds and tens of thousands of genetic targets. Physically imaging every possible combination is logistically and financially prohibitive.

While computational models like Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) have attempted to simulate these responses, they often produce blurry images or fail to generalize to novel, "unseen" drugs. There is a critical need for an in-silico tool that can accurately simulate these cellular responses, effectively creating a "digital twin" for cell morphology to accelerate drug discovery.

The MorphDiff Framework

MorphDiff leverages the power of Latent Diffusion Models (LDMs), adapted for biological data, to generate high-fidelity microscopic images. The framework operates in two primary stages:

  1. Morphology VAE (MVAE): A Variational Autoencoder compresses complex, 5-channel "Cell Painting" images (highlighting DNA, RNA, mitochondria, ER, and AGP) into a low-dimensional latent space, simplifying the generative task.
  2. Transcriptome-Guided Diffusion: The core model learns to denoise these latent representations, conditioned on L1000 gene expression profiles. By using the transcriptome as a biological "prompt," the model generates physical morphological states that correspond to specific molecular perturbations.
Overview of the MorphDiff pipeline, including dataset curation, model architecture (MVAE and LDM), inference modes (G2I and I2I), and downstream applications.

The model supports two inference modes: Gene-to-Image (G2I), which generates a perturbed image from gene expression data alone, and Image-to-Image (I2I), which predicts transitions by editing a "control" cell image based on a target gene expression profile.

Results and Discussion

MorphDiff was validated across three massive datasets: CDRP (1,028 drug perturbations), JUMP (130 genetic perturbations), and LINCS (61 drug perturbations). The model demonstrated superior fidelity, outperforming baselines like MorphNet and IMPA across standard image metrics such as FID and Inception Score, as well as biological metrics like CellProfiler feature correlation.

Quantitative benchmarking of MorphDiff against baseline models and qualitative analysis of feature preservation and MOA clustering.

Key Biological Insights

  • Transcriptome as a Universal Bridge: Gene expression was found to be a more informative condition for predicting morphology than chemical structure (SMILES strings), as it captures the direct biological effect of a drug rather than just its shape.
  • Enhanced MOA Retrieval: MorphDiff-generated images achieved MOA identification accuracy comparable to real ground-truth images, outperforming the previous state-of-the-art (IMPA) by 16.9% and gene-expression-only retrieval by 8.0%.
  • Functional Similarity: The model successfully grouped drugs with the same biological target even when they were chemically dissimilar, suggesting it learns deep biological principles rather than simple pattern memorization.

Conclusion

MorphDiff represents a significant step toward reliable in-silico phenotypic screening. While diffusion models are computationally intensive, the high fidelity and generalizability of MorphDiff offer a powerful tool for predicting the effects of novel perturbations. Future work will focus on improving inference speed and incorporating temporal and concentration-dependent data to further refine these cellular "digital twins."

Summary updated

Abstract

Investigating cell morphology changes after perturbations using high-throughput image-based profiling is increasingly important for phenotypic drug discovery, including predicting mechanisms of action (MOA) and compound bioactivity. The vast space of chemical and genetic perturbations makes it impractical to explore all possibilities using conventional methods. Here we propose MorphDiff, a transcriptome-guided latent diffusion model that simulates high-fidelity cell morphological responses to perturbations. We demonstrate MorphDiff’s effectiveness on three large-scale datasets, including two drug perturbation and one genetic perturbation dataset, covering thousands of perturbations. Extensive benchmarking shows MorphDiff accurately predicts cell morphological changes under unseen perturbations. Additionally, MorphDiff enhances MOA retrieval, achieving an accuracy comparable to ground-truth morphology and outperforming baseline methods by 16.9% and 8.0%, respectively. This work highlights MorphDiff’s potential to accelerate phenotypic screening and improve MOA identification, making it a powerful tool in drug discovery. Wang, Fan, Guo, and colleagues present MorphDiff, a transcriptome-guided latent diffusion model that accurately predicts cell morphological responses to perturbations, enhancing MOA identification and phenotypic drug discovery.
BibTeX
@inproceedings{wang2025prediction,
  author = {Xuesong Wang and Yimin Fan and Yucheng Guo and Chenghao Fu and Kinhei Lee and Khachatur Dallakyan and Yaxuan Li and Qijin Yin and Yu Li and Le Song},
  title = {{Prediction of cellular morphology changes under perturbations with a transcriptome-guided diffusion model}},
  booktitle = {Nature Communications},
  year = {2025},
  url = {https://openreview.net/forum?id=bV9wVj1mlV},
}
Figure