MorphDiff

Image-based profiling with Cell Painting — a six-dye, five-channel fluorescence assay that captures the morphological state of individual cells — has become a workhorse of phenotypic drug discovery. The obstacle is combinatorial: the space of chemical and genetic perturbations is far larger than any screening campaign can image, so most perturbations of interest are never observed morphologically. Transcriptomic profiling, by contrast, is comparatively cheap and broadly available through resources such as the L1000 platform.

MorphDiff bridges these two modalities. It is a transcriptome-guided latent diffusion model that takes an L1000 gene expression profile measured after a perturbation and generates the corresponding single-cell morphology images, without ever seeing that perturbation during training. Developed by BioMap Research with the Chinese University of Hong Kong and MBZUAI, it was published in Nature Communications in September 2025.

The model sits alongside a small family of generative morphology tools. Where IMPA performs style-transfer-style image translation from control images conditioned on a learned perturbation embedding, MorphDiff conditions directly on a measured transcriptome — an experimentally grounded signal that carries perturbation-specific biology rather than an identifier looked up from a training vocabulary. It is generative in a sense complementary to embedding models such as OpenPhenom-S/16, and shares its diffusion backbone lineage with MorphGen.

#Key Features

• Transcriptome conditioning: L1000 expression is projected by a fixed linear encoder and injected into the denoising U-Net through cross-attention keys and values, so generation is driven by measured biology rather than a perturbation label.
• Two inference modes: G2I generates perturbed morphology from gene expression alone; I2I starts from an unperturbed control image and edits it toward the perturbed state using SDEdit, requiring no retraining.
• Unseen-perturbation generalization: evaluated on held-out out-of-distribution splits and leave-one-out target and MOA splits, with statistical tests confirming the held-out sets are genuinely distributionally distinct from training data.
• Native five-channel output: the model generates all five Cell Painting channels (DNA, ER, RNA, AGP, Mito) directly, whereas most image-translation baselines are restricted to three-channel RGB.
• MOA retrieval from generated images: DeepProfiler embeddings of MorphDiff outputs support mechanism-of-action retrieval at accuracy comparable to ground-truth morphology.

#Technical Details

MorphDiff adapts the Stable Diffusion v1 framework in two stages. A morphology VAE with adversarial patch discriminator, per-channel LPIPS loss, and KL regularization compresses 128×128×5 Cell Painting crops into a 32×32×4 latent; a DDPM-style latent diffusion model with a cross-attention U-Net then learns to denoise those latents conditioned on gene expression. Single cells are segmented from plate-level images with CellProfiler 4.2.5.

Training and evaluation span three datasets: 1,028 drug perturbations from CDRP (U2OS, 977 L1000 landmark genes), 130 gene-overexpression perturbations from JUMP (U2OS, 12,328 genes), and 61 drugs from LINCS (A549). Splits hold out 10% of perturbations as out-of-distribution — 96 for CDRP, 13 for JUMP. Against MorphNet, DMIT, DRIT++, StarGAN, VQGAN, MDTv2, and IMPA, MorphDiff leads on FID, CMMD, Inception Score, density, and coverage. On JUMP OOD, over 70% of the 232 CellProfiler features generated by MorphDiff are statistically indistinguishable from ground truth; 87.6% of G2I samples reach R² > 0.5 against measured features. Wasserstein distances between perturbation groups correlate at 0.97 with ground truth. On the CDRP Target_MOA set (10 targets, 69 drugs, 35 MOAs), top-5 MOA retrieval from generated images beats IMPA by 16.9% and expression-based retrieval by 8.0%.

#Applications

MorphDiff is aimed at phenotypic screening groups that have transcriptomic readouts for far more compounds than they have imaging data. Given an L1000 profile, it produces the morphological readout in silico, letting teams triage candidate compounds, hypothesize mechanisms of action before committing plate time, and interpret hits through CellProfiler features that map onto recognizable organelle-level changes. The I2I mode is useful when matched control images from the same plate are available, since it anchors generation to the batch's own imaging conditions.

#Impact

The work establishes cross-modal generation between transcriptome and morphology as a practical route to expanding perturbation screens, and demonstrates that generated images retain enough perturbation-specific signal to support downstream MOA analysis rather than merely looking plausible. The authors are explicit about the main limitation: accuracy degrades as test compounds become chemically more distant from the training set, a constraint shared with CellOT, IMPA, and other perturbation-response models. Evaluation is entirely computational, benchmarked against held-out measured images rather than new experiments. Code is released under an MIT license, though no pretrained checkpoint has been published, so reproducing the trained model requires running both training stages from the public datasets.