MORPH

Genetic perturbation screens such as Perturb-seq have made it possible to knock out or activate genes and read out the single-cell consequences at scale, but the space of possible perturbations is far larger than any experiment can cover. A screen that measures a few hundred single-gene perturbations leaves millions of untested combinations and countless untested cellular contexts. MORPH addresses this by learning to predict the single-cell outcome of a genetic perturbation, including perturbations, combinations, and cell contexts never seen during training.

MORPH — a modular framework for predicting responses to perturbational changes — was developed by Chujun He, Jiaqi Zhang, Munther Dahleh, and Caroline Uhler at the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard, together with the Laboratory for Information and Decision Systems (LIDS) and the Institute for Data, Systems, and Society (IDSS) at MIT. It was posted to bioRxiv in July 2025.

What distinguishes MORPH is that it operates across data modalities. The same framework predicts both molecular readouts, such as single-cell transcriptomes, and microscopy-based, image-derived cell profiles, positioning it as a general engine for perturbation-response prediction rather than a transcriptomics-only tool. Its attention mechanism also makes predictions interpretable, surfacing the latent features and gene programs a perturbation acts through.

#Key Features

• Cross-modality prediction: A single framework predicts perturbation outcomes for both single-cell transcriptomics and imaging-based (Cell Painting-style) profiles, rather than being specialized to one readout.
• Zero-shot generalization: MORPH forecasts responses to single-gene perturbations, combinatorial perturbations, and perturbations in new cellular contexts that were never observed during training.
• Discrepancy-based VAE with attention: Separate encoders map a control cell and a perturbation into a shared latent space, where attention modules dynamically identify the latent features most relevant to the given perturbation.
• Interpretable gene programs: The attention weights reveal which latent programs a perturbation engages, allowing the model to infer gene regulatory relationships alongside its predictions.
• Transfer learning from a genome-wide checkpoint: A checkpoint pretrained on genome-wide Perturb-seq data can be adapted to a new dataset, the workflow the authors recommend as the optimal starting point over training from scratch.

#Technical Details

MORPH combines a discrepancy-based variational autoencoder with an attention mechanism. For each pairing of a control cell and a genetic perturbation, dedicated encoders project both into latent representations; within that latent space, attention modules weight the features most relevant to the perturbation before decoding the predicted post-perturbation profile. This modular design is what lets the same architecture serve transcriptomic and imaging readouts. The framework was evaluated on three single-cell Perturb-seq datasets spanning different scales and cell lines — a K562 essential-gene screen, an RPE1 screen, and a genome-wide K562 screen — using multiple sources of prior biological knowledge. Across the applications tested, MORPH outperformed the available baseline perturbation-prediction models, and its pretraining on genome-wide Perturb-seq data provides the checkpoint used for transfer learning to new experiments. Running the demos requires a single GPU and roughly 20 GB of disk space.

#Applications

MORPH is aimed at researchers designing and interpreting genetic perturbation screens in functional genomics and drug-target discovery. Before an experiment, it can prioritize which perturbations or combinations are most likely to be informative, reducing the combinatorial cost of screening; after an experiment, it can impute responses for conditions or contexts that were not directly measured. Because it spans transcriptomic and imaging readouts, it fits into both sequencing-based Perturb-seq pipelines and image-based high-content screening. Its interpretable attention over gene programs additionally supports hypothesis generation about gene regulatory structure.

#Impact

MORPH extends single-cell perturbation modeling beyond the transcriptome-only setting that characterizes models such as GEARS, CPA, and scGen, offering one framework that generalizes across perturbations, combinations, cellular contexts, and data modalities while remaining interpretable. Coming from Caroline Uhler's group, a leading contributor to causal and representation-learning approaches in single-cell biology, it arrives as the community increasingly treats genome-wide Perturb-seq atlases as pretraining substrates for transferable perturbation-response models. As a preprint, its results await peer review, and the pretrained checkpoint is distributed through the project's GitHub repository rather than a dedicated model hub.