Morphologically Intelligent Perturbation Prediction with FORM

TL;DR

The paper addresses the challenge of modelling perturbation responses at subcellular detail by introducing FORM, a 3D generative framework that operates directly on multicellular morphologies. It combines a multichannel VQGAN morphology encoder with a diffusion-based Perturbation Trajectory Module to produce both unconditional and perturbation-conditioned 3D cell morphologies, and pairs this with downstream signaling predictions such as ERK-KTR readouts. A new benchmarking suite, MorphoEval, quantifies structural, statistical, and functional fidelity, and the approach demonstrates generalization across unseen cancer subtypes and recapitulates known genotype–drug interactions. The work advances in silico perturbation testing by linking 3D morphology to signalling and morphology dynamics, offering a platform for high-resolution virtual cell simulations and drug-discovery insights.

Abstract

Understanding how cells respond to external stimuli is a central challenge in biomedical research and drug development. Current computational frameworks for modelling cellular responses remain restricted to two-dimensional representations, limiting their capacity to capture the complexity of cell morphology under perturbation. This dimensional constraint poses a critical bottleneck for the development of accurate virtual cell models. Here, we present FORM, a machine learning framework for predicting perturbation-induced changes in three-dimensional cellular structure. FORM consists of two components: a morphology encoder, trained end-to-end via a novel multi-channel VQGAN to learn compact 3D representations of cell shape, and a diffusion-based perturbation trajectory module that captures how morphology evolves across perturbation conditions. Trained on a large-scale dataset of over 65,000 multi-fluorescence 3D cell volumes spanning diverse chemical and genetic perturbations, FORM supports both unconditional morphology synthesis and conditional simulation of perturbed cell states. Beyond generation, FORM can predict downstream signalling activity, simulate combinatorial perturbation effects, and model morphodynamic transitions between states of unseen perturbations. To evaluate performance, we introduce MorphoEval, a benchmarking suite that quantifies perturbation-induced morphological changes in structural, statistical, and biological dimensions. Together, FORM and MorphoEval work toward the realisation of the 3D virtual cell by linking morphology, perturbation, and function through high-resolution predictive simulation.

Figures (7)

• Figure 1: Overview of FORM.a) Single-cell 3D volumes are processed through the Form Encoder, and the resulting embeddings are used to train the perturbation trajectory module. b) The perturbation trajectory module samples from stochastic noise to generate morphologies under a specified perturbation condition. c) Conditioning on a control input, the model generates the corresponding post-treatment morphology and quantifies morphological changes relative to the control. d) Predicted morphologies can be further used to simulate intracellular signalling activity directly from structure. e)Form also supports modelling of morphodynamic changes, enabling prediction of morphological evolution between perturbation transition states.
• Figure 2: Unconditional generative synthesis with FORM.a) Workflow for analysing unconditional samples. Each generated volume is converted into a mesh for the extraction of morphological descriptors and simultaneously passed through the Form Encoder to obtain feature embeddings for downstream performance evaluation. b) Representative 3D volumes generated under different perturbation settings, with corresponding orthogonal views (axial, coronal, sagittal). c) Comparison of Form-generated samples with state-of-the-art baselines (HA-GAN and MedicalDiffusion) using maximum intensity projections across three representative perturbations: nocodazole, blebbistatin, and binimetinib.
• Figure 3: Conditional generation with Form.a) Conditional generation from an untreated (DMSO) input cell. The leftmost column shows the real control; subsequent columns show Form-generated post-treatment morphologies under different perturbation prompts. Top: 3D volumetric renderings. Middle: mesh reconstructions. Bottom: relative percentage change in key morphological descriptors versus the untreated control. b) Distributions of morphological descriptors across all generated samples ($N=1{,}000$ per perturbation) benchmarked against real counterparts ($N=1{,}000$). c) Hierarchical clustering of Form Encoder embeddings for real and generated samples across perturbations, showing that generated cells co-cluster with their corresponding real treatment groups. d) Cross-subtype generalisation: using the WM266-4 binimetinib model, DMSO controls from TNBC cell lines (231, 468, 159) were used as conditioning inputs to generate binimetinib-treated morphologies; a hierarchical clustermap of Form Encoder embeddings across all cell lines and perturbation conditions (real and generated) shows that generated samples co-cluster with their corresponding real groups.
• Figure 4: Morphodynamic evolution of Form-generated cells. Morphological descriptor trajectories (eccentricity, sphericity, protrusivity, etc.) are shown as a function of denoising timestep ($t = 300 \rightarrow 0$), capturing how cellular shape evolves during the generative process. Below, representative 3D renderings of generated cells at selected timesteps illustrate the corresponding structural transitions, linking quantitative descriptor changes with visually interpretable morphology.
• Figure 5: FORM models perturbation-induced changes in ERK signalling activity.a)Form is used to conditionally generate KTR activity maps for each gene knockdown from the RNAi library. From these predictions, ERK-KTR ratios are calculated, averaged per knockdown, and Z-score normalised. Representative examples are shown, where cytoplasm and nucleus inputs are used to synthesise the corresponding ERK-KTR signal. b) Z-score normalised predicted ERK-KTR ratios are compared to experimentally measured pERK intensities across knockdowns. Each point represents a gene, illustrating the alignment between predicted signalling states and true biochemical measurements. c) Schematic of the simulation pipeline: untreated gene knockdowns are conditionally transformed to model the impact of binimetinib treatment, enabling the generation of predicted ERK activity maps. d) Lollipop plot showing the percentage increase in ERK-KTR ratio following simulated binimetinib treatment across a range of gene knockdowns, reflecting predicted ERK inhibition.