Generative 3D cell-morphology model from the Bakal lab and Sentinal4D that simulates how single cells reshape under each drug it was trained on.
No providers recorded yet. Browse all providers
A melanoma cell treated with the myosin-II inhibitor blebbistatin grows long, thin protrusions; treated with nocodazole, it collapses into a sphere. Photographed as a flat 2D projection, much of that difference — how far a protrusion reaches along the optical axis, where the nucleus sits inside the cell body — is lost. Most generative models of perturbation-induced morphology, MorphoDiff among them, work on 2D images, so what they learn is a compressed shadow of the cell. FORM is a generative model of perturbed cell shape that operates natively on 3D volumes.
FORM (not to be confused with the RNA-modification model m6A-FORM) was developed by Reed Naidoo, Chris Bakal and colleagues in the Bakal lab at The Institute of Cancer Research, London and its spinout Sentinal4D, and posted to arXiv in October 2025. The same first author and lab later built PhenoSeq.
The mechanism has two stages. A multi-channel VQGAN compresses each 3D single-cell volume into discrete tokens, with separate codebooks for the cytoplasm and the nucleus, so that the two compartments are represented as related but distinct structures rather than a single blended intensity field. A latent diffusion model then learns the distribution of those tokens for a given perturbation. Rather than learning an explicit untreated-to-treated mapping, FORM models each treatment as its own morphological landscape and moves a cell between landscapes by partially noising it and denoising it under a different condition.
A separate FORM model — VQGAN plus a dual-channel 3D UNet denoising diffusion model with spatial and depth-wise attention — is trained for each drug perturbation. The VQGAN is optimised with reconstruction, commitment and adversarial losses, then frozen and used as the tokeniser for the diffusion stage. Training data were over 65,000 WM266-4 melanoma cells expressing CAAX-EGFP, H2B-iRFP670 and ERK-KTR-Ruby, embedded in collagen and imaged by stage-scanning oblique plane microscopy after treatment with binimetinib (MEK), blebbistatin (myosin II), H1152 (ROCK), PF228 (FAK), palbociclib (CDK4/6) or nocodazole (microtubules). That compound panel is the one the Results section gives; the Methods section lists a partly different one, adding CK666 and MK1775 and omitting palbociclib.
Applied without retraining to a separate RNAi screen of more than 35,000 cells covering 167 Rho GTPase-pathway genes, FORM encoder embeddings recovered known gene relationships better than 2D OpenPhenom embeddings across CORUM, HuMAP, Reactome, SIGNOR and STRING (CORUM recall 0.556 versus 0.333). For unconditional generation of nocodazole, blebbistatin and binimetinib populations, FORM scored an inverse FID of 0.822, F1 of 0.57 and coverage of 0.741, ahead of the 3D medical-imaging generators HA-GAN and MedicalDiffusion on all three. Predicted ERK-KTR ratios across the knockdowns correlated with independently measured nuclear pERK at ρ = −0.50, the expected inverse relationship, with a concordance index of 0.68.
FORM is aimed at phenotypic drug discovery: simulating how a population of cells is likely to reshape under a compound, and comparing predicted 3D shape changes against known mechanisms. Its scope is the drugs it was trained on. Each drug perturbation has its own independently trained checkpoint, fitted on WM266-4 melanoma cells, so a compound outside that panel requires training a new model. What the paper demonstrates is transfer to unseen cell lines rather than to unseen drugs: the binimetinib model trained on melanoma cells generated treated morphologies for three triple-negative breast cancer lines it had not seen, and the outputs clustered with the real treated cells of each line. Combining real RNAi knockdown cells with simulated binimetinib treatment reproduced the known result that DOCK5 depletion sensitises cells to MEK inhibition. Sentinal4D uses FORM inside its SentiCORE platform, so outside researchers access it through the company, not as a downloadable model.
FORM moves generative perturbation modelling of cell morphology from 2D images to 3D volumes, and its link from shape to ERK activity argues that 3D structure carries recoverable signalling information. The paper also introduces MorphoEval, a benchmarking suite that scores generated morphologies on shape descriptors, distributional metrics computed in the FORM encoder's own feature space, and downstream signalling predictions. The limits are clear. FORM is an arXiv preprint without peer review, it trains one model per perturbation rather than a single model conditioned on arbitrary compounds, the evaluation covers a handful of drugs in one melanoma line plus a small transfer to breast cancer lines, and no code, weights or imaging data have been released. It sits alongside V3Cell, which builds 3D organoid surrogates from brightfield video, in an emerging push toward three-dimensional virtual-cell models.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.