This can be a sponsored article delivered to you by MBZUAI.
Should you’ve ever tried to guess how a cell will change form after a drug or a gene edit, you realize it’s half science, half artwork, and principally costly trial-and-error. Imaging 1000’s of situations is sluggish; exploring tens of millions is not possible.
A new paper in Nature Communications proposes a distinct route: simulate these mobile “after” photos immediately from molecular readouts, so you may preview the morphology earlier than you decide up a pipette. The workforce calls their mannequin MorphDiff, and it’s a diffusion mannequin guided by the transcriptome, the sample of genes turned up or down after a perturbation.
At a excessive stage, the concept flips a well-recognized workflow. Excessive-throughput imaging is a confirmed technique to uncover a compound’s mechanism or spot bioactivity however profiling each candidate drug or CRISPR goal isn’t possible. MorphDiff learns from instances the place each gene expression and cell morphology are recognized, then makes use of solely the L1000 gene expression profile as a situation to generate practical post-perturbation photos, both from scratch or by remodeling a management picture into its perturbed counterpart. The declare is that aggressive constancy on held-out (unseen) perturbations throughout massive drug and genetic datasets plus beneficial properties on mechanism-of-action (MOA) retrieval can rival actual photos.
This analysis led by MBZUAI researchers begins from a organic remark: gene expression finally drives proteins and pathways that form what a cell appears like below the microscope. The mapping isn’t one-to-one, however there’s sufficient shared sign for studying. Conditioning on the transcriptome gives a sensible bonus too: there’s merely much more publicly accessible L1000 knowledge than paired morphology, making it simpler to cowl a large swath of perturbation house. In different phrases, when a brand new compound arrives, you’re more likely to discover its gene signature which MorphDiff can then leverage.
Underneath the hood, MorphDiff blends two items. First, a Morphology Variational Autoencoder (MVAE) compresses five-channel microscope photos right into a compact latent house and learns to reconstruct them with excessive perceptual constancy. Second, a Latent Diffusion Mannequin learns to denoise samples in that latent house, steering every denoising step with the L1000 vector through consideration.
Wang et al., Nature Communications (2025), CC BY 4.0
Diffusion is an efficient match right here: it’s intrinsically sturdy to noise, and the latent house variant is environment friendly sufficient to coach whereas preserving picture element. The workforce implements each gene-to-image (G2I) technology (begin from noise, situation on the transcriptome) and image-to-image (I2I) transformation (push a management picture towards its perturbed state utilizing the identical transcriptomic situation). The latter requires no retraining because of an SDEdit-style process, which is helpful whenever you need to clarify adjustments relative to a management.
It’s one factor to generate photogenic footage; it’s one other to generate biologically devoted ones. The paper leans into each: on the generative facet, MorphDiff is benchmarked in opposition to GAN and diffusion baselines utilizing commonplace metrics like FID, Inception Rating, protection, density, and a CLIP-based CMMD. Throughout JUMP (genetic) and CDRP/LINCS (drug) take a look at splits, MorphDiff’s two modes usually land first and second, with significance checks run throughout a number of random seeds or impartial management plates. The result’s constant: higher constancy and variety, particularly on OOD perturbations the place sensible worth lives.
The larger image is that generative AI has lastly reached a constancy stage the place in-silico microscopy can stand in for first-pass experiments.
Extra attention-grabbing for biologists, the authors step past picture aesthetics to morphology options. They extract a whole lot of CellProfiler options (textures, intensities, granularity, cross-channel correlations) and ask whether or not the generated distributions match the bottom fact.
In side-by-side comparisons, MorphDiff’s function clouds line up with actual knowledge extra carefully than baselines like IMPA. Statistical checks present that over 70 p.c of generated function distributions are indistinguishable from actual ones, and feature-wise scatter plots present the mannequin accurately captures variations from management on probably the most perturbed options. Crucially, the mannequin additionally preserves correlation construction between gene expression and morphology options, with larger settlement to floor fact than prior strategies, proof that it’s modeling greater than floor fashion.
Wang et al., Nature Communications (2025), CC BY 4.0
The drug outcomes scale up that story to 1000’s of remedies. Utilizing DeepProfiler embeddings as a compact morphology fingerprint, the workforce demonstrates that MorphDiff’s generated profiles are discriminative: classifiers educated on actual embeddings additionally separate generated ones by perturbation, and pairwise distances between drug results are preserved.
Wang et al., Nature Communications (2025), CC BY 4.0
That issues for the downstream activity everybody cares about: MOA retrieval. Given a question profile, can you discover reference medication with the identical mechanism? MorphDiff’s generated morphologies not solely beat prior image-generation baselines but additionally outperform retrieval utilizing gene expression alone, they usually strategy the accuracy you get utilizing actual photos. In top-k retrieval experiments, the typical enchancment over the strongest baseline is 16.9 p.c and eight.0 p.c over transcriptome-only, with robustness proven throughout a number of okay values and metrics like imply common precision and folds-of-enrichment. That’s a robust sign that simulated morphology incorporates complementary info to chemical construction and transcriptomics which is sufficient to assist discover look-alike mechanisms even when the molecules themselves look nothing alike.
MorphDiff’s generated morphologies not solely beat prior image-generation baselines but additionally outperform retrieval utilizing gene expression alone, they usually strategy the accuracy you get utilizing actual photos.
The paper additionally lists some present limitations that trace at potential future enhancements. Inference with diffusion stays comparatively sluggish; the authors recommend plugging in newer samplers to hurry technology. Time and focus (two elements that biologists care about) aren’t explicitly encoded resulting from knowledge constraints; the structure might take them as further situations when matched datasets turn into obtainable. And since MorphDiff relies on perturbed gene expression as enter, it may well’t conjure morphology for perturbations that lack transcriptome measurements; a pure extension is to chain with fashions that predict gene expression for unseen medication (the paper cites GEARS for instance). Lastly, generalization inevitably weakens as you stray removed from the coaching distribution; bigger, better-matched multimodal datasets will assist, as will conditioning on extra modalities akin to buildings, textual content descriptions, or chromatin accessibility.
What does this imply in observe? Think about a screening workforce with a big L1000 library however a smaller imaging price range. MorphDiff turns into a phenotypic copilot: generate predicted morphologies for brand spanking new compounds, cluster them by similarity to recognized mechanisms, and prioritize which to picture for affirmation. As a result of the mannequin additionally surfaces interpretable function shifts, researchers can peek below the hood. Did ER texture and mitochondrial depth transfer the best way we’d count on for an EGFR inhibitor? Did two structurally unrelated molecules land in the identical phenotypic neighborhood? These are the sorts of hypotheses that speed up mechanism looking and repurposing.
The larger image is that generative AI has lastly reached a constancy stage the place in-silico microscopy can stand in for first-pass experiments. We’ve already seen text-to-image fashions explode in client domains; right here, a transcriptome-to-morphology mannequin reveals that the identical diffusion equipment can do scientifically helpful work akin to capturing refined, multi-channel phenotypes and preserving the relationships that make these photos greater than eye sweet. It gained’t exchange the microscope. But when it reduces the variety of plates you need to run to seek out what issues, that’s money and time you may spend validating the hits that depend.
