A generative framework for predicting cellular morphological and transcriptomic perturbation responses Authors: Rui Peng, Ziru Liu, Yuanxu Gao, Jinzhuo Wang Published: 2026 ABSTRACT SummaryPredicting coordinated cellular responses to perturbations is crucial for phenotypic drug discovery and building virtual cell models. Current methods typically predict isolated modalities or require empirical post-perturbation data, precluding true de novo virtual screening. To address this, we introduce MultiVCDiff, a unified multimodal generative diffusion framework that simultaneously predicts high-fidelity cellular morphology images and transcriptomic profiles directly from chemical or genetic perturbations. This framework was trained on a comprehensive trimodal corpus encompassing 1,118 chemical and 130 genetic perturbations across four datasets. Evaluated under strict zero-shot settings on unseen drugs, MultiVCDiff outperforms state-of-the-art single-modality baselines. It successfully overcomes morphological mode collapse, generates highly accurate gene expression signatures, and captures complex biological relationships to reliably retrieve drug mechanisms of action. By eliminating the dependency on prior post-perturbation readouts, MultiVCDiff enables scalable in silico hypothesis generation, significantly accelerating drug discovery and advancing holistic virtual cell modeling.Keywords: virtual cell modeling, multimodal generative model, diffusion model, perturbation response prediction, virtual screening, Cell Painting, L1000, cellular morphology synthesis, gene expression prediction, in silico phenotypic screening Introduction Predicting cell’s responses to perturbations is important to understanding the underlying biological mechanisms and realizing the vision of AI virtual cell (AIVC).1 The AIVC is a visionary computational model designed to simulate cellular responses to external perturbations by integrating multiscale and multimodal data into a coherent representation of cellular state. A core function of virtual cell is to predict the complex changes when a cell is perturbed.1 Among them, the most informative changes are cellular morphology, which provides a high-content view of the cell’s structural state, and gene expression, which captures underlying molecular programs. These modalities are intimately linked, offering complementary insights into cellular function.2 Therefore, modeling them jointly is an essential and critical step toward building realistic virtual cell simulations that can accurately capture the full spectrum of perturbation responses. Recent advances in high-throughput screening have made large-scale data collection increasingly feasible. Platforms such as Cell Painting3 captures high-content morphological images, while transcriptomic assays like L10004 measures corresponding gene expression programs under thousands of conditions. Despite these powerful technologies, experimental approaches remain fundamentally constrained by cost, scale, and feasibility, making it impossible to exhaustively profile the cellular response to all combinations of perturbations, dosages, and cell types. This gap highlights the urgent need for computational frameworks that can learn from existing data to generalize and predict unobserved cellular responses, enabling in silico hypothesis generation and virtual screening at a scale unattainable through experiments alone. A number of recent studies have explored the use of machine learning to predict cellular responses to perturbations. Currently, most of these efforts are primarily confined to generating a single target modality. In the image domain, methods such as MorphoDiff,5 IMPA,6 StarGAN,7 DMIT,8 and MDTv29 focus on generating high-resolution images of cellular morphology. In the transcriptomic domain, models like ChemCPA,10 biolord,11 and PDGrapher12 are specifically designed to predict gene expression responses, disentangle single-cell data, or infer therapeutic targets based on transcriptomic profiles. More recently, sophisticated multi-modal models have been developed; they generate one modality conditioned by another. For instance, MorphDiff,13 MorphNet,14 and TRIDENT15 successfully utilize a transcriptome-guided approach to predict cellular morphology changes using gene expression profiles as the condition. While promising for specific tasks, existing computational approaches face two fundamental bottlenecks. First, many models are limited to predicting a single target modality in isolation, failing to capture the coordinated biological changes across different cellular dimensions. Second, cross-modal generation models (e.g., Morphdiff) inherently rely on post-perturbation data from one modality (such as gene expression) to synthesize the other (such as cellular morphology). This dependency poses a critical barrier to genuine de novo virtual screening. For completely novel, untested candidate perturbations, neither phenotypic nor molecular experimental readouts are available, rendering such generation models inapplicable in these practical scenarios. Therefore, a unified framework capable of simultaneously simulating both transcriptomic and morphological consequences directly from the perturbation itself (e.g., a compound’s chemical structure or a genetic target) would address a critical unmet need in the development of scalable and realistic virtual cell systems. In this study, we address this challenge by introducing MultiVCDiff, a perturbation-conditioned diffusion framework that leverages transformer to learn the joint distribution of both response modalities and generate them simultaneously. Unlike cross-modal generation frameworks that require one measured modality as input to reconstruct the other, MultiVCDiff starts from the perturbation representation and directly produces a coherent pair of morphology and gene-expression outputs. Its core self-attention mechanism operates across a joint representation of morphology and gene expression, enabling the model to explicitly learn their cross-modal associations while remaining conditioned on a given perturbation. Moreover, to rigorously train and evaluate this framework across diverse biological contexts, we construct a collection of four harmonized trimodal datasets by systematically bridging L1000 transcriptomic profiles and Cell Painting morphological images via shared perturbations across multiple public resources (Figure 1A). Specifically, this collection comprises 98 drug perturbations from BBBC021,16 959 small-molecule perturbations from CDRP,17 130 genetic perturbations from JUMP,18 and 61 drug perturbations from LINCS.19 After matching with L1000, we denote these paired datasets as CPgenes-BBBC021, CPgenes-CDRP, CPgenes-JUMP, and CPgenes-LINCS, respectively. Overview of the CPgenes datasets and the MultiVCDiff framework (A) We systematically bridge large-scale public L1000 (transcriptomics) and Cell Painting (morphology) assays by identifying shared chemical and genetic interventions. This process curates a comprehensive collection of four trimodal datasets (CPgenes-BBBC021, CPgenes-CDRP, CPgenes-JUMP, and CPgenes-LINCS), with each matched sample containing a cell morphology image, its corresponding gene expression profile, and the specific perturbation. (B) Paired inputs of cellular morphology images and gene expression profiles are passed through modality-specific encoders to obtain their latent representations. Notably, in the model diagram, the “snowflake” icon indicates that a module’s parameters are frozen and do not participate in training, whereas the “flame” icon denotes that the module’s parameters are trainable and actively optimized. (C) The specific perturbation is converted into a numerical embedding via a versatile perturbation encoder. (D) The generation process starts from random noise representing both the cellular image and the gene data, which are merged into a single, unified sequence. Guided by the specific perturbation embedding, a denoising transformer model then works backwards in steps to progressively remove this noise. The self-attention mechanism within the transformer allows the model to learn the complex relationships and dependencies between the cell’s morphology and its gene expression. Once the process is complete, the final, clean representations are separated and passed through their respective decoders to reconstruct the high-fidelity cellular morphology image and gene expression profile. Through comprehensive evaluations, we demonstrate that MultiVCDiff achieves state-of-the-art fidelity in both morphological and transcriptomic predictions. For cellular morphology, the model effectively overcomes the mode collapse prevalent in existing generative baselines, accurately rendering distinct, perturbation-specific phenotypes induced by chemical and genetic interventions. Concurrently, for transcriptomic responses, MultiVCDiff significantly outperforms specialized single-cell architectures, generating highly accurate and biologically interpretable gene expression signatures. Beyond modality-specific fidelity, we show that the joint multimodal landscape learned by MultiVCDiff successfully captures higher-order functional relationships, enabling highly accurate mechanism of action (MOA) retrieval for unseen drugs. Taken together, our findings establish MultiVCDiff as a powerful and scalable in silico engine for virtual cell modeling and drug screening. Results Overview of MultiVCDiff Generative models can learn the joint distribution of complex, high-dimensional data, making them a perfect match for our goal of simultaneously modeling morphology and transcriptomics. The potential of the joint distribution learning has been successfully demonstrated by RNA-CDM,20 a framework that generates realistic whole-slide image conditioned on gene expression profile, effectively translating transcriptomic information into visual histology. Among these generative techniques, diffusion models have emerged as the state-of-the-art for producing high-fidelity outputs,21,22,23,24 making them ideal for the task of in silico exploration of perturbational landscapes. Building upon these foundations, we introduce MultiVCDiff, a unified multimodal diffusion framework designed to simultaneously predict cellular morphology and transcriptomic profiles in response to diverse chemical and genetic perturbations. Our pipeline consists of three core modules: (1) a pair of modality-specific encoders for image and RNA data, (2) a conditional denoising transformer that operates on the joint multimodal space, and (3) a versatile perturbation encoder that provides conditional guidance. Crucially, this perturbation module is highly adaptable, designed to seamlessly process and embed not only small-molecule compounds (chemical perturbations) but also gene-level interventions (genetic perturbations). The overall architecture of MultiVCDiff is illustrated in Figures 1B–1D. The training process for MultiVCDiff begins by independently encoding paired inputs: a high-content cellular morphology image and its corresponding gene expression profile (Figure 1B). We employ a pretrained vision transformer (ViT)25 to transform the image into an image representation χimage, and a dedicated multilayer perceptron (MLP) encoder concurrently transforms the RNA profile into a dense embedding χrna. These initial embeddings, χimage0 and χrna0, are then subjected to a parallel forward diffusion process. Gaussian noise is incrementally added over a series of timesteps T, gradually corrupting the information until the embeddings approximate pure isotropic Gaussian noise at χimageT and χrnaT. To guide the generative process toward a specific perturbation outcome, MultiVCDiff incorporates a versatile perturbation encoder (Figure 1C) capable of processing both chemical and genetic interventions. For chemical perturbations, the encoder converts standard SMILES strings into numerical embeddings using the RDKit toolkit, effectively capturing the structural and physicochemical properties of the small molecules. For genetic perturbations, we leverage scGPT,26 a state-of-the-art single-cell foundation model, to extract dense, context-aware embeddings of the targeted genes. These distinct modalities of intervention are seamlessly mapped into a shared representation space. The resulting perturbation embedding χperturb then serves as the critical conditional signal, guiding the multimodal denoising process to generate the corresponding phenotypic and molecular responses. The core of MultiVCDiff is the multimodal backward denoising process, which learns the intricate relationship between cellular morphology and gene expression under perturbation. The process begins by concatenating the noisy image χimageT and RNA embedding χrnaT into a single, joint multimodal sequence χmmT. This unified sequence is then processed by a denoising Transformer, where the self-attention mechanism encourages the model to learn complex dependencies not only within each modality but, crucially, between them. This effectively creates a form of cross-modal attention, enabling the model to explicitly learn the intrinsic associations between specific morphological features and their corresponding gene expression patterns. This learning process is critically conditioned on both the perturbation embedding χperturb and the current timestep t, which forces the model to learn the precise mapping between a given perturbation and its corresponding joint multimodal features. By iteratively predicting and removing the noise from t = T down to t = 0, the model reverses the forward diffusion process, generating a clean, coherent multimodal embedding χmm0 that captures these learned interdependencies. Finally, this denoised joint embedding is separated back into its constituent image and RNA components χˆimage0 and χˆrna0, which are passed through their respective decoders to reconstruct the final high-fidelity cellular morphology image and gene expression profile (Figure 1D; “STAR Methods”). Through this complete noising and denoising process, the model learns the conditional joint distribution of morphology and gene expression given a specific perturbation, enabling it to simultaneously generate both modalities. During inference, the generation process is initiated by sampling two random tensors from a standard Gaussian distribution, serving as the initial noisy embeddings χmmT. By providing a specific perturbation embedding χperturb as a condition, MultiVCDiff executes the full backward denoising process to generate a paired set of morphology and transcriptomic profiles that correspond to the given perturbation (“STAR Methods”). MultiVCDiff generates high-fidelity and perturbation-specific morphological images Within the AIVC framework, cellular morphology provides an information-rich, high-content phenotypic readout that delivers critical insights into biological processes and cellular responses to exogenous perturbations. The accurate modeling of such morphological changes is essential for developing a virtual cell capable of authentically simulating cellular behavior. To rigorously evaluate the foundational capacity of MultiVCDiff in generating high-fidelity cellular morphology images, we comprehensively compared the quality and biological relevance of its outputs against a diverse panel of eight state-of-the-art generative baselines. These baselines were strategically selected to represent three primary modeling paradigms: (1) cross-modal predictors that map gene expression profiles to morphological phenotypes (MorphDiff and MorphNet), (2) perturbation-conditioned generators that predict phenotypic responses using style-transfer or specific perturbation embeddings as guidance (IMPA and MorphoDiff), and (3) foundational image synthesis architectures representing advanced computer vision frameworks for multi-domain generation and latent diffusion (StarGAN, DMIT, MDTv2, and a fine-tuned unconditional Stable Diffusion). To ensure a fair comparison, all models were trained for 100,000 steps under identical settings. From the original 98 BBBC021 compounds, 44 trimodal samples with matched L1000 RNA profiles were utilized exclusively for joint multimodal training. The remaining 54 compounds lacking transcriptomic data were strictly reserved for held-out, morphology-only generalization testing. Across six representative perturbations, MultiVCDiff generated images that most closely resembled the ground-truth data in global field organization, and channel-specific phenotypes (Figure 2A). The agreement was not limited to overall image texture but extended to perturbation-dependent differences in cell density, cell shape, and subcellular structures. For instance, in response to microtubule inhibitors like colchicine, MultiVCDiff correctly predicted a significant visual change that a more intense and diffuse signal in the tubulin channel (green). This is biologically consistent with colchicine’s known mechanism, which is to inhibit microtubule polymerization by binding to tubulin.27 When polymerization is blocked, the tubulin proteins cannot assemble into a structured cytoskeletal network and instead become dispersed, leading to the widespread green signal our model accurately portrayed. By contrast, the comprehensive suite of eight baseline models failed to produce meaningful, perturbation-specific phenotypes for these drugs (Figure 2A). The generative vision models, including MorphoDiff and unconditional Stable Diffusion, frequently produced generic high-density cellular textures with attenuated perturbation-to-perturbation variation, suggesting that they largely captured the overall image distribution without preserving the specific phenotype induced by each drug. Other methods, including StarGAN, IMPA, and MorphNet, struggled heavily with image fidelity. Even sophisticated models like MorphDiff, DMIT, and MDTv2 failed to reproduce critical drug-induced events such as cell rounding, cytoskeletal dispersion, and population sparsity. Importantly, their outputs also varied less consistently across the six representative perturbations than those of MultiVCDiff, indicating weaker conditional separation among drug responses. To demonstrate that this robust generative capacity extends beyond a single biological context, we further expanded our comparative analysis to the remaining three trimodal datasets. For large-scale chemical interventions, MultiVCDiff consistently produced highly accurate, perturbation-specific morphological responses across both the CPgenes-LINCS (Figure S1) and CPgenes-CDRP (Figure S2) cohorts. Notably, we also evaluated the model’s performance on genetic perturbations using the CPgenes-JUMP dataset. By leveraging the scGPT foundation model to extract context-aware embeddings for targeted genes, MultiVCDiff successfully translated genetic interventions into their corresponding phenotypic manifestations (Figure S3). Across all four diverse datasets, MultiVCDiff accurately identified and rendered the distinct morphological changes induced by both diverse drug and genetic perturbations. In stark contrast, the baseline methods persistently exhibited varying degrees of mode collapse across these cohorts, failing to capture the specific cellular consequences of these broader interventions. The ability of MultiVCDiff to preserve both perturbation-specific morphology and channel-level biological structure underscored its superior accuracy in modeling cellular morphology responses from perturbation input alone. Evaluation of the fidelity of cellular morphologies generated by MultiVCDiff (A) Representative drug perturbations demonstrate that MultiVCDiff’s outputs closely mirror the real images and accurately capture complex, drug-specific phenotypes. In contrast, the eight baseline models fail to capture the specific morphological effects induced by each perturbation. (B) We use two well-established metrics to quantitatively evaluate the quality of images generated by different models. MultiVCDiff consistently achieves the lowest scores across both metrics. ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001. (C) In the high-dimensional feature space visualized by UMAP, embeddings of images generated by MultiVCDiff are tightly intermingled with those from real, perturbed images, forming a joint cluster that is distinctly separated from the unperturbed control cells. To quantitatively substantiate these visual observations, we benchmarked the generated images using two well-established metrics: Fréchet Inception Distance (FID)28 and Kernel Inception Distance (KID).29 As illustrated in Figure 2B, MultiVCDiff demonstrated a profound and statistically significant improvement over the comprehensive suite of eight baseline models. Both FID and KID measure the distributional similarity between generated and real images, with lower scores indicating a closer match to the true data distribution. Across all evaluated perturbations, MultiVCDiff consistently achieved the lowest scores. This indicated that its generated image distribution most closely matched the ground-truth, confirming a much higher degree of realism and perturbation-specific accuracy. The substantial performance gap between our model and the competing baselines (p < 0.001 to p < 0.0001) provided strong quantitative evidence for the superior visual quality observed in Figure 2A. Conversely, the significantly higher FID and KID scores of the baselines, particularly foundational models like unconditional Stable Diffusion and MorphoDiff, quantitatively confirmed their tendency to suffer from mode collapse, generating generic images that severely diverge from the specific target phenotype. Taken together, this statistical evidence provided robust confirmation that MultiVCDiff uniquely generates images with superior fidelity and stronger alignment to the target perturbation’s phenotypic space. Finally, we moved beyond pixel-level metrics to analyze their high-dimensional representations. We projected both real and generated images into a deep feature space using a pre-trained ViT model and visualized the resulting embeddings via UMAP30 (Figure 2C). The analysis for representative compounds like vinblastine, docetaxel, and camptothecin revealed a striking alignment between our model’s output and the ground-truth data. In Figure 2C, the embeddings from MultiVCDiff-generated images and real perturbed images were tightly intermingled, occupying a shared and overlapping manifold. This near-perfect alignment was further corroborated by the marginal density plots, which showed an almost identical distribution for the generated and real embeddings along both UMAP axes, confirming they were virtually indistinguishable in this feature space. Crucially, this joint cluster of real and generated perturbed cells maintained a clear and significant separation from the control (unperturbed) cell population, confirming that MultiVCDiff not only produced visually realistic images but also generated images that are biologically correct in a high-dimensional feature space. MultiVCDiff captures biologically interpretable morphological signatures and generalizes to unseen perturbations Building upon the validation of high-fidelity image generation, we next assessed the capacity of MultiVCDiff to capture biologically interpretable morphological signatures and to generalize to unseen drug perturbations. To evaluate whether MultiVCDiff captures MOA-specific morphological programs, we directly compared the generated phenotypes for three structurally and functionally distinct compounds: doxorubicin, docetaxel, and aphidicolin (Figure 3A). It revealed that MultiVCDiff learned a robust, conditional mapping from a drug’s identity to its unique morphological fingerprint, successfully distinguishing between diverse MOAs. For example, our model’s prediction for the topoisomerase II inhibitor doxorubicin was highly biologically consistent. Doxorubicin functions by trapping the topoisomerase II enzyme in a complex with broken DNA strands, which prevents their re-ligation and halts replication.31,32,33 This induced replication stress leads to cell-cycle arrest, a state often characterized by enlarged nuclei. MultiVCDiff accurately portrayed this specific phenotype, generating images in which the DAPI-stained (blue) nuclei were visibly prominent and enlarged. In contrast, the phenotypes generated for docetaxel and aphidicolin were markedly different. Specifically, MultiVCDiff produced docetaxel-perturbed cells that exhibited reduced cellular spreading and a predominantly purple overall hue. Conversely, the predicted morphology for aphidicolin was characterized by a significantly more prominent signal in the red channel, which corresponds to the actin cytoskeleton. Together, the clear visual divergence among these predicted phenotypes indicates that MultiVCDiff successfully captures and generates MOA-specific morphological programs directly from the perturbation condition. Biological consistency and generalization of MultiVCDiff’s morphological predictions (A) Comparison of the generated phenotypes for three structurally and functionally distinct compounds. The clear morphological divergence among the generated images demonstrates that MultiVCDiff accurately captures and produces distinct, MOA-specific morphological signatures directly from the perturbation condition. (B) We use CellProfiler to extract 6,345 morphological features from the images. The boxplots compare the distributions of a representative feature (“AreaOccupied”) across the whole cell, cytoplasm, and nucleus for three different drugs. (C) The model’s ability to generalize is tested by predicting the morphological response to unseen 54 drugs that were held out from the training set. Image quality metrics were compared against baseline models. ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001. Moving beyond deep feature representations, we employed CellProfiler34 to perform a large-scale morphological analysis. This process generated a high-dimensional feature vector of 6,345 morphological measurements per image, quantifying a wide spectrum of cytological properties. These included cellular geometry (shape, size, and perimeter), textural information, and fluorescent signal intensity, all measured at the level of the whole cell, cytoplasm, and nucleus. We then compared the feature distributions of images generated by MultiVCDiff against ground-truth perturbed images and unperturbed controls (Figure 3B). For representative compounds such as docetaxel and emetine, the analysis showed a remarkable alignment between the feature distributions of the generated and true images. The boxplots for key features like “AreaOccupied” demonstrated that the distributions from generated samples consistently exhibited a significant shift away from the control population, closely mirroring the true phenotypic effect of the drug treatment. This quantitative evidence substantiated our qualitative observations, proving that MultiVCDiff accurately recapitulated the nuanced, multi-scale morphological changes induced by drug perturbations. The ultimate test for a predictive model in drug discovery is its ability to generalize to novel candidates. We, therefore, rigorously evaluated MultiVCDiff’s performance on a challenging out-of-distribution (OOD) task using 54 compounds from the CPgenes-BBBC021 dataset that were entirely held out during training. For a comprehensive comparison, we benchmarked our model against the full suite of eight state-of-the-art baselines. To ensure a strictly fair OOD evaluation and eliminate any potential data leakage, all baseline models were subjected to the exact same zero-shot setting and were completely blinded to these 54 compounds during training. As illustrated in Figure 3C, MultiVCDiff consistently outperformed the competing baselines across FID and KID. This result highlighted the powerful generalization capability of our model. By learning a coherent joint distribution between molecular and morphological data from the matched training corpus, MultiVCDiff developed a fundamental understanding of structure-to-phenotype relationships. This allowed the model to successfully extrapolate and make accurate morphological predictions for novel chemical structures driven solely by perturbation embeddings. This capacity for robust extrapolation is crucial for real-world applications, enabling reliable in silico screening and accelerating the identification of promising therapeutic drugs. MultiVCDiff generates high-fidelity and biologically interpretable transcriptional profiles A central goal of the AIVC framework is to bridge the gap between phenotypic changes and their underlying molecular drivers.1 Beyond generating high-fidelity morphology, we evaluated MultiVCDiff’s capacity to predict perturbation-specific transcriptional responses, a critical component for elucidating drug mechanisms and discovering novel therapeutics. To conduct a rigorous assessment of prediction accuracy and explicitly address potential data leakage, we evaluated our model on the CPgenes-LINCS dataset, which comprises 61 distinct small-molecule perturbations. We meticulously partitioned this dataset into a training set of 31 compounds and a strictly held-out test set of 30 compounds. Crucially, we enforced a stringent structural similarity constraint during this split to establish a true OOD setting. As illustrated by the boxplot in Figure 4B, the structural similarity between any drug in the unseen test set and the training cohort is strictly constrained below 0.2. This quantification ensures that the test compounds are chemically highly divergent from those encountered during training. We then compared the predicted log fold change (logFC) of gene expression against the ground truth across these 30 unseen OOD perturbations. The heatmap in Figure 4A revealed a highly perfect alignment between the predicted and ground-truth logFC for the top differentially expressed (DE) genes. MultiVCDiff accurately predicts biologically interpretable transcriptomic signatures for unseen perturbations (A) Heatmaps display the logFC of the top DE genes across the 30 strictly unseen drug perturbations from the CPgenes-LINCS dataset. The predicted logFC patterns demonstrate a striking, gene-level alignment with the ground-truth data (overall correlation = 0.872). (B) A boxplot quantifies the structural similarity between the 30 unseen test drugs and the training set. The maximum similarity strictly below 0.2, rigorously confirming a true zero-shot OOD setting. (C) Violin plots for representative unseen drugs compare the predicted and ground-truth distributions of log fold changes for key DE genes, highlighting the model’s distributional consistency. (D) Boxplots for representative unseen drugs assess the directional concordance of these expression changes, confirming that the upregulation and downregulation trends predicted by the model are highly consistent with the ground truth. (E) GO enrichment analysis is performed on the DE genes predicted by MultiVCDiff in response to the unseen HC-030031 perturbation. (F) The molecular function network visualizes the interconnected gene programs that are robustly identified and captured by our model under HC-030031 perturbation. Remarkably, even when tasked with predicting the responses of novel drugs sharing less than 0.2 structural similarity with the training set, our model achieved an impressive overall correlation of 0.872 between the predicted RNA expression and the ground truth. Furthermore, our model accurately captured the highly specific transcriptional signature unique to each unseen compound. The key genes exhibited diverse patterns of upregulation and downregulation in response to different drugs, a biological complexity that was precisely recapitulated in the model’s predictions. To statistically validate these observations, we next examined the distributional consistency of the predicted responses. We compared the logFC distributions of the top DE genes across 500 generated samples versus the ground-truth samples. As shown in Figure 4C, the predicted distributions for representative drugs closely aligned with the observed distributions. This indicated that MultiVCDiff accurately learned the variance and central tendencies of gene expression changes induced by each perturbation. Furthermore, we assessed the directional concordance of these changes in Figure 4D, which confirmed that the upregulation and downregulation trends predicted by the model are highly consistent with the ground truth. This high-fidelity, gene-level accuracy across structurally novel chemical perturbations indicated that MultiVCDiff did not merely memorize training profiles. Instead, the model learned a sophisticated and biologically relevant mapping from chemical structure to its precise impact on the cellular transcriptome, highlighting its powerful zero-shot extrapolation capabilities and providing strong quantitative evidence for its fidelity in generating statistically sound transcriptomic profiles. The ultimate measure of a predictive model’s utility lies in its ability to generate biologically interpretable results. To this end, we investigated whether the predicted transcriptional signatures could reliably reflect known drug MOAs. We selected HC-030031 as a compelling case study. HC-030031 is a well-established, selective antagonist of the TRPA1 channel, which effectively inhibits TRPA1-mediated calcium influx.35,36 TRPA1 functions as a potent cellular redox sensor that is highly responsive to oxidative stress signals, such as hydrogen peroxide. Crucially, HC-030031 represents a strict OOD test case. As depicted in Figure 4B, the median structural similarity between HC-030031 and the 31 drugs comprising our training set is merely 0.132. This rigorously confirms that the model had never encountered a remotely similar chemical scaffold during its training phase. We performed Gene Ontology (GO) term enrichment analysis on the DE genes predicted by MultiVCDiff in response to the HC-030031 perturbation (Figure 4D). The analysis revealed that the predicted gene signatures were significantly enriched in biological processes intimately tied to oxidative stress. Most notably, the top enriched terms included “response to reactive oxygen species,” “response to hydrogen peroxide,” and “regulation of reactive oxygen species metabolic process.” This finding perfectly aligns with the primary MOA of HC-030031. Because TRPA1 functions as a critical cellular redox sensor, its blockade by HC-030031 fundamentally alters the cell’s normal capacity to perceive and mitigate oxidative stress. Therefore, by accurately predicting the modulation of these specific pathways, MultiVCDiff successfully captured the precise downstream transcriptional consequences of TRPA1 antagonism. The molecular function network (Figure 4E) further corroborated these transcriptomic findings, visualizing the interconnected gene programs that were robustly identified by our model. This strong alignment between the predicted gene programs and the established core mechanism of HC-030031 provides compelling evidence that MultiVCDiff can successfully generalize to structurally novel, unseen drugs under rigorous OOD settings. It demonstrates that the model accurately predicts zero-shot gene expression profiles that are not only statistically sound but also profoundly biologically interpretable. MultiVCDiff outperforms state-of-the-art models in transcriptional perturbation prediction To benchmark the OOD transcriptional prediction capabilities of our multimodal framework, we expanded our comparative analysis using the strictly held-out test set of 30 unseen compounds from the CPgenes-LINCS dataset. We compared MultiVCDiff against three recent, state-of-the-art architectures specialized in predicting single-cell perturbational responses: (1) chemCPA,10 which utilizes molecular representations to predict drug effects, (2) biolord,11 a deep generative framework designed to disentangle cellular attributes, and (3) PDGrapher,12 a causally inspired graph neural network. By evaluating performance exclusively on these 30 novel chemical structures, we ensured a rigorous zero-shot setting where none of the models had prior exposure to the test compounds. We comprehensively benchmarked all models using the established Cell-Eval framework,37 which provides a robust suite of metrics to assess the identification of DE genes from single-cell perturbation experiments (see STAR Methods for detailed metric descriptions). Across the strictly held-out test set of 30 unseen drug perturbations, MultiVCDiff demonstrated consistently superior or highly competitive performance on a suite of eight distinct evaluation metrics when compared to the three baselines (Figure 5A). MultiVCDiff’s improvements were particularly pronounced against chemCPA and biolord across all metrics. For instance, in the precision of DE genes, MultiVCDiff achieved a mean score of 0.626 (±0.140), significantly outperforming chemCPA (0.474 ± 0.116), biolord (0.542 ± 0.103), and the causally inspired PDGrapher (0.601 ± 0.141). Furthermore, MultiVCDiff showed a marked advantage in capturing the overall quality of DE gene classification and ranking. It achieved a mean AUPRC of 0.729 (±0.175), demonstrating a substantial improvement over chemCPA (0.553 ± 0.135) and biolord (0.605 ± 0.145), while maintaining a statistically significant edge over PDGrapher (0.697 ± 0.171). This highlighted the model’s robust ability to identify true positive DE genes with high confidence, even for novel chemical structures. MultiVCDiff also excelled at predicting the magnitude and direction of expression changes. It obtained a mean Spearman correlation for log fold changes of 0.754 (±0.107), significantly exceeding chemCPA (0.569 ± 0.088) and biolord (0.633 ± 0.091), and outperforming PDGrapher (0.744 ± 0.107). Moreover, it achieved a highly accurate mean score of 0.849 (±0.083) in direction match. While PDGrapher showed competitive performance in this specific directional metric (0.857 ± 0.073, no significant difference), MultiVCDiff vastly outperformed the other two baselines, confirming its overall reliability in predicting whether a gene is upregulated or downregulated. This strong performance across a diverse set of evaluation criteria in a strict zero-shot setting demonstrated that by learning a joint multimodal space, MultiVCDiff generated transcriptomic profiles with fidelity that surpasses or matches the latest models exclusively specialized for single-cell transcriptomic predictions. This result strongly suggests a synergistic interplay between the morphological and genomic states: incorporating the high-content phenotypic context of a cell enables a more accurate, robust, and generalizable prediction of its transcriptional response to novel perturbations. MultiVCDiff demonstrates superior fidelity in predicting both gene-level and drug-level transcriptional responses for novel perturbations (A) MultiVCDiff is compared against chemCPA, biolord, and PDGrapher across eight distinct metrics from the Cell-Eval framework. Evaluated on the strictly held-out test set of 30 unseen OOD drugs from the CPgenes-LINCS dataset. ∗∗p < 0.01, ∗∗∗p < 0.001, NS, not significant. (B and C) We conduct a granular head-to-head comparison between MultiVCDiff (red) and PDGrapher (blue), the most competitive baseline, across four key metrics. The scatterplots clearly illustrate a predominant advantage for MultiVCDiff, which consistently outperforms PDGrapher in predicting the overlap of DE genes and the DE Spearman log fold change (B), as well as the recall and Pearson delta of DE genes (C). (D) The drug similarity matrix is constructed by calculating the pairwise Pearson correlation between the average gene expression profiles predicted for each of the 30 unseen drugs. (E and F) The structural concordance of the drug similarity matrices shown in (D) is quantitatively assessed using four distinct metrics: PCC and Spearman correlation (E), alongside the Mantel test and RV coefficient (F). Furthermore, we performed a more granular head-to-head comparison between MultiVCDiff and PDGrapher, the most competitive baseline in our evaluation (Figures 5B and 5C). The results clearly showed a predominant advantage for MultiVCDiff across multiple key metrics. For instance, in predicting the overlap of DE genes, MultiVCDiff outperformed PDGrapher in 378 out of 500 evaluated instances, which accounts for 75.6% of the samples. This advantage was consistently pronounced across other crucial metrics, where MultiVCDiff was superior in 70.4% (352/500) of instances for the recall of DE genes, 65.4% (327/500) for Pearson delta, and 62.4% (312/500) for DE Spearman log fold change. This consistent and significant outperformance confirmed the robust and reliable fidelity of MultiVCDiff’s individual gene-level predictions over state-of-the-art single-modality methods. Moving beyond the fidelity of individual gene-level predictions, we next sought to determine if MultiVCDiff could capture higher-order biological structures at the drug level. Accurately modeling the relationships between different chemical perturbations is a more stringent test of a model’s capabilities, as it reflects a deeper understanding of the perturbation landscape rather than just isolated response profiles. This ability is crucial for practical applications such as identifying compounds with similar mechanisms of action or for drug repurposing. To assess this within our strict zero-shot setting, we constructed drug similarity matrices by calculating the pairwise Pearson correlation between the average gene expression profiles predicted for each of the 30 unseen OOD drugs. A qualitative comparison showed that the similarity matrix generated from MultiVCDiff’s predictions strikingly recapitulated the structural patterns of the ground-truth matrix, an alignment that was degraded in the matrices generated by PDGrapher and biolord (Figure 5D). We then quantified this structural concordance using four distinct metrics: Pearson correlation coefficient (PCC), Spearman correlation, the Mantel test,38 and the RV coefficient.39 On all four metrics, MultiVCDiff achieved the highest scores by a significant margin. Specifically, MultiVCDiff demonstrated a 35.6% improvement in PCC and a 39.5% improvement in the Mantel test score over the next-best model, PDGrapher (Figures 5E and 5G). Similar substantial gains were observed in the Spearman correlation (+36.3%) and the RV coefficient (+25.7%). This provided compelling evidence that MultiVCDiff learned a more biologically faithful representation of the perturbation landscape, correctly modeling both the unique transcriptional signature of each novel drug and the subtle, complex relationships between them. Discussion In this study, we sought to advance a central goal of the AIVC initiative: building a generative virtual cell model capable of capturing coordinated cellular responses across multiple molecular and phenotypic layers. Existing perturbation modeling approaches generally fall into restricted paradigms. The most prevalent category focuses on predicting a single downstream modality directly from a given perturbation. For instance, several state-of-the-art models specialize in generating cellular morphology from chemical or genetic interventions (e.g., MorphoDiff, IMPA, StarGAN, DMIT, and MDTv2), while others are exclusively designed to predict transcriptional responses (e.g., chemCPA, biolord, and PDGrapher). While these methods are highly valuable for modality-specific predictions, they fundamentally treat each biological readout in isolation. Consequently, they cannot explicitly construct the intrinsic relationships between transcriptomic and morphological responses. This limitation is particularly critical because a core objective of a virtual cell is not merely to predict individual, disjointed phenotypes, but rather to understand how molecular alterations are translated into higher-order, multimodal cellular states. The second category of methods begins to address this issue by modeling one modality conditioned on another. A representative example is MorphDiff, which uses perturbation-induced transcriptomic profiles to guide the generation of cellular morphology. Such approaches are important because they establish a cross-modal mapping between RNA states and morphological phenotypes, providing a more biologically meaningful view of cellular response. However, this framework also has a critical practical limitation: it requires the post-perturbation gene expression profile as an input during inference. In realistic virtual screening settings, especially for newly designed or untested compounds, post-perturbation transcriptomic measurements are unavailable. As a result, although MorphDiff can model the mapping from RNA to morphology, it cannot be directly deployed for true large-scale screening of novel perturbations. MultiVCDiff was designed to overcome both limitations simultaneously. Rather than predicting only a single modality or relying on an experimentally observed intermediate modality, our framework takes the perturbation itself as the sole input and jointly generates both transcriptomic and morphological responses. In the drug setting, this means that only the molecular structure is required at inference time, without any need for measured post-perturbation RNA or image data. This design provides two key advances. First, by jointly generating both modalities within one unified framework, MultiVCDiff explicitly builds a bridge between transcriptomic and morphological responses, which is essential for virtual cell modeling. Second, because MultiVCDiff does not require any post-perturbation readouts as input, it can infer downstream cellular responses directly from the perturbation itself, using only a drug structure or a gene representation at inference time. This makes the framework directly applicable to realistic in silico virtual screening scenarios, where experimental transcriptomic and morphological measurements are unavailable for newly designed compounds or genetic perturbations. In this sense, MultiVCDiff goes beyond conventional cross-modal conditional generation by enabling multimodal response prediction from the perturbation alone, thereby providing a more practical and conceptually complete framework for virtual screening and virtual cell modeling. A core finding of our work is that jointly modeling cellular morphology and gene expression with MultiVCDiff significantly outperforms specialized single-modality approaches. Through rigorous zero-shot, OOD benchmarking, we demonstrated that MultiVCDiff successfully avoids mode collapse, surpassing eight state-of-the-art baselines in morphological generation as well as top-tier models like chemCPA, biolord, and PDGrapher in transcriptional prediction. This success is rooted in our transformer architecture, where self-attention mechanisms effectively capture cross-modal dependencies within a joint embedding space. Furthermore, high-content morphological information acts as a potent contextual constraint to enhance the accuracy of molecular predictions, while the transcriptomic state ensures biological fidelity in the generated phenotypes. Beyond achieving quantitative accuracy, MultiVCDiff learns fundamental mappings from structural perturbations to actual biological consequences rather than merely memorizing chemical structures. Ultimately, this robust biological interpretability enables highly accurate drug similarity matrix construction and MOA retrieval for completely unseen drugs, establishing MultiVCDiff as a reliable engine for actionable hypothesis generation in drug repurposing and novel target discovery. Limitations of the study While MultiVCDiff represents a significant advance in constructing generative multimodal virtual cells, several limitations and potential caveats of the current framework must be acknowledged. First, MultiVCDiff predicts cellular responses as static endpoints (e.g., at a fixed 24-h post-treatment mark). However, biological responses are inherently continuous and dynamic. The current model cannot seamlessly simulate the temporal kinetics of morphological and transcriptomic evolution. Second, MultiVCDiff currently does not explicitly encode dosage information or combinatorial effects. However, extending our framework to accommodate these factors is highly feasible. Because MultiVCDiff is built upon a flexible conditional latent diffusion model powered by a Transformer backbone, future iterations could seamlessly model dose-dependent effects by projecting specific dosage values into dense embeddings and concatenating them with our existing structural perturbation embeddings. Similarly, combinatorial therapies could be straightforwardly handled by aggregating multiple distinct perturbation embeddings, such as through an attention-based fusion module or direct vector addition to form a unified conditional guidance signal for the denoising transformer. Third, while MultiVCDiff implicitly captures cross-modal relationships through its self-attention mechanism, it does not currently integrate explicit biological priors, such as gene regulatory networks. These underlying networks fundamentally govern how cells behave and adapt following external perturbations. Therefore, incorporating advanced inference frameworks like KINDLE40 or CellOracle41 could enhance our model’s causal interpretability. Resource availability Lead contact Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Jinzhuo Wang (wangjinzhuo@pku.edu.cn). Materials availability This study did not generate new, unique reagents. Data and code availability •All datasets analyzed in this study are publicly available via their respective official repositories: BBBC021 (https://bbbc.broadinstitute.org/BBBC021), CDRP (https://github.com/gigascience/paper-bray2017), JUMP (https://github.com/jump-cellpainting/datasets/), and LINCS (https://github.com/broadinstitute/lincs-cell-painting). The processed and curated CPgenes dataset generated for this study is deposited at Zenodo: https://doi.org/10.5281/zenodo.19640132. All datasets analyzed in this study are publicly available via their respective official repositories: BBBC021 (https://bbbc.broadinstitute.org/BBBC021), CDRP (https://github.com/gigascience/paper-bray2017), JUMP (https://github.com/jump-cellpainting/datasets/), and LINCS (https://github.com/broadinstitute/lincs-cell-painting). The processed and curated CPgenes dataset generated for this study is deposited at Zenodo: https://doi.org/10.5281/zenodo.19640132. •The source code of MultiVCDiff is available at GitHub (https://github.com/prsigma/MultiVCDiff) and archived at Zenodo: https://doi.org/10.5281/zenodo.19639206. The source code of MultiVCDiff is available at GitHub (https://github.com/prsigma/MultiVCDiff) and archived at Zenodo: https://doi.org/10.5281/zenodo.19639206. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Acknowledgments This research was supported by the National Key Research and Development Program of China (2024YFF0507404) and Prevention and Control of Emerging and Major Infectious Diseases-National Science and Technology Major Project (2025ZD01901802). Author contributions Conceptualization, P.R. and J.W.; data acquisition, P.R. and J.W.; analysis, P.R., Z.L., and J.W.; methodology, P.R. and J.W.; supervision, J.W.; funding acquisition, J.W.; project administration, J.W.; writing, P.R., Y.G., and J.W. Declaration of interests The authors declare no competing financial interests. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERDeposited dataBBBC021Caie et al.16https://bbbc.broadinstitute.org/BBBC021CDRPBray et al.17https://github.com/gigascience/paper-bray2017/blob/master/download_cil_images.shJUMPChandrasekaran et al.18https://github.com/jump-cellpainting/datasets/LINCSNatoli et al.19https://github.com/broadinstitute/lincs-cell-paintingCPgenesThis paperhttps://doi.org/10.5281/zenodo.19640132Software and algorithmsPython version 3.10Python Software Foundationhttps://www.python.orgScalable Diffusion Models with TransformersPeebles et al.21https://github.com/facebookresearch/DiTMultiVCDiffThis paperhttps://doi.org/10.5281/zenodo.19639206 Method details Problem formulation of MultiVCDiff The goal of MultiVCDiff is to learn the conditional joint probability distribution p(I, R∣P) of a cellular morphology image I and a transcriptomic profile R, given a specific perturbation P. To achieve this, we employ a latent diffusion model that operates on a unified, multimodal representation of the cell state. The model is trained to reverse a forward diffusion process that incrementally adds Gaussian noise to the latent embeddings, thereby learning to generate coherent, paired data from a random noise distribution, conditioned on the perturbation. Multimodal representation of MultiVCDiff We define a single data sample as a triplet (I,R,P), where I∈RH×W×C is a cellular morphology image, R∈RNgenes is the corresponding gene expression vector, and P represents the perturbation. Our objective is to model the conditional distribution Pθ(χimage0,χrna0|χperturb), where χ0 denotes the initial, noise-free latent embeddings derived from the data, and χperturb is the embedding of the perturbation. By learning this conditional joint probability distribution, our model Pθ can simultaneously generate paired data for both modalities by sampling from this distribution. Crucially, this joint modeling approach enables the model to implicitly learn the intricate interrelationships between morphological and transcriptomic responses, capturing how molecular changes influence as phenotypic outcomes under specific perturbational contexts. To create a shared representational space, we first encode the image and RNA modalities into latent vectors of a fixed dimension Dlatent. The cell image I is encoded using a pretrained ViT encoder ℇimage, which transforms it into a sequence of patch embeddings χimage0=ℇimage(I)∈RNimg×Dlatent, where Nimg is the number of image patches. Concurrently, the gene expression profile R is encoded via a dedicated MLP encoder ℇrna, yielding χrna0=ℇrna(R)∈RNrna×Dlatent, where Nrna is the sequence length of the RNA embedding (Figure 1B). These two sequences are then concatenated to form the initial joint multimodal latent representation χmm0=[χimage0;χrna0]∈R(+Nrna)×Dlatent, where [·; ·] denotes concatenation along the sequence dimension. The conditioning signal is derived from a versatile perturbation encoder, ℇperturb, designed to accommodate both chemical and genetic interventions (Figure 1C). For chemical perturbations, which are represented as SMILES strings, we utilize the RDKit toolkit to compute Morgan fingerprints. These fingerprints effectively capture the structural properties and functional groups of the small molecules. For genetic perturbations, the target genes are encoded using scGPT, a large-scale generative foundation model pretrained on massive single-cell data, to extract context-rich genetic embeddings. These modality-specific features are subsequently projected through linear layers into a dense embedding χperturb=ℇperturb(P)∈RNpert×Dlatent. This embedding provides the necessary context to guide the generative process, enabling the model to produce outputs specific to the drug’s structural properties. Training process of MultiVCDiff The training of MultiVCDiff is structured around a forward-diffusion process that adds noise and a learned reverse process that removes it. The model learns by optimizing an objective derived from the variational bound on the negative log likelihood. Forward diffusion process The training process is anchored by a fixed forward diffusion process q, which incrementally corrupts the initial multimodal latent embedding χmm0 over T discrete timesteps. This process is defined as a Markov chain that gradually adds Gaussian noise according to a pre-defined variance schedule {βt∈(0,1)}t=1T: A key property of this Markov chain is that we can sample the latent state χmmt at any arbitrary timestep t in a closed form, conditioned only on the initial state χmm0. Using the notation αt = 1–βt and α¯t=∏s=1tαs, the distribution of χmmt is given by: This allows us to directly generate a noisy sample for any timestep t by sampling a standard Gaussian noise variable ϵ∼N(0,I) and applying the reparameterization: This equation represents the explicit noising process, which provides the noisy inputs for training the model. Learned reverse process and objective function The objective of training is to learn a reverse process pθ(χmmt−1|χmmt), that can invert the diffusion, effectively learning to denoise the corrupted latents. This is achieved by optimizing the variational bound on the negative log likelihood, which can be expressed as: Since the forward process is fixed, the first term is a constant and can be ignored during training. The core of the learning lies in the KL divergence terms, which train the reverse process pθ. Each term for t > 1 compares the reverse process transition pθ(χmmt−1|χmmt)=N(χmmt−1;μθ(χmmt,t,χperturb),∑θ(χmmt,t,χperturb)) with the tractable forward process posterior q(χmmt−1|χmmt,χmm0). Following the DDPM framework,22 we set the covariance to untrained constants ∑θ=σt2I and parameterize the mean μθ to predict the noise ϵ that was added during the forward process: Here, ϵθ is our denoising Transformer. This parameterization transforms the KL divergence terms in the loss into a weighted mean squared error between the true and predicted noise. For simplicity and improved performance, we use a simplified objective that removes these weights: This objective function directly trains the model ϵθ to predict the noise ϵ from the noised latent χmmt, conditioned on the timestep t and the perturbation χperturb. Inference process of MultiVCDiff The inference process generates a novel, paired sample (Iˆ,Rˆ) by reversing the diffusion process. The procedure begins by sampling an initial latent variable from the prior distribution χmmT∼N(0,I). Given a perturbation embedding χperturb, the model then iteratively denoises this latent variable for t = T,T-1, …,1, using the learned conditional distribution pθ(χmmt−1|χmmt). Each step of this reverse Markov chain computes a slightly less noisy latent χmmt−1 from the previous latent χmmt. Using our noise-prediction parameterization, the update step is given by: where z∼N(0,I) is a random Gaussian noise sample for t > 1, and z = 0 for t = 1. The term ϵθ(χmmt,t,χperturb) is the output of our trained denoising Transformer, which predicts the noise component in χmmt. The coefficients αt and α¯t are derived from the fixed variance schedule {βt∈(0,1)}t=1T of the forward process. The standard deviation of the added noise δt is set to β˜t=1−α¯t−11−α¯tβt. This iterative procedure resembles annealed Langevin dynamics, where the model ϵθ effectively provides a learned score function to guide the sampling trajectory. After T steps, this process yields the denoised joint latent representation χmm0. This is then de-concatenated into its constituent unimodal components [χˆimage0;χˆrna0]=χmm0. Finally, these latent representations are passed through their respective decoders Dimage and Drna, to reconstruct the final high-fidelity outputs: Iˆ=Dimage(χˆimage0) and Rˆ=Drna(χˆrna0). Training details MultiVCDiff is built upon a cross-modal Transformer architecture. The Transformer backbone consists of 28 blocks, with a hidden size of 1152 and 16 attention heads. The model is designed to handle multimodal inputs, starting with cellular morphology images (512 × 512 pixels), transcriptomic profiles (977 landmark genes), and specific perturbations (encompassing both chemical and genetic interventions). Each modality is first projected into a shared 1152-dimensional latent space before being processed by the Transformer backbone. Specifically, morphology images are passed through an encoder to obtain a 4 × 32 × 32 latent map, which is subsequently converted into a sequence of tokens. The 977-dimensional gene expression vector and the modality-specific perturbation features, either 1024-bit Morgan fingerprints for small molecules or 512-dimensional scGPT embeddings for target genes, are transformed into 1152-dimensional embeddings through their respective linear projection layers. The Transformer is then conditioned on the diffusion timestep and the resulting unified perturbation embedding using an adaptive layer norm zero (adaLN-Zero) mechanism. All models are implemented in PyTorch. The training is conducted on a high-performance computing cluster equipped with eight NVIDIA A100 GPUs, each with 80GB of VRAM. We use the AdamW optimizer with a constant learning rate of 1e−4. The model is trained for a total of 100,000 steps, which takes approximately 3–4 days to complete. Construction of the CPgenes datasets To power the development of a multimodal virtual cell, we construct the CPgenes datasets, a comprehensive, large-scale, trimodal corpus designed for generative modeling. This collection systematically bridges two powerful high-throughput screening technologies by identifying shared perturbations between four extensive public Cell Painting resources for cellular morphology (the BBBC021, CDRP, JUMP, and LINCS datasets) and the L1000 Connectivity Map (CMap) project for gene expression. The morphology data originated from distinct biological contexts: MCF7 breast cancer cells (BBBC021), U2OS osteosarcoma cells (CDRP and JUMP), and A549 lung cancer cells (LINCS). To extract high-resolution cellular crops, the raw multichannel images from these collections were carefully preprocessed. For the BBBC021 dataset, the DAPI, tubulin, and actin channels were merged into the blue, green, and red channels of a single RGB image, respectively. For the CDRP, JUMP, and LINCS datasets, the full five-channel images were utilized to maximize phenotypic information. All composite images were segmented and cropped to a uniform size of 512 × 512 pixels. The two modalities were seamlessly integrated by matching a shared set of small-molecule perturbagens and genetic interventions present in both assays, yielding four harmonized sub-datasets (CPgenes-BBBC021, CPgenes-CDRP, CPgenes-JUMP, and CPgenes-LINCS). To create a unified transcriptomic signature for each perturbation, we retrieved all corresponding gene expression profiles from the L1000 database and computed their average, resulting in a single, representative vector for each chemical compound or genetic target. This averaged RNA expression profile was then assigned to every morphological sample associated with that specific perturbation. To ensure a sufficiently large and balanced dataset for deep learning, the collection of processed images for each perturbation was augmented to generate an adequate number of distinct samples. This rigorous curation process resulted in a massive final corpus of matched, trimodal samples, which was subsequently partitioned into training and testing sets using an 8:2 ratio. Each sample in this integrated dataset tightly links three components: the specific perturbation embedding (chemical or genetic), a high-resolution morphological image, and the corresponding averaged gene expression profile, providing the necessary foundation for training MultiVCDiff. Image evaluation metrics Fréchet Inception Distance The FID provides a sophisticated measure of similarity between a collection of real images and those produced by a generative model. Rather than comparing raw pixel values, it evaluates perceptual quality by analyzing the statistical resemblance of high-level features extracted from the images. The methodology leverages the feature extraction capabilities of the Inception V3 neural network,42 which is pre-trained on the extensive ImageNet database.43 For every real and generated image, a 2048-dimensional embedding is produced by capturing the activations from the network’s final pooling layer. This process yields two distinct sets of feature vectors: one for the real images and one for the generated ones. These collections of feature vectors are then treated as samples from two distinct multivariate normal distributions. The FID is calculated as the Fréchet distance between these two Gaussians, a measure that considers both the difference in their means (mr for real, mg for generated) and their covariance matrices (Cr,Cg). The formal expression is: Here, the first term captures the squared distance between the mean feature vectors, while the second term, involving the trace (Tr), accounts for the differences in the covariance structures. Ultimately, a smaller FID value signifies that the statistical properties of the generated image features are more closely aligned with those of the real images, implying a higher degree of realism and quality in the generated output. Kernel Inception Distance The KID is another metric used to quantify the dissimilarity between the probability distributions of real and generated images. Similar to FID, it operates on feature vectors extracted from a pre-trained Inception network. However, instead of assuming the features follow a Gaussian distribution and comparing their moments, KID measures the distance between distributions by comparing sets of independently drawn samples from each. It computes the squared Maximum Mean Discrepancy between the feature sets. As with FID, a lower KID value signifies a better alignment between the generated and real image distributions, indicating higher-quality image synthesis. RNA evaluation metrics To systematically assess our model’s fidelity in predicting transcriptomic responses, we adopt metrics from the Cell-Eval framework. This framework provides a comprehensive quantitative suite for evaluating model performance on key experimental outputs, including gene expression and differential expression statistics. DE overlap accuracy This metric offers a direct measure regarding the identification of specific genes affected by a perturbation. It quantifies the intersection between the model’s top k predicted DE genes and the true top k DE genes from experimental data. A higher value therefore indicates that the model not only captures global expression trends but also prioritizes the correct targets at the top of its ranked list. Top-k precision This metric directly evaluates the precision of the model’s most confident predictions. It is defined as the fraction of the top k predicted DE genes that are truly significant in the experimental data. This metric measures the reliability of the genes it flags as most important. Direction match This metric verifies the direction match of expression changes. For the set of genes identified as significant by both the model and the ground truth, this metric calculates the fraction for which the predicted and observed fold changes share the same sign (i.e., both are up-regulated or both are down-regulated). Spearman correlation of log fold changes This metric assesses the overall correlation between predicted and observed log fold changes by computing the Spearman correlation for all genes. This provides a broad measure of whether the model accurately captures the relative up- or down-regulation across the full set of measured genes. Recall This metric is used to evaluate the model’s ability to identify the complete set of truly affected genes. Specifically, recall quantifies the fraction of all genuinely significant DE genes that the model successfully recovers. A high recall is essential for ensuring that critical biological signals are not overlooked. AUPRC and AUROC These metrics evaluate the model’s ability to correctly classify genes as either significant or non-significant. The AUPRC is particularly important as it measures the trade-off between precision and recall. This is especially informative in scenarios where significant DE genes are sparse, reflecting the model’s ability to recover true positives with high precision. Pearson delta correlation This metric assesses a model’s ability to accurately predict the change in gene expression induced by a perturbation. It calculates the Pearson correlation coefficient between the predicted and observed “expression deltas”, which are defined as the difference in transcriptomes between the perturbed cells and the control cells. By focusing specifically on the perturbation effect, this metric provides a direct evaluation of how well the model captures the biological impact of the perturbation itself. Cellprofiler feature construction To quantitatively capture the rich phenotypic information contained in the Cell Painting images, we developed and applied a custom image analysis pipeline using CellProfiler (v5.0). This pipeline was designed to perform robust segmentation of cellular compartments and extract a comprehensive set of morphological features. The analysis began with the identification and segmentation of three cellular compartments: nuclei, cells, and cytoplasm. First, nuclei were identified as primary objects from the blue channel images, corresponding to DNA staining. Second, the cell boundaries were delineated as secondary objects by propagating outwards from the identified nuclei, using the green channel to define the cell edges. Finally, the cytoplasm was defined as the tertiary object by subtracting the nuclear region from the corresponding cell region. A critical quality control step was integrated into the pipeline to filter out and exclude any cells or nuclei touching the border of the image, ensuring that all measurements were derived from complete, intact cells. Following segmentation, we executed a suite of CellProfiler measurement modules to extract an extensive set of features from each identified compartment (nuclei, cytoplasm, and cells) across all three channels. The extracted features encompassed several categories. Morphological measurements included basic descriptors of size and shape such as area, perimeter, major and minor axis length, eccentricity, and solidity. Intensity features comprised statistics on the distribution of pixel intensities within each compartment, such as mean and median intensity, standard deviation, median absolute deviation, and intensity quartiles. To quantify patterns and spatial variations in staining, texture features were calculated using Haralick metrics, including contrast, correlation, entropy, and sum variance, at multiple spatial scales. The pipeline also captured relational and distributional features by measuring the spatial relationships between cells, the correlation of signals between different channels, and the radial distribution of intensity from an object’s center to its edge. Per-object measurements were then aggregated across all cells in an image by calculating the mean, median, and standard deviation to form a single profile for each image. By applying this comprehensive pipeline, we successfully extracted a total of 6,345 distinct morphological features for each image, creating a detailed quantitative profile that captures the subtle effects of drug perturbations on cellular phenotype. Matrix similarity comparison metrics Pearson correlation coefficient This metric is used to evaluate the model’s ability to capture the linear relationships between the predicted and ground-truth drug similarity matrices. Specifically, after flattening the values from each matrix into a one-dimensional vector, the PCC quantifies the strength and direction of the linear association between these two resulting vectors. Spearman correlation This metric evaluates the model’s ability to correctly preserve the rank-order of drug similarities. To calculate this, the similarity values from each matrix are flattened into one-dimensional vectors. Spearman correlation then assesses the monotonic relationship between these two vectors by comparing the ranks of their corresponding elements. Mantel test This test is used to formally assess the statistical significance of the correlation between the two complete similarity matrices. The Mantel test quantifies the overall association between the two matrices and uses permutations to determine if this correlation is stronger than what would be expected by chance. A significant result provides robust statistical evidence that the model has learned the true underlying structure of the drug or gene perturbation landscape. RV coefficient This metric is used to evaluate the global similarity between the two high-dimensional drug similarity matrices, acting as a multivariate extension of the squared correlation coefficient. The RV coefficient measures the overall congruence between the geometric configurations of the two matrices, yielding a single score between 0 and 1. A high RV coefficient demonstrates that the multidimensional structure of the predicted drug landscape is highly consistent with the ground-truth landscape, confirming the model’s ability to capture complex multivariate relationships. Quantification and statistical analysis Statistical tests and significance Statistical analyses were performed using Python 3.10 with the SciPy (v1.10.1) and statsmodels (v0.13.5) libraries. To compare the performance metrics (e.g., image quality scores and transcriptomic Cell-Eval scores) between MultiVCDiff and baseline models, two-sided Student’s t-tests were employed for normally distributed data, while Wilcoxon rank-sum tests (Mann-Whitney U tests) were utilized when data did not meet the assumptions of normality. Assumptions of normality and homogeneity of variances were assessed using the Shapiro-Wilk test and Levene’s test, respectively, prior to selecting the appropriate statistical approach. For GO enrichment analysis, significance was calculated using the hypergeometric test with Benjamini-Hochberg procedure for False Discovery Rate (FDR) correction. Statistical significance was defined as a p-value <0.05. In all figures, significance levels are denoted using asterisks: ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001. NS. indicates not significant (p ≥ 0.05). In the text and bar plots, data are generally presented as mean ± standard deviation to indicate dispersion and precision. For boxplots, the center line indicates the median, the box limits represent the upper and lower quartiles (75th and 25th percentiles), and the whiskers extend to 1.5 times the interquartile range. Points beyond the whiskers are considered outliers. Definition of sample size n In morphological image evaluations (e.g., FID, KID, and CellProfiler feature distributions), n represents the number of independent generated samples or real ground-truth samples per perturbation condition (typically n = 500). In transcriptomic evaluations (e.g., Cell-Eval metrics), n represents the number of distinct perturbation conditions evaluated (e.g., n = 30 strictly held-out out-of-distribution novel drugs). Published: May 21, 2026 Footnotes Supplemental information can be found online at https://doi.org/10.1016/j.crmeth.2026.101459. Supplemental information Document S1. Figures S1–S3 Document S2. Article plus supplemental information References Associated Data Supplementary Materials Document S1. Figures S1–S3 Document S2. Article plus supplemental information Data Availability Statement •All datasets analyzed in this study are publicly available via their respective official repositories: BBBC021 (https://bbbc.broadinstitute.org/BBBC021), CDRP (https://github.com/gigascience/paper-bray2017), JUMP (https://github.com/jump-cellpainting/datasets/), and LINCS (https://github.com/broadinstitute/lincs-cell-painting). The processed and curated CPgenes dataset generated for this study is deposited at Zenodo: https://doi.org/10.5281/zenodo.19640132. All datasets analyzed in this study are publicly available via their respective official repositories: BBBC021 (https://bbbc.broadinstitute.org/BBBC021), CDRP (https://github.com/gigascience/paper-bray2017), JUMP (https://github.com/jump-cellpainting/datasets/), and LINCS (https://github.com/broadinstitute/lincs-cell-painting). The processed and curated CPgenes dataset generated for this study is deposited at Zenodo: https://doi.org/10.5281/zenodo.19640132. •The source code of MultiVCDiff is available at GitHub (https://github.com/prsigma/MultiVCDiff) and archived at Zenodo: https://doi.org/10.5281/zenodo.19639206. The source code of MultiVCDiff is available at GitHub (https://github.com/prsigma/MultiVCDiff) and archived at Zenodo: https://doi.org/10.5281/zenodo.19639206. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.