Visualizing the genotype-phenotype link: Predicting drug-induced tissue dynamics with RNA-based diffusion models

Introduction: Oncology drug development is highly challenging because the chance of obtaining FDA approval is only 4.1%. This underscores the need for predictive models that may de-risk this process and improve the odds of success. We hypothesize that accurate modeling of tissue structures based on gene expression profiles may shed light on genotype-to-phenotype relationships that are crucial for uncovering fundamental biological mechanisms and predicting how drugs affect tissue architecture, with prospects in guiding therapeutic strategies and informing early-phase trial design. Here, we present an RNA-aware diffusion model that produces realistic histological states of tissue samples based on rich RNA vector representations, capturing specific changes in gene signatures. Methods: RNA embedding based on expression data of 20,062 genes obtained using a variational autoencoder trained on RNA-seq samples from public sources was integrated via cross-attention to condition the diffusion model. The diffusion model was then fine-tuned on paired H&E-RNA-seq samples from TCGA, CPTAC, and GTEx. Board-certified pathologists were consulted to verify the biological relevance of the generated images and the correspondence between the visual histological structures and altered signatures. Results: Our model generated 512×512-pixel tissue crops (0.5 microns per pixel) with a visually indistinguishable FID of 15. The encoder performance metrics were an MSE of 0.0008 and a median R2 of 0.885. In a simulated H&E analysis, drug-induced gene expression changes were reflected in the anticipated tissue slides over time, showing the dynamics of tertiary lymphoid structures (TLS) and follicles corresponding to changes in TLS and B-cell signature. As such, the model could capture biologically meaningful tissue-level responses, enabling in silico modeling of tissue alterations resulting from specific gene or signature modifications, including those induced by therapeutic interventions. Conclusion: By translating gene expression profiles into depictable tissue structures, our model enables the prediction of how drugs or drug combinations impact tissue morphology through their effects on gene expression. This approach allows us to gather insights into the underlying mechanisms of action for drugs and drug combinations, thereby promoting the discovery of promising single agents or combinations, such as immune checkpoint inhibitors, T- or NK-cell engagers, or PD-1/VEGF bispecific antibodies, that are tailored to specific diagnoses. Our model is poised to improve drug candidate selection, optimize study design, and reduce drug development costs, aiding both preclinical discovery and early-phase clinical development.