Predicting condition-aware drug-induced transcriptional responses via a latent diffusion model.

Abstract

Accurate prediction of condition-aware drug-induced transcriptional responses is essential for drug discovery and precision medicine. Current computational models, including encoder-decoder architectures and generative adversarial network-based approaches, exhibit reasonable performance but frequently neglect biological characteristics and fail to generalize to unseen conditions. Thus, this study presents a latent diffusion model that combines a variational autoencoder (VAE) with a diffusion process. The VAE compresses gene expression (GE) profiles into a low-dimensional latent space, where the diffusion process learns the joint probability distribution over latent GE representations and noisy intermediates, thereby enabling more effective capture of gene-gene correlations. The model incorporates multiple perturbation conditions, including cell line, compound, dose, and time, to enhance prediction performance on unseen conditions. The reverse diffusion process predicts both the mean and variance of the posterior distribution, improving the fidelity of the generated GE profiles. The proposed model achieved the highest reconstruction performance in the unseen compound split with a Pearson correlation coefficient of 0.870 ± 0.001 and an R2 score of 0.739 ± 0.001, outperforming previous approaches. The model demonstrated superior preservation of gene-gene correlations, as confirmed by heatmap analysis. To evaluate biological relevance, we predicted half-maximal inhibitory concentration using generated GE, outperforming baseline methods. Latent space analysis revealed that the model preserved cell line identity and continuous dose-time variation. Gene set enrichment analysis confirmed that predicted GE reproduced known pathway-level responses to perturbation. These results demonstrate diffusion-based generative models as effective tools for modeling transcriptional responses in drug discovery and precision medicine. Source code and dataset are available at https://doi.org/10.5281/zenodo.18871024.