New AI Model Predicts Gene Expression from Tissue Images

Researchers have introduced a novel computational approach for predicting spatial gene expression from routine histology images, addressing the high cost and limited scalability of spatial transcriptomics (ST). The method, detailed in a preprint on arXiv (2608.14330), reformulates morphology-to-transcriptomics prediction as a conditional generation problem in transcriptional program space, rather than predicting individual genes independently. By using consensus non-negative matrix factorization (cNMF), the model extracts a low-dimensional set of transcriptional programs that capture coordinated expression variation in training data, then trains a conditional diffusion model to generate these programs from tissue morphology. This approach leverages established transcriptomic modeling practices and avoids the heterogeneous gene selection strategies that complicate fair comparisons across existing methods. The work aims to improve the accuracy and generalizability of predicting gene expression from histology, potentially enabling broader clinical and research applications where ST is impractical. The preprint was announced as a new submission on arXiv, with the identifier 2608.14330.