GeneCMamba: Autoregressive spatial transcriptome prediction from histopathology via state-space sequence modeling and adaptive cross-modal gating

Spatial transcriptomics provides spatially resolved molecular profiles but remains constrained by cost, throughput, and technical variability. Predicting spatial gene expression from routine hematoxylin-and-eosin (H&E) images offers a scalable alternative, although cross-modal alignment remains challenging because morphology and transcription are probabilistically related and represented at different scales. We propose GeneCMamba, a contrastive learning-guided generative framework that formulates spatial gene expression prediction as autoregressive gene-token generation. A dual-encoder architecture first aligns histopathological and transcriptomic representations in a shared latent space. A state-space model decoder then captures long-range dependencies among gene tokens with linear computational complexity, while a cross-modal gating mechanism uses gene queries to retain informative morphological features and suppress staining artifacts and irrelevant background. Experiments on six tissue datasets comprising approximately 23,500 spatial samples showed that GeneCMamba outperformed all evaluated baselines. On the breast dataset, it achieved a token accuracy of 82.42% and an F1 score of 0.78 and converged five epochs earlier than the comparison models. Ablation studies confirmed the complementary contributions of the decoder and gating mechanism. Moreover, spatial reconstruction of FN1 reproduced heterogeneous expression patterns in stroma-rich regions, supporting the biological relevance of the predictions. GeneCMamba therefore provides an efficient framework for inferring spatial molecular profiles from routine histopathology images.