Abstract 4189: Quantifying uncertainty in virtual spatial transcriptomics using Bayesian neural networks

Abstract Introduction: Spatial transcriptomics (ST) enables whole-transcriptome profiling within intact tissue architecture but remains costly and batch-dependent. Deep-learning-based virtual RNA inference from hematoxylin and eosin (H&E) slides offers a way to collect ST at scale without profiling, yet the reliability of such predictions is unclear. Uncertainty may arise from assay noise, biological limits on what morphology can reveal, or variation in gene abundance across tissue structures. Uncertainty estimation complements accuracy by helping distinguish these sources and clarify why some genes are more challenging to infer. Here, we integrate virtual RNA inference with Bayesian neural networks (BNN) to quantify and characterize uncertainty in spatial transcriptomic predictions. Methods: An in-house dataset of 65 paired H&E-stained colorectal and triple-negative breast cancer whole-slide images with matched Visium ST data yielded 289,569 spots and corresponding 512×512-pixel H&E patches for 991 spatially variable genes. Using 5-fold patient-level cross-validation, a BNN with Bayesian convolutional and multilayer perceptron layers was trained via variational inference, with Kullback-Leibler divergence regularizing weight distributions. During inference, predictive posterior sampling produced means and log-variances for each gene from image patches, enabling separation of epistemic (uncertainty reducible with more data or improved modelling) and aleatoric (irreducible noise driven by biological variability or assay limits) uncertainty. Gene set enrichment was used to characterize genes ranked by uncertainty. Results: High epistemic uncertainty genes were enriched for signalling and regulatory pathways such as Histone H4-K12 acetylation and positive regulation of apoptosis. High aleatoric uncertainty mapped to stress-response programs including cellular response to ionizing radiation, medium-chain fatty acid biosynthesis, and myeloid cell activation. Low-uncertainty genes were dominated by stable metabolic and cell-cycle pathways, including aerobic respiration, electron transport, and mitochondrial NADH→ubiquinone transport. Conclusion: These results suggest that the uncertainty framework highlights gene programs that are more amenable to histology-based inference, with metabolic pathways exhibiting the most reliable morphological association. Further algorithmic refinement and external validation will be key to establishing the utility of uncertainty modelling for cohort design in large-scale ST studies. Citation Format: Thanosan Prathifkumar, Nehan Mohamed, Anvith Kakkera, Zarif Azher, Xiaoying Liu, Fred Kolling, Laurent Perreard, Scott Palisoul, Louis J. Vaickus, Joshua Jay Levy. Quantifying uncertainty in virtual spatial transcriptomics using Bayesian neural networks [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 4189.