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

Key Points

• To quantify and characterize uncertainty in virtual RNA inference using Bayesian neural networks in spatial transcriptomics.
• Utilized a dataset of 65 paired H&E-stained images and Visium ST data, yielding 289,569 spots.
• Employed Bayesian neural networks with convolutional and multilayer perceptron layers for variational inference.
• Used 5-fold patient-level cross-validation for model training and predictive posterior sampling during inference.
• Identified high epistemic uncertainty genes associated with signalling and regulatory pathways.
• Mapped high aleatoric uncertainty to stress-response programs and biological variability.
• Observed that low-uncertainty genes were primarily involved in stable metabolic and cell-cycle pathways.