Deep learning-based 3D spatial transcriptomics with X-Pression

Abstract

Spatial transcriptomics technologies currently lack scalable and cost-effective options to profile tissues in three dimensions. Technological advances in microcomputed tomography enabled non-destructive volumetric imaging of tissue blocks with sub-micron resolution at a centimetre scale. Here, we present X-Pression, a deep convolutional neural network-based frame-work designed to reconstruct 3D expression signatures of cellular niches from volumetric microcomputed tomography data. By training on a singular 2D section of a paired spatial transcriptomics experiment, X-Pression achieves high accuracy and is capable of generalising to out-of-sample examples. We utilised X-Pression to demonstrate the benefit of 3D examination of tissues on a paired SARS-CoV-2 vaccine efficacy spatial transcriptomics and microcomputed tomography cohort of a recently developed live attenuated SARS-CoV-2 vaccine. By applying X-Pression to the entire mouse lung, we visualised the sites of viral replication at the organ level and the simultaneous collapse of small alveoli in their vicinity. In addition, we assessed the immunological response following vaccination and virus challenge infection. X-Pression offers a valuable and cost-effective addition to infer expression signatures without the need for consecutive 2D sectioning and reconstruction, providing new insights into transcriptomic profiles in three dimensions.