Research Project
Full description Spatial transcriptomics is a breakthrough technology that enables spatially-resolved measurement of molecular profiles in tissues. However, the high cost of generating data has limited its widespread adoption. Predicting gene expression profiles from histology images can be an effective and cost-efficient in-silico spatial transcriptomics solution but is computationally challenging and current methods are limited in model performance. To assess achievements, bottlenecks and potential solutions in this increasingly important field, we systematically review deep learning models developed for predicting gene expression profiles from histology images. We analysed different algorithms, model architectures, and data processing pipelines and assessed how these factors contribute to model performance. We then performed extensive experiments to evaluate the performance variation when these models were applied to in-distribution and out-of-distribution data and across a range of cancer and non-cancer datasets and technologies, suggesting factors that contribute to generalisation/robustness. Lastly, we empirically tested new model approaches and discussed future directions for model improvement. Our results shed insight on key features in a neural network model that either improve or not the performance and generalisation of in-silico spatial transcriptomics, which could contribute immensely to new understanding of tissue biology and new applications in digital pathology.