Generalization of deep learning models for predicting spatial gene expression profiles using histology images: A breast cancer case study

Spatial transcriptomics is a breakthrough technology that enables spatially-resolved measurement of molecular profiles in tissues, opening the opportunity for integrated analyses of morphology and transcriptional profiles through paired imaging and gene expression data. However, the high cost of generating data has limited its widespread adoption. Predicting gene expression profiles from histology images only 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 advance research in this emerging and important field, this study makes the following contributions. We first provide a systematic review of deep learning methods for predicting gene expression profiles from histology images, highlighting similarities and differences in algorithm, model architecture, and data processing pipelines. Second, we performed extensive experiments to evaluate the generalization performance of the reviewed methods on several spatial transcriptomics datasets for breast cancer, where the datasets are generated using different technologies. Lastly, we propose several ideas for model improvement and empirically investigate their effectiveness. Our results shed insight on key features in a neural network model that either improve or not the performance of in-silico spatial transcriptomics , and we highlight challenges in developing algorithms with strong generalization performance. Key Messages We comprehensively compared the performance of existing methods for predicting spatial gene expression profiles from histology images We assessed the roles of different algorithms, model architectures, and data processing pipelines to model performance We performed extensive experiments to evaluate the generalization of the models on in-distribution and out-of-distribution spatial transcriptomics datasets We proposed several strategies for improving existing models and empirically investigated their effectiveness