Abstract
Spatial Transcriptomics (ST) integrates histology with spatially resolved gene expression, offering rich insights into tissue architecture and function. However, its clinical and large-scale deployment is hindered by high costs, technical complexity, and limited accessibility. To address this, computational pathology methods have emerged to predict gene expression directly from histology images, typically framing the task as a multi-output regression problem mapping image patches to gene expression profiles. While several Convolutional Neural Network (CNN) models have been proposed, little is known about how performance is influenced by (a) the number of trainable parameters and (b) the patch size used for prediction. Moreover, existing studies rely primarily on quantitative metrics and overlook biological relevance of predictions. In this study, we systematically evaluated multiple convolution based models (including a Vision Transformer (ViT) model) with different patch sizes on the Xenium based Autoimmune Machine Learning Challenge (AMLC) dataset. We assessed model performance on both globally expressed genes and subsets enriched for immune or disease associated pathways. Our findings reveal that compact CNNs trained on larger patches outperform deeper models, offering superior accuracy in predicting gene expression, especially for biologically important genes. These insights provide practical guidance for designing efficient and biologically meaningful models in the emerging field of image-based gene expression prediction.