Abstract
Spatial Transcriptomics is a novel technology that aligns histology imageswith spatially resolved gene expression profiles. Although groundbreaking, itstruggles with gene capture yielding high corruption in acquired data. Givenpotential applications, recent efforts have focused on predictingtranscriptomic profiles solely from histology images. However, differences indatabases, preprocessing techniques, and training hyperparameters hinder a faircomparison between methods. To address these challenges, we present asystematically curated and processed database collected from 26 public sources,representing an 8.6-fold increase compared to previous works. Additionally, wepropose a state-of-the-art transformer based completion technique for inferringmissing gene expression, which significantly boosts the performance oftranscriptomic profile predictions across all datasets. Altogether, ourcontributions constitute the most comprehensive benchmark of gene expressionprediction from histology images to date and a stepping stone for futureresearch on spatial transcriptomics.