Spatially Resolved Transcriptomic Correlation Improves the Prediction of Anticancer Drug Candidates

Abstract Predicting the anticancer effects of compounds is critical for computational drug evaluation. However, traditional transcriptomics-based virtual screening relies on bulk RNA profiling, causing intratumoral spatial heterogeneity to be averaged across the entire tissue and thereby limiting predictive accuracy. This limitation is particularly consequential for triple-negative breast cancer, which lacks receptor-guided treatment options and exhibits pronounced spatial heterogeneity that can obscure region-specific therapeutic vulnerabilities when assessed by bulk analysis alone. To address this limitation, we developed a spatially resolved computational framework that integrates spatial transcriptomics to improve the prediction of anticancer drugs. Using a data set of 1291 diverse compounds, we compared the predictive performance of conventional disease-specific bulk signatures with that of localized spatial gene expression signatures. Whereas bulk signatures yielded insufficient predictive performance, incorporating spatial information improved prediction accuracy. Analysis of the predicted drug rankings further demonstrated that the spatial layer substantially improved the prioritization of known anticancer drugs while also identifying unclassified and nononcology compounds with potential anticancer effects. These findings indicate that the predicted therapeutic relevance of a compound depends on localized transcriptomic states rather than representing a static property across an entire heterogeneous tumor section. Ultimately, the proposed method provides a robust, region-aware computational framework for accurately characterizing the anticancer properties of compounds and offers a more nuanced paradigm for transcriptomics-based drug evaluation.