Abstract 1205: Trajectory-aware spatial transcriptomics deconvolution via image representation of RNA-seq data.

Abstract Introduction: Intra-tumor heterogeneity across malignant clones, immune infiltrates and hypoxic niches drives spatially variable radiosensitivity within the same lesion. Spot-based spatial transcriptomics (ST) affords location-resolved gene expression across whole sections, yet each spot aggregates multiple cell types and states, obscuring the micro-ecologies that define the tumour microenvironment (TME). Although single-cell-resolution ST (e.g., imaging-based transcriptomics) can localize transcripts at sub-cellular scales, its cost, assay time and specialized infrastructure currently limit routine deployment in clinical settings. Accurate deconvolution of spot profiles into cell-type/state proportions is therefore essential to derive mechanistic biomarkers such as subtype-specific radiosensitivity scores or hypoxia indices that can guide biologically adaptive radiotherapy. However, existing deconvolution approaches that directly match ST expression to pre-measured reference profiles are vulnerable to technical variation, including batch effects and gene dropouts, particularly when references and ST are generated on different platforms. In this study, we present an optimal transport (OT) framework for ST deconvolution that leverages the geometry of the gene-expression manifold to stabilize estimates, yielding robust quantification of spot composition. Methods: We represent each spot’s gene expression profile as a “genomap” image, ensuring a consistent 2D arrangement of genes across samples. From these per-spot genomap embeddings, we learn a graph using our previously developed “genoTrajectory” method to capture the manifold of spots. To estimate cell-type compositions, we align the graph to a single-cell reference atlas using optimal transport that integrates both gene-expression similarity and manifold geometry, where the loss function consists of transcriptomic dissimilarity between spots and reference cells, and the geodesic distances between the graphs. Preliminary Results: We applied the geometry-aware deconvolution to a breast cancer dataset with paired single-cell-resolution ST (Xenium) and spot-based Visium measurements. Using Xenium as the reference ground truth, our method achieved a mean Pearson correlation of r = 0.68 between predicted per-spot cell-type proportions and the Xenium-derived proportions. Conclusion: Current results demonstrate promising performance for characterizing tumor microenvironment. Future work will expand validation to additional datasets, benchmarking against state-of-the-art methods, and extend analysis to the immune landscape and hypoxia status, with the goal of linking these features to radiotherapy response. Figure 1 Genomaps and genoTrajectory derived from spot-based ST. Citation Format: Junyan Liu, Md Tauhidul Islam, Lei Xing. Trajectory-aware spatial transcriptomics deconvolution via image representation of RNA-seq data [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 1205.