Abstract 4218: ProteoBridge: Bridging skipped sections via histology-based protein prediction

Abstract Background: The progress of spatial multimodal platforms has enabled high-resolution mapping of various types of molecular and morphological characteristics within a tissue. Proteins influence cellular phenotypes and the organization of the tumor microenvironment, with their abundance and patterns highly correlated in adjacent serial sections. Spatial Multimodal platforms such as Singular Genomics G4X, which capture H&E-stained images, multiplexed protein expression, and targeted RNA transcriptomics from the same tissue slide, now create a foundation for three-dimensional (3D) spatial multimodal profiling and 3D tumor microenvironment modeling. However, comprehensive multiplex protein imaging across entire tissue stacks is costly, labor-intensive, and impractical, leading to sparse proteomic sampling that limits accurate 3D reconstruction. Methods: We developed ProteoBridge, a deep learning-based framework that predicts protein expression patterns across serial tissue sections to densify the proteomic information needed for 3D molecular modeling. For each tissue stack, the model was trained on the first section, which contains H&E, multiplex protein images, and RNA transcriptomics. The inputs include H&E tiles and RNA transcriptome data, while the outputs are multi-channel protein maps for the same field of view. The combination of targeted RNA transcriptomics and H&E stained morphology helps define cell types and states, allowing the model to learn morphology-to-protein relationships based on transcriptional programs. The training focused on reducing mean absolute error across protein channels to establish histology-protein correspondences for unmeasured sections. Results: We evaluated ProteoBridge on tissue stacks with multiple profiled sections. Image-level similarity between predicted and measured protein maps was assessed. Using single-slide supervision, ProteoBridge accurately reproduced marker intensities and preserved cross-plane consistency, capturing both intensity values and spatial organization of proteins.. By filling in skipped planes, ProteoBridge generates a more continuous 3D representation of the tumor proteome suitable for downstream 3D molecular modeling and visualization. Conclusion: ProteoBridge requires only routine H&E staining on the remaining sections in a stack to infer protein intensities across serial planes, reducing cost and turnaround time while expanding effective proteomic coverage. Using the G4X platform's multimodal data and extending protein predictions to unmeasured areas allows for cost-effective 3D reconstruction of the tumor proteome, enhancing molecular modeling of the tumor microenvironment and its heterogeneity.Generative AI assisted only with language editing of this abstract. The authors retain sole responsibility for the scientific content and conclusions, having reviewed and approved the final version. Citation Format: Minji Kim, Sunho Park, Seock-Jin Chung, Inyeop Jang, Jean R. Clemenceau, Soyoung Im, Eric Sha, Hwanil Choi, Soonyoung Lee, Jongseong Jang, Sam C. Wang, Tae Hyun Hwang. ProteoBridge: Bridging skipped sections via histology-based protein prediction [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 4218.