Abstract PS2-09-05: Feasibility of Virtual RNA Inference from Histopathology to Identify Gene Signatures of Neoadjuvant Chemotherapy and Immune Checkpoint Inhibitor (ICI) Response in Triple-Negative Breast Cancer (TNBC)

Abstract Background: TNBC remains a biologically aggressive subtype with limited biomarkers to predict pathologic response in the neoadjuvant setting. While spatial transcriptomics (ST) can reveal features of the tumor microenvironment (TME) associated with response and resistance, its clinical application is limited by cost and scalability. As part of ongoing efforts to investigate mechanisms of resistance to neoadjuvant chemotherapy and ICI in breast cancer, we evaluated the feasibility of virtual RNA inference (VRI)—a deep learning approach for prediction of spatial gene expression from routine H&E-stained slides. We hypothesized that inferred ST patterns localized to the tumor could predict pathologic response. Methods: To develop VRI, development cohort of 25 TNBC cases from Dartmouth Health was assembled, comprising primarily of surgical specimens with matched Visium ST data. All slides were sectioned at 5 μm and scanned at 40x (Leica Aperio GT450). A Vision Transformer model (UNI), pretrained on large-scale histopathology datasets, was fine-tuned to predict expression values for a 1,000-gene panel at the level of individual Visium spots (∼55 μm resolution), resulting in over 125,000 paired image-expression samples. Spearman correlations were computed per gene using cross-validation.For the validation cohort, H&E slides from retrospective TNBC resections (Cedars-Sinai, 20x, Leica Aperio GT450, Hamamatsu Nanozoomer S360, 01/2018-07/2024) were collected from patients who received neoadjuvant chemotherapy±ICI (N=58 total, 28 pCR, 20 ICI). VRI was applied to predict spatial gene expression. A deep-learning-based semantic segmentation model trained on over 2,000 annotated cases was used to segment tumor and various TME compartments (e.g., tumor stroma, tumor infiltrating lymphocytes). Inferred gene expression was averaged over intratumoral regions to generate patient-level profiles. Differential gene expression, adjusting for ICI, cross-validation, and pathway analysis were used to identify spatially localized gene signatures associated with complete vs. partial pathologic response to ICI. Results: In the development cohort, VRI achieved a median Spearman correlation of 0.55 across 1,000 genes, demonstrating reliable inference of ST from histology. Application to the Cedars-Sinai validation cohort identified a tumor-specific gene expression signature associated with complete pathologic response. A few of many notable predictive genes included LAD1, MDK, Ki67, and CDKN2A, enriched in hallmark cancer pathways such as E2F targets (p < 0.001), G2M checkpoint (p < 0.001), and epithelial-mesenchymal transition (p = 0.03), among others. A composite score derived from these pathways effectively stratified responders from non-responders, achieving an AUC of 0.82 (sensitivity=0.77, specificity=0.73). Ongoing works include expansion of include cohort size, additional subgroup analysis to factor in treatment regimens (e.g., ICI), and adjust for clinical covariates such as stage, grade, residual tumor, and tumor cellularity. Conclusion: This pilot study demonstrates the feasibility of using VRI to infer tumor-specific gene expression patterns predictive of neoadjuvant chemotherapy response from standard H&E slides in TNBC. Inferred spatial transcriptomic features capture biologically relevant response signatures without the need for molecular assays. Future work will formulate inferred ST into spatially resolved TME features—such as immune cell distribution and stromal context—that may inform drug response and guide future biomarker discovery. Citation Format: Y. Yuan, V. Pujara, J. Lownik, G. Murray, J. Bitar, L. Chen, P. Najafzadeh, K. Dabirian, D. Lin, F. Kolling IV, J. Marotti, L. Vaickus, K. Yao, L. Vahdat, K. Le, Z. Azher, J. Levy. Feasibility of Virtual RNA Inference from Histopathology to Identify Gene Signatures of Neoadjuvant Chemotherapy and Immune Checkpoint Inhibitor (ICI) Response in Triple-Negative Breast Cancer (TNBC) [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS2-09-05.