Abstract PS3-06-18: Deep Learning-guided Spatial Transcriptomics Identifies KRAS signaling as the Dominant Mechanism of Resistance to CDK4/6 Inhibitors

Abstract Background: Resistance to CDK4/6 inhibitors remains poorly understood in many patients. While ESR1 mutations and PI3K/AKT pathway alterations have been linked to resistance, these explain only a subset of cases and drugs targeting these mutations show short-lived clinical benefit. We hypothesized that both intratumoral heterogeneity and the tumor microenvironment contribute to therapeutic response, and that AI-based pathology and spatial transcriptomics could uncover dominant resistance pathways. Methods: We developed a fine-tuned GigaPath foundation model (Xu et al., Nature, 2024) to predict progression-free survival (PFS), using 511 H&E-stained slides from 178 metastatic breast cancer patients treated with first-line CDK4/6 inhibitors (Palbociclib,108; Abemaciclib,70) plus endocrine therapy. Grad-CAM visualizations identified spatial regions associated with PFS. Spatial transcriptomics (Visium v2, 10x Genomics) was applied to 12 pre-treatment samples to compare gene expression between AI-identified PFS-long (sensitive) and PFS-short (resistant) regions. Additionally, 7 matched pre- and post-treatment tumor pairs underwent spatial transcriptomic profiling using Visium HD. GSEA was performed to analyze enriched pathways. Results: The model achieved an AUC of 0.84 in distinguishing cases with PFS <6 months versus ≥15 months. Grad-CAM identified intratumoral regions associated with sensitivity and resistance in 9 samples. In these regions, GSEA revealed that resistant areas were enriched for KRAS signaling (e.g., KRAS.600.LUNG.BREAST_UP.V1_UP and HALLMARK_KRAS_SIGNALING_UP), whereas sensitive areas were enriched for gene sets associated with suppressed KRAS activity (e.g., KRAS.600_UP.V1_DN) (Table). In the 7 pre- and post-treatment tumor pairs, spatial transcriptomics within tumor regions detected 29 pathways in at least 5 out of 7 tumor pairs. Among them, eight were related to KRAS signaling—including KRAS.LUNG.BREAST_UP.V1_UP and KRAS.BREAST_UP.V1_UP—five were associated with polycomb group proteins (BMI1, MEL18), and five involved tumor suppressor pathways (p53, BRCA1, PTEN, ATM). These results indicate that KRAS signaling was the most frequently enriched pathway following treatment. Conclusion: Two independent spatial transcriptomic analyses-comparing AI-identified regions and treatment timepoints—consistently identified KRAS signaling as the dominant mechanism of resistance to CDK4/6 inhibitors in breast cancer. These findings provide a compelling rationale to explore KRAS pathway inhibition to overcome resistance to CDK4/6 inhibitors in breast cancer. Citation Format: M. Tanioka, T. Nakayama, T. Marutani, Y. Niki, M. Tohda, Y. Hirose, L. Zehao, T. Daisuke, K. Morooka, A. Nakaya, T. Shien. Deep Learning-guided Spatial Transcriptomics Identifies KRAS signaling as the Dominant Mechanism of Resistance to CDK4/6 Inhibitors [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 PS3-06-18.