Citations

Abstract Guideline-concordant breast cancer treatment selection is increasingly complex due to rapid evolution in biomarker-driven therapies. Large language models (LLMs), such as GPT-4, offer a potential tool to support clinical decision-making. However, hallucinations and deviation from evidence-based guidelines remain significant concerns. This study evaluated whether retrieval-augmented generation (RAG) using National Comprehensive Cancer Network (NCCN) guidelines improves GPT-4 performance in treatment recommendation tasks. We developed a retrieval-augmented GPT-4 agent (RAG-GPT) by indexing 2025 NCCN Breast Cancer Guidelines into a vector database using sentence embeddings. Forty clinical vignettes were created from published case reports to reflect diverse scenarios, including HER2-low, triple-negative, and BRCA-mutated disease across all stages. Two models—a baseline GPT-4 (without augmentation) and the RAG-GPT—were prompted with structured templates requesting treatment plans, rationale, clinical trial options, and citations.Outputs were independently scored by two blinded oncology-trained reviewers using the modified Generative AI Performance Score (mG-PS). The mG-PS scored outputs for 2 categories, guideline concordance (Gold standard = 1.0, Acceptable = 0.5, Non-concordant = 0.0) and hallucination errors (Severe = -1.0, Moderate = -0.5, Mild = -0.25, None = 0.0). Guideline Concordance was deemed based on NCCN guidelines and what an oncologist would due based on using the guidelines for each case. Hallucinations are a term used for falsehoods in Artificial Intelligence, like made-up facts or studies, incorrect information, improper doses, or inappropriate or wrong treatments. To ensure consistency and minimize inter-rater variability, reviewers were provided with comprehensive scoring guidelines and detailed instructions for each evaluation metric. Scores were normalized between -1.0 and 1.0 for improved evaluation of tools. Readability and rationale clarity were rated on a 5-point Likert scale. Inter-rater reliability was assessed using Cohen’s kappa. The RAG-GPT demonstrated a significantly higher mean mG-PS score (0.599 ± 0.102, 95% CI: 0.498-0.702) compared to the baseline GPT-4, which not only scored lower but had a negative mean (-0.21 ± 0.143, 95% CI: -0.353 to -0.067). For the mG-PS Cohen’s Kappa test resulted in a score of .69 with substantial aggreament. An independent sample t-test showed a difference (t(78) = 29.1, p < 0.0001). Severe hallucinations were also absent in RAG-GPT outputs (0/40) but occurred in 8/40 baseline GPT-4 outputs. The evaluation of readability and rationality also showed notable differences, with the augmented LLM scoring significantly higher (M = 4.35, SD = 0.66) compared to the naïve LLM (M = 3.05, SD = 1.06). An independent samples t-test confirmed this difference t(65.3) = 6.58, p < 0.0001, further supporting the benefit of augmented Large Language models. Cohen’s Kappa test resulted in a score of .76 for a substantial agreement for readability and rational clarity scale. Limitations include small sample size; further fine-tuning and larger real-world validation studies remain necessary. Linking GPT-4 to NCCN guidelines via retrieval-augmented generation significantly improves treatment accuracy and reduces hallucination severity in breast cancer care scenarios. While further refinement is needed, the reduced incidence of severe hallucinations, increase in guideline-concordant recommendations, and enhanced readability findings support further development and prospective validation of guideline-anchored LLMs for clinical use. Citation Format: C. Yost, A. Aseem, B. Callas, S. Monick, L. Bonilla, R. Nguyen, R. Galamaga, S. Osborn, Y. Kumar, N. Ertz-Archambault. Retrieval-augmented GPT-4 improves NCCN-concordant breast cancer treatment recommendations [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 PS3-06-29.

Abstract Background: While synoptic reports have standardized breast cancer data, a wealth of critical information remains locked in unstructured free-text, including nuanced biomarker details and procedural notes. This data is not easily computable, creating a significant barrier to understanding comprehensive population-level statistics and subtle clinical trends. Traditional data extraction methods are often insufficient to capture this complexity. We developed and validated a novel, flexible AI-powered system using an agent-based architecture to extract, normalize, and discretize this valuable unstructured data, enabling previously infeasible population-scale analytics. Methods: Our system employs a modular, agent-based framework where each "agent" is a specialized large language model (LLM) tasked with a specific function in a sequential workflow. The process begins with an initiator agent that parses the synoptic report. Subsequent agents then perform targeted tasks, such as identifying and normalizing tumor characteristics, procedural details, and biomarker information from free-text fields. A final agent aggregates this information into a structured, computable format. To rigorously test our approach, we first developed a synthetic data generator to create realistic, anonymized report snippets with known ground-truth values. We then evaluated the performance of several leading LLMs of various sizes (e.g., gemini, llama3.3, phi4, deepseek-r1, gemma3) on this synthetic dataset. Results: Our early analysis on simulated synoptic reports demonstrated high accuracy in data extraction and normalization. The gemini-1.5-pro-002 model achieved an overall accuracy of 96.9%, with a sensitivity of 99.3%, specificity of 94.3%, and an F1-score of 97.0%. The llama3.3:70b model also showed strong performance with a 98.7% accuracy, (F1: 98.8%, sensitivity: 99.8%, specifically: 97.6%). Gemma3:27b was the least performant model tested (F1: 66.2%), substantially worse than much smaller models (deepseek-r1:1.5b and phi4:14b). A detailed error analysis of false positives and negatives has allowed us to iteratively refine the agent logic. By the time of the symposium, we will have analyzed a large cohort of real, de-identified synoptic reports and will present comprehensive performance metrics and initial findings on population-level trends. Conclusion: This AI-powered, agent-based system offers a powerful and scalable solution to overcome the limitations of structured data in synoptic reports. By accurately extracting and normalizing previously inaccessible free-text information, our approach has the potential to provide clinicians and researchers with near real-time insights into population-level statistics and trends in breast cancer. This will facilitate a more profound understanding of disease patterns and ultimately support evidence-based advancements in patient care and clinical research. Citation Format: S. N. Hart. A Novel Agent-Based AI System to Unlock Unstructured Data in Synoptic Reports for Advanced Population Health Analysis in Breast Cancer [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 PS3-06-11.

Abstract Introduction: Accurate detection and prioritization of tumor-specific somatic variants is foundational to the development of personalized immunotherapies. Traditional pipelines for neoepitope identification rely heavily on short-read next-generation sequencing (NGS) and DNA-based variant calling, followed by RNA expression filtering and HLA-binding prediction. However, this approach is limited by time-intensive workflows, potential false positives from DNA-only calls, and poor capture of complex isoforms or RNA-level editing. Moreover, the accuracy of neoepitope prediction is constrained by conventional sequence-based HLA binding algorithms that do not account for structural and biophysical features of peptide-MHC interactions, particularly across underrepresented HLA alleles. Methods: To overcome these limitations, we developed a streamlined, sequencing-to-vaccine framework that leverages long-read RNA sequencing for direct detection of expressed somatic variants, including fusions, indels, and single nucleotide variants. Using Oxford Nanopore Technologies’ cDNA-PCR approach, we enable comprehensive, full-length transcript characterization from both fresh and frozen breast tumor samples. Our bioinformatic pipeline, NeoTarget, integrates alignment, variant calling, and neoepitope inference from tumor RNA in a single, modular workflow. A core innovation of this pipeline is HLA-Inception, a deep convolutional neural network trained to predict peptide-MHC class I binding using features derived from molecular dynamics-based modeling of electrostatic potential at the peptide-HLA interface, enhancing generalizability across diverse HLA subtypes. Results: Expanding on previous results, the current NeoTarget pipeline identifies, evaluates, and prioritizes single nucleotide variants, indels, and fusions. To validate this system, we used a set of 9 human breast cancer tumors benchmarked with short read sequencing. Identified neoantigens included most variants shown by clinical short read sequencing and also novel structural variants (i.e., fusions) only revealed with long read sequencing with high confidence. Conclusion: Here, we describe the design and performance of the NeoTarget pipeline, and show proof-of-concept feasibility of applying this fast and cost-effective workflow for HLA-restricted epitope discovery in breast tumors using long-read RNA sequencing. Overall, these methods can be applied to rapidly identify and prioritize neoantigens for personalized neoantigen-based immune therapies. Citation Format: F. Batalini, S. Henry, Y. Zhang, R. Foster, J. M. Gayton, A. Hostal, E. A. Wilson, A. Singharoy, M. Gustafson, B. Pockaj, J. Park, K. S. Anderson. Somatic variant detection using long read RNA sequencing reveals HLA-restricted targets for immunotherapy [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 PS3-06-08.

Background Immune checkpoint inhibitors (ICIs) improve survival in advanced non-small cell lung cancer (NSCLC), yet current biomarkers such as programmed death-ligand 1 (PD-L1) expression and response criteria (Response Evaluation Criteria in Solid Tumors, RECIST, V.1.1) align poorly with long-term survival. Radiomics has been proposed as a source of novel biomarkers, but standard radiomic approaches suffer from limited biological interpretability and poor generalizability across treatment settings. We address these gaps by developing the Quantitative Vessel Tortuosity (QVT) score, a biologically interpretable imaging biomarker that quantifies tumor vascular complexity—a known mediator of immune evasion—from routine imaging. We hypothesized that QVT score would improve prognostication and enable treatment response monitoring in ICI-treated NSCLC, independent of current biomarkers. Methods This retrospective, multicenter study analyzed 1,301 CT scans from 682 patients with ICI-treated NSCLC. An automated pipeline segmented lesions and tumor-associated vasculature within each scan, extracting 910 QVT features measuring vascular shape and complexity. Unsupervised clustering of these features in a discovery cohort (N=375) was performed to identify fundamental vascular phenotypes. A continuous QVT score was then derived using regularized logistic regression to map patients along this phenotypic spectrum. QVT score was externally validated in ICI monotherapy (N=172) and chemoimmunotherapy (N=135) cohorts. In a longitudinal cohort (n=143), early on-treatment QVT score changes were evaluated for overall survival (OS) association. Results Two robust vascular phenotypes emerged in the discovery cohort: a highly vascularized, chaotic “QVT High” phenotype with poor post-ICI OS and a “QVT Low” phenotype with normalized vasculature and improved ICI outcomes. The continuous QVT score was prognostic for ICI monotherapy (HR=1.17 per 0.1 increase, p=0.0028) and chemoimmunotherapy (HR = 1.23 per 0.1 increase, p = 4.9×10⁻⁵). High QVT status remained prognostic for both treatments after adjustment for PD-L1 and clinical variables (adjusted HR range: 2.13–2.38, p≤0.002). Early decreases in QVT score during therapy, indicating vascular normalization, were associated with improved OS (HR=1.93, p=0.0022) independent of RECIST best overall response and tumor volume change. Conclusions QVT score is a novel, biologically interpretable imaging biomarker that quantifies vascular complexity. It enables automated, non-invasive prediction and monitoring of ICI outcomes by capturing treatment-induced vascular remodeling. Integrating QVT score into clinical decision-making and drug development can address critical gaps in precision oncology.

Abstract Background: Women younger than 40 generally fall outside the routine screening population for breast cancer (BC). We aimed to understand factors influencing diagnosis and treatment delays in these patients, and the impact of delays on disease-free survival (DFS). Methods: We included women aged ≤ 40 years at diagnosis of stage 0-IV BC enrolled in the Young Women’s BC Study from 2006-2016. Sociodemographics, initial BC symptoms, and disease status were collected via annual surveys. Medical record review was used to capture tumor characteristics and recurrence status. We used logistic regression to estimate associations between patient and tumor factors and three delay types: (1) self-delay (≥ 90 days from first symptom to presentation to a provider), (2) diagnosis delay (≥ 90 days from healthcare visit to diagnosis), and (3) treatment delay (≥ 60 days from diagnosis to first treatment). Odds ratios (OR) and 95 % confidence intervals (CI) were estimated in multivariable models after choosing factors via backwards selection (p≤0.05 retained). Hazard ratios (HR) and 95% CIs for the association between delays and DFS, overall and by stage, were estimated via Cox proportional hazard models, adjusted for tumor and treatment factors. Time to event was measured from diagnosis or treatment to recurrence, death, or last follow-up, whichever occurred first. Results: Of 1,297 patients, 1,071 who completed the long-form baseline survey were included. Median time from first symptom to presentation was 14 days (interquartile range (IQR)=2-45), presentation to diagnosis was 14 days (IQR=7-30), and diagnosis to treatment was 41 days (IQR=25-57). Overall, 14% (n=153) had a self-delay and 11% (n=120) had a diagnosis delay. Of 839 patients with self-detected tumors, 18% had a self-delay and 13% had a diagnosis delay. Among those with Stage 1-3 BC (N=928), 22% (n=205) had a treatment delay. In univariable models, income <50K (p=0.01), lower financial comfort (p=0.01), and not living with a partner (p=0.03) were associated with higher odds of self-delay, while pregnancy within 1 year pre-diagnosis (p=0.05) and first-degree family history of any cancer (p=0.04) or BC (p=0.03) were associated with lower odds. Self-delays were more likely for women diagnosed with stage 4 BC (p=0.004) or with HER2+ tumors (p=0.03) and less likely for those with triple negative (TN) BC (p<0.001). In the multivariable model, family history of BC (OR=0.40, 95% CI=0.17-0.80), income <50K (OR=1.97, 95% CI=1.22-3.13) and TN subtype (OR v. luminal A=0.30, 95% CI=0.13-0.60) remained associated. Odds of diagnosis delay were higher among women who were younger (p=0.01) or with Stage 4 BC (p=0.004), and lower for those with luminal B (p=0.04), TN (p=0.01), or high grade tumors (p=0.03). In the multivariable model, younger age (OR per 1-year=0.94, 95% CI=0.89-0.99), and subtype (luminal B v. A OR=0.60, 95% CI=0.38-0.95; TN v. luminal A OR=0.35, 95% CI=0.16-0.70) remained associated. Multivariable models for self-and diagnosis delays remained similar when restricted to self-detected patients.Treatment delays were less likely among unemployed individuals (p=0.001), and women with a positive germline BC risk variant (p=0.002), luminal B (p=0.005) or TN (p<0.001) subtypes, and high-grade tumors (p=0.03). Only tumor factors remained significant in the multivariable model, including subtype (luminal B v. A OR=0.59, 95% CI=0.41-0.86, TN v. luminal A OR=0.25, 95% CI=0.13-0.43), and stage (stage 3 v. 1 OR=0.37, 95% CI=0.20-0.65). Overall and by stage at diagnosis, delays were not statistically significantly associated with DFS before or after adjusting for tumor and treatment factors. Conclusions: Delays in breast cancer care are influenced by multiple psychosocial and clinical factors, and disparities exist by socioeconomic status. It is critical to decrease self- and diagnosis delays to prevent stage 4 breast cancer. Citation Format: B. X. Zhang, T. Sella, Y. Zheng, G. J. Kirkner, K. J. Ruddy, S. M. Rosenberg, V. F. Borges, L. Schapira, S. E. Come, J. Peppercorn, A. H. Partridge, K. D. Brantley. Delays in breast cancer diagnosis and treatment in young women [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 PS3-02-16.

Abstract Background: Tertiary lymphoid structures (TLS), which are organized ectopic lymphoid aggregates resembling secondary lymphoid organs, have been associated with improved outcomes across various cancer types. However, limited data exist on the prognostic relevance of TLS in HER2+ BC pts treated with trastuzumab. In our prior work, we demonstrated that the presence of functional TLS, defined by lymphoid aggregates with elevated expression of BCL6 or IL21R, correlated with improved outcomes in the NCCTG N9831 trial. In this study, we evaluated TLS-related gene and outcomes in the real-world database. Methods: 1344 HER2+ (HR-HER2+: 560; HR+HER2+: 784) BC samples were analyzed via NGS (592-gene panel, NextSeq; WES/WTS, NovaSeq; Caris Life Sciences). Above (high) or below (low) the 50th percentile expression levels of TLS-related genes were used to calculate overall survival (OS). IL21R-high (H) and -low (L) tumors were classified by RNA expression above or below the 50th percentile. Real-world OS was derived from insurance claims and calculated from date of biopsy to last contact using Kaplan-Meier. Statistical significance was assessed using chi-square and Mann-Whitney U with multiple comparison adjustments (q < 0.05). Results: Among pts with HR-HER2+ BC, high expression of IL21R, CXCL9, CD38, and CXCL10 were associated with improved OS (Table). In contrast, among pts with HR+HER2+ BC, high expression of CXCR3, PDCD1, TBX21, CD200 and IL21R, were associated with improved OS. In both HR-HER2+ and HR+HER2+, IL21R-H groups had 1) higher infiltration of B cells (HR-HER2+: 4.4% vs 2.9%; HR+HER2+: 5.6% vs 4.2%), M1 MΦ (HR-HER2+: 3.9% vs 2.6%; HR+HER2+: 3% vs 2.1%), M2 MΦ (HR-HER2+: 4% vs 3.2%; HR+HER2+: 5.7% vs 4.4%), NK cells (HR-HER2+: 2.8% vs 2.5%; HR+HER2+: 2.7% vs 2.4%), CD8+ T cell (HR-HER2+: 0.9% vs 0%; HR+HER2+: 0.7% vs 0%), Tregs (HR-HER2+: 2.9% vs 1.4%; HR+HER2+: 2.8% vs 1.5%), all q<.05; 2) higher T cell inflamed score (HR-HER2+: 87.5 vs -95; HR+HER2+: 71 vs -86), IFNy score (HR-HER2+: -0.19 vs -0.48; HR+HER2+: -0.27 vs -0.49), all q<.05; 3) higher expression levels of MHC class I genes (HLA-A, HLA-B, HLA-C, HLA-E, FC: 1.6-2.3; all q<.05), MHC class II genes (HLA-DQA1, HLA-DOA, HLA-DPA1, HLA-DMA, FC: 2.2-3.2; all q<.05) compared to IL21R-L group. Conclusion: Consistent with our prior findings, overexpression of TLS-related genes was associated with improved outcomes in HER2+ BC in real-world data. Notably, the prognostic impact of TLS-related genes differed by hormone receptor (HR) status. Among these, IL21R emerged as the only gene significantly associated with clinical outcomes across both datasets and HR subgroups. Higher IL21R expression is associated with more B cell, M1 MΦ, CD8 T cell and NK cells. These findings underscore the potential importance of the IL21/IL21R axis in HER2+ BC and warrant further investigation to validate its role and therapeutic relevance. Citation Format: S. Chumsri, S. Deshmukh, Y. Liu, S. Wu, J. Xiu, N. Norton, S. Shipra Gandhi, E. Torres, M. Lustberg, E. Perez, A. Nassar, G. Sledge, E. Thompson, K. Knutson. Tertiary lymphoid structure (TLS)-related gene expression and outcomes in HER2-Positive (HER2+) breast cancer (BC) in the Real-World Database [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 PS2-09-08.

Abstract Introduction: Patients with ER+/HER2- breast cancer associated with germline pathogenic variants (PVs) in BRCA1, BRCA2 and CHEK2 exhibit distinct clinicopathological features compared to sporadic cases. However, the optimal endocrine therapy strategies for these patients remain undefined. The Breast Cancer Index (BCI) is a validated gene expression-based signature that stratifies patients based on the risk of overall (0-10 years) and late (post-5 years) distant recurrence and predicts the likelihood of benefit from extended endocrine therapy (EET) in early-stage, HR+ breast cancer. Here we utilize BCI to investigate endocrine sensitivity and the predicted EET benefit in a cohort of ER+/HER2- breast cancer patients with BRCA1, BRCA2 and CHEK2 PVs. Methods: BRCA1, BRCA2 and CHEK2 PV carriers with ER+/HER2- breast cancer diagnosed at Mayo Clinic between 1996 and 2024 at Mayo Clinic were identified from a registry tracking germline test result. For comparison, we utilized the BCI Registry, a prospective, multi-institutional tumor registry of patients with early-stage ER+/HER2- breast cancer undergoing BCI testing, regardless of germline PV status. The analysis was restricted to women with invasive, early-stage locoregional breast cancer and available pre-treatment tumor specimens. BCI testing was performed by RT-PCR blinded to clinical outcome. BCI prognostic and BCI predictive categories were compared between the Mayo Clinic germline PV cohort and the BCI Registry using descriptive statistics and Fisher’s exact test. Results: The final analysis included 123 PVcarriers (BRCA1: 36; BRCA2: 40; CHEK2: 47) from the MayoClinic cohort and 2,563 women from the BCI Registry. The proportion of patientsdiagnosed before age 50 was higher among PV carriers—52.8% in BRCA1,50.0% in BRCA2, and 21.3% in CHEK2—compared to 7.0% in the BCIRegistry. PV carriers were also more likely to be premenopausal andnode-positive at diagnosis. Based on BCI score, a significant higher proportionof BRCA1 (75.0%, p=0.002) andBRCA2 (70.0%, p=0.01) PV carriers wereclassified as high risk, compared to 48.5% of patients in the BCI Registry, butno significant difference was seen for CHEK2 PV carriers (46.8%,p=0.88). A significantly higher proportion of BRCA1 PV carriers werepredicted to benefit from EET compared to the BCI Registry cohort (77.8% vs.33.9%, p<0.001), whereas no significant differences in BCI Predictivecategories were observed for BRCA2 or CHEK2 PV carriers. Overall,compared to breast cancer cases in the BCI registry, a significant differencein the distribution of the combined BCI prognostic and predictive categorieswas noted for BRCA1 (p<0.001)and BRCA2 (p=0.02) PV carriers,but not for CHEK2 PV carriers (p=0.66). In particular, BRCA1 andBRCA2 PV carriers were enriched for high-risk tumors with a high likelihoodof EET benefit (66.7% and 37.5%, respectively vs 25.1%, p<0.001 and p=0.02),whereas only a small non-significant increase was observed for CHEK2 PVcarriers (29.8% vs 25.1%, p=0.66) compared to breast cancer cases in the BCIRegistry. Conclusions: Genomicclassification by BCI reveals distinct patterns of endocrine sensitivity andpredicted EET benefit among germline PV carriers, particularly in BRCA1and BRCA2 PV carriers. These findings warrant further investigation andmay inform personalized endocrine therapy strategies for women with germlinePV-associated ER+/HER2- breast cancer. Citation Format: S. Yadav, B. O'Neal, S. Yasir, N. Siuliukina, A. Karki, Y. Zhang, A. K. Anderson, K. Treuner, M. P. Goetz, F. J. Couch. Endocrine Sensitivity and Predicted Benefit of Extended Endocrine Therapy based on Breast Cancer Index (BCI) in BRCA1, BRCA2 and CHEK2 Pathogenic Variant Carriers with ER+/HER2- Breast Cancer [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 PS2-08-25.

Abstract Background: Triple-negative breast cancer (TNBC) is an aggressive yet immunogenic subtype of breast cancer (BC) and the only one with FDA-approved immune checkpoint inhibitor (ICI) - pembrolizumab. Currently, PD-L1 (22C3) is the companion biomarker for predicting pembrolizumab benefit, but its utility is limited by spatial/temporal heterogeneity, assay variability, and modest correlation with treatment response. Chemokines such as CXCL9, CXCL10, and CCL5 recruit CD8+ T cells into the tumor microenvironment and have been associated with improved response to ICI, but their predictive value alongside PD-L1 remains unclear. Here, we examined the relationship between these chemokines, PD-L1, and pembrolizumab efficacy in TNBC. Methods: We analyzed 3662 TNBC tumors profiled by NGS (592, NextSeq; WES/WTS, NovaSeq; Caris Life Sciences). Immune cells were estimated using WTS deconvolution (Quantiseq). PD-L1 positivity was defined as >10% CPS (PharmDx 22C3). CXCL9/CXCL10/CCL5-high (H) and -low (L) TNBC were classified by RNA expression above or below 50th percentile. Real-world median overall survival (mOS) was derived from insurance claims and calculated from start of pembrolizumab to last contact using Kaplan-Meier. Statistical significance was assessed by chi-square and Mann-Whitney U with multiple comparison adjustments (q<0.05). Results: TNBC expressed higher levels of CXCL9, CXCL10, and CCL5 (N = 3662, median TPM: 5.3, 14.8, and 15.1) compared to HER2+ (N = 1344; 4.8, 10.1, 10.6) and HR+HER2- BC (N = 6097; 2.7, 6.6, 9.4), all q<0.05. Moderate correlations were observed between PD-L1 CPS and each chemokine (r = 0.41, p<.01). PD-L1-positive tumors had better mOS (24.3 months (m) vs 18.6 m, HR 0.75, 95% CI 0.6-0.92, p=0.007) compared to PD-L1-negative TNBC post-pembrolizumab. Among PD-L1-negative tumors, those with CXCL9-H had better mOS (26.9 m vs 15.8 m for CXCL9-L, HR 0.61, 95% CI 0.45-0.82, p=0.001) post-pembrolizumab. Among PD-L1-positive tumors, CXCL10-H and CCL5-H was associated with improved mOS (CXCL10-H: 25.7 m vs. 20.7 m for CXC10-L, p = 0.03; CCL5-H: 26.7 vs. 18.5 m for CCL5-L, p = 0.03) post-pembrolizumab. PD-L1 negative CXCL9-H, PD-L1 positive CXCL10-H and PD-L1 positive CCL5-H groups showed higher infiltration of CD8+ T cell, B cells, M1/M2 macrophages, Tregs. They also exhibited higher T cell inflamed score, IFNy score and MAPK activation score vs. CXCL9-L, CXCL10-L and CCL5-L groups (Table). Conclusions: In TNBC, a chemokine signature comprising CXCL9, CXCL10 and CCL5 refines PD-L1 based prediction of pembrolizumab benefit. Specifically, high CXCL10 or CCL5 expression in PD-L1-positive tumors, and high CXCL9 expression in PD-L1-negative tumors, were associated with improved survival post pembrolizumab and immune-rich microenvironment. These findings support the use of chemokine profiling to complement PD-L1 and enhance patient selection for ICI therapy in TNBC. Citation Format: S. Gandhi, S. Deshmukh, S. Wu, J. Xiu, G. Sledge Jr., D. Trapani, J. Leone, S. Chumsri, M. Lustberg, C. Hong, S. Yao, K. Takabe, M. Ernstoff, M. Opyrchal, P. Chalasani, C. Paulos, G. Lesinski, A. Madabhushi, S. Badve, K. Kalinsky, P. Kalinski. Chemokine signatures improve PD-L1-based prediction of pembrolizumab response in triple-negative breast cancer [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 PS2-08-14.

Abstract Background: Male breast cancer (MaBC) is a rare malignancy, accounting for approximately 1% of all breast cancer cases. Our previous studies suggest that MaBC is biologically distinct from female breast cancer (FeBC), exhibiting a unique mutational profile and immune-genomic microenvironment. While the impact of these molecular alterations on site-specific metastases has been explored in FeBC, it remains poorly understood in MaBC. This gap limits accurate risk stratification and prediction of disease trajectory in male patients. In this study, we investigated site-specific genomic and immune alterations across metastatic sites in MaBC. Methods: 238 MaBC samples were tested by NGS (592, NextSeq; WES, NovaSeq) and WTS (NovaSeq; Caris Life Sciences, Phoenix, AZ). We performed analysis in hormone receptor-positive (HR+) and HER2-negative (HER2-) subtypes which are most common in male. HR+HER2- was defined as estrogen receptor (ER) or progesterone receptor (PR) positive and HER2 IHC 1+ or 2+ with negative chromogenic in situ hybridization (CISH) assay. Immune cells were estimated using WTS deconvolution (Quantiseq). Gene expression profiles were analyzed for T-cell inflamed score, interferon-gamma (IFN-gamma) score and MAPK Activation Score. We classified patients by site of biopsy/specimen site. Statistical significance was assessed with chi-square and Mann-Whitney U tests, p<0.05 is considered statistically significant. Results: Our cohort included 86 HR+HER2- MaBC samples biopsied from primary breast cancer (pBC) (n=39), bone (n=15), liver (n=8), lung (n=8), lymph nodes (n=7), and skin (n=9) mets. We report here only the data on statistically significant differences, but the overall landscape has been depicted across all defined-sites investigated. pBC showed a significantly lower prevalence of PIK3CA mutations (23.6% vs. 53.3%, p<0.05) compared to bone mets. ESR1 mutations were absent in pBC but present in 25% of liver and 25% of lung mets (all comparisons p<0.05). Immune profiling revealed reduced infiltration of M2 macrophages (4.3% vs. 6%, p<0.05) and neutrophil (2.2% vs. 6.9%, p<0.05) in pBC relative to bone Mets. Further, CD4+ T cells were lower in pBC compared to lymph nodes (0% vs. 2.4%, p<0.05). pBC had lower expression of immune checkpoint genes (PDCD1LG2, HAVCR2 and CD47) compared to other mets. pBC exhibited lower MAPK pathway activation score (-1.6 vs. 0.17, p=0.03) and a trend toward lower T cell-inflamed gene expression scores (-65 vs. 8, p=0.06) compared to bone mets. Bone mets had higher infiltration of M2 macrophage compared to lung (6% vs. 4.4%, p<0.05) and lymph node (6% vs. 4.5%, p<0.05). Neutrophil infiltration was elevated in bone versus liver (6.9% vs. 2%, p<0.05) but T regulatory cells were lower in bone compared to lymph nodes (1% vs. 2.4%, p<0.05). Bone showed reduced expression of CD47 (median TPM 50.6 vs. 109.9, p<0.05) compared to skin. Further bone had lower IFN-gamma scores (-0.41 vs. -0.17, p<0.05) compared to lymph node. Data on prior number and types of cancer treatment to give a context to the findings will be presented at the meeting. Conclusions: Our study highlights that each MaBC metastatic site harbors a unique genomic and immune landscape that can influence tumor behavior and potentially impact patient prognosis. Better understanding of metastatic site-specific molecular features in larger cohorts can help in the future research efforts and guiding treatment decisions. Citation Format: D. Trapani, S. Deshmukh, S. Wu, J. Xiu, N. Lin, G. Curigliano, P. Spanheimer, P. Advani, P. Barreto Coelho, M. Lustberg, G. Sledge Jr, S. Tolaney, J. Leone. Site specific Genomic and Immune Landscapes Underpinning in metastatic Male Breast Cancer [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 PS2-08-04.

Abstract Background: Stromal tumor infiltrating lymphocytes (sTILs) are associated with improved response to chemotherapy and better overall survival (OS) in triple-negative breast cancer (TNBC). Quantification of sTILs via hematoxylin and eosin (H&E) tumor slides offers a widely available method to define a simple, effective surrogate for host anti-tumor immunity. Digitization of H&E slides has rapidly gained acceptance for use in diagnostic pathology and research, opening the door for image-based computational approaches to enumerate sTILs, among other features. Methods: We describe QuTILs - a multilayer perceptron-based framework to detect/enumerate TILs trained on >700,000 cells via The Cancer Genome Atlas Program (TCGA) slide images, implemented in the widely used open-source QuPath digital pathology suite. QuTILs was applied to digital H&E images using standard computing in primary breast tumors from phase 3 breast cancer clinical trials - CALGB 40502 (testing), CALGB 40601 and 40603 (validation) - in total 1,729 total images from 962 unique patients. The aims of this investigation were: 1) to assess the relationship between TIL patterns with overall or event-free survival outcomes; 2) to benchmark diverse image-based characteristics of TIL abundance and spatial architecture; and 3) to evaluate the association between image-based TIL levels and RNA-seq-based immune deconvolution metrics (CIBERSORT, EPIC, quanTIseq). Results: QuTILs average processing time was 7.9 minutes for whole-slide total TIL count and area density on a standard laptop workstation. In CALGB 40502, there was positive but modest correlation between pathologist TILs and QuTILs total TILs %, higher in the TNBC subgroup of CALGB 40502 (Pearson r = 0.55). In Cox proportional hazards models, the CALGB 40502 TNBC cohort showed significant univariate association for total TILs % with increasing decile (higher total TIL%) associated with reduced hazard (hazard ratio [HR]: 0.712, 95% CI: 0.559-0.906; P = 0.006), which remained significant in multivariable model incorporating age, race, and treatment arm (HR: 0.713, 95% CI: 0.556-0.915; P = 0.008). In validation TNBC clinical trial CALGB 40603, marginal TILs % was significant in the univariable Cox model (HR: 0.786, 95% CI: 0.639-0.966). Evaluation of TILs % by QuTILs with RNAseq-based deconvolution metrics in cohort-stratified and deconvolution method-stratified multivariable models showed that decile total TILs % by QuTILs was shown to significantly predict outcome in the EPIC (HR: 0.668, 95% CI: 0.466-0.957) and quanTIseq (HR: 0.646, 95% CI: 0.446-0.936) sets, with a similar magnitude as in clinical Cox, suggesting that RNA-based deconvolution estimates may not supplement the prognostic power of image-based TIL estimates. Exploratory analyses of pre- versus post-neoadjuvant therapy H&E images in CALGB 40603 and CALGB 40601 revealed a significant decrease in QuTILs total TIL% in CALGB 40603 for all pairs (n = 118 pairs, pre-median=9.52%, post-median=6.33%; Wilcoxon signed rank P < 0.001) but no significant change in QuTILs total TIL% among all pairs in CALGB 40601 (n = 82 pairs, pre-median=15.92%, post-median=15.10%; P = 0.886). Conclusions: QuTILs provides a computationally efficient open-source workflow for sTIL identification from digital H&E images that can be deployed on a standard laptop. We present the first evaluation of digital sTIL enumeration in CALGB 40502, CALGB 40603, and CALGB 40601 phase III breast cancer clinical trials and demonstrate significant prognostic association with QuTIL total TIL % among TNBCs across studies. Citation Format: M. Vater, R. Salgado, E. Blige, H. Lefebvre, A. Strahan, M. Gatti-Mays, H. Rugo, K. Ballman, M. Watson, Y. Wen, R. Leon-Ferre, K. Niazi, W. Symmans, L. Carey, K. AbdulJabbar, Z. Li, D. Stover. Qutils: open-source image-based infiltrating immune cell detection and outcomes in calgb 40502, calgb 40603, and calgb 40601 phase iii breast cancer clinical trials [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 PS2-07-27.