Citations

Abstract Breast cancer accounts for nearly a quarter of all cancers in women with an estimated 2.3 million new diagnoses and over 670,000 deaths annually. Triple-negative breast cancer (TNBC) is an aggressive subtype characterized by lack of estrogen receptor alpha (ERα) and progesterone receptor (PR) expression in the absence of HER2 amplification. TNBC accounts for ∼15% of all incident breast cancers and is associated with poor patient outcomes. Problematically, TN tumors exhibit vast inter-patient molecular heterogeneity making the identification of novel therapeutic vulnerabilities difficult and the development of alternative treatment strategies problematic. We and others have demonstrated that estrogen receptor beta (ERβ) is expressed in about 20% of TN tumors and potently suppresses proliferation, invasion, and migration in TNBC cells in vitro and in vivo. We thus hypothesized that endocrine therapies that selectively activate ERβ would be an effective treatment approach for this subset of patients. We also anticipated that some tumors would display de novo or acquired resistance to ERβ targeted therapies despite ERβ protein expression. We therefore sought to define mechanisms of resistance to ERβ targeted therapy and to identify synthetic lethal vulnerabilities. We generated the first models of ERβ resistant TNBC through chronic exposure to estradiol or an ERβ specific agonist (LY500307) and through a CRISPR genome wide knockout screen. We identified the LncRNA XIST as one of the most upregulated transcripts in multiple models of ERβ resistant disease, and knockdown of XIST expression was shown to resensitize resistant cell to estradiol and LY500307 treatment. XIST was not found to be directly regulated by ERβ in TNBC further suggesting that its overexpression is a bona fide mechanism of resistance to ERβ mediated suppression of proliferation. The phenomena of chronic ligand exposure leading to activation of XIST expression in ERβ+ cells is not shared with ERα suggesting that the mechanism by which XIST expression is elevated during the development of resistance is distinct between the two sister receptors. We further characterized XIST expression in patient samples of TNBC and found that XIST expression varies, but largely exhibits a bimodal distribution with most tumors displaying either low or high transcript levels. In conclusion, we have identified the X-inactivating transcript as a driver of resistance to ERβ targeted therapies in TNBC. Future studies aimed at uncovering the consequences of XIST over-expression on global gene expression, functionality of ERβ, and aggressiveness of TNBC are ongoing. Citation Format: Michael Emch, Kirsten Aspros, Matthew P. Goetz, John Hawse. The LncRNA XIST mediates resistance to estrogen receptor beta targeted therapies in triple negative breast cancer [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 2293.

Abstract Background: High-dose interleukin-2 (IL-2) is administered after tumor-infiltrating lymphocyte (TIL) infusion to promote in-vivo expansion and persistence, yet optimal dosing and the clinical significance of out-of-specification (OOS) TIL products remain unclear. We investigated the impact of number of infused IL-2 doses and manufacturing OOS status on outcomes in metastatic melanoma treated with commercial TIL therapy. Methods: We retrospectively analyzed 31 patients treated across three Mayo Clinic sites. Planned IL-2 was 600,000 IU/kg every 12 hours for ≤6 doses; actual dosing varied based on clinical tolerance. Patients were grouped as low (0-2) vs high (≥3) IL-2 doses. TIL products were classified as within-specification or OOS per commercial release criteria. Survival associations were evaluated using Kaplan-Meier and multivariable Cox regression. Results: In our 31-patient cohort, the median age was 59 years (range 17-77), 58% were male, and 84% had ECOG 0. At TIL infusion, M-stage distribution was M1a 6%, M1b 10%, M1c 61%, and M1d 23%. LDH was elevated above the upper limit of normal in 42%. The median number of IL-2 infusions was 4 (range 0-6); 58% of patients received ≥3 doses, while 2 patients (6%) received none. Out-of-specification (OOS) TIL products occurred in 6 patients (19%). Best responses included complete response (26%) and partial response (23%), for an overall response rate of 48% and a disease-control rate of 61% (CR + PR + SD). Among responders, the median duration of response was 84 days (95% CI 73-95). At data cutoff, the median follow-up was 253 days (95% CI 137-370). The most common grade 3/4 toxicities occurring >14 days after TIL infusion were lymphopenia (39%), anemia (16%), and thrombocytopenia (10%). Median progression-free survival (PFS) was 99 days (95% CI 73-125) with 20 progression events, and median overall survival (OS) was 193 days (95% CI 32-354) with 13 deaths. Number of infused IL-2 doses was dependent on patients tolerability. Receiving ≥3 IL-2 doses was associated with improved overall survival (394 vs 101 days; p < 0.001) and a trend toward longer progression-free survival (110 vs 59 days; p = 0.055). OOS products predicted inferior PFS (30 vs 110 days; p = 0.003) and showed a trend toward worse OS (120 vs 394 days; p = 0.079). In multivariate analysis, ≥3 IL-2 doses independently improved PFS (HR 0.20, 95% CI 0.06-0.69; p = 0.011) and OS (HR 0.09, 0.02-0.36; p = 0.001), while OOS status increased risk of progression (HR 8.10, 2.24-29.25; p = 0.001) and death (HR 6.16, 1.35-28.20; p = 0.019). Conclusions: Higher number of infused doses IL-2 and meeting commercial TIL release criteria were independently associated with improved outcomes. Citation Format: Mohamed A. Aboelatta, Jeffrey E. Johnson, Jabra G. Zarka, Matyáš Benada, James W. Jakub, Roxana S. Dronca, Ruqin Chen, Mahesh Seetharam, Deepti Behl, Svetomir N. Markovic, Lisa A. Kottschade, Heather N. Montane, Matthew S. Block, Anastasios Dimou, Robert R. McWilliams, Paula Gill, Yi Lin, Arkadiusz Z. Dudek. Impact of number of IL-2 infusions and product out of specification status on outcomes of tumor infiltrating lymphocyte therapy for melanoma [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 3706.

Abstract Gastrointestinal (GI) cancers are rising in people < 50 and remain major causes of cancer mortality despite therapeutic advances. Chimeric antigen receptor T (CART) cell therapy has transformed treatment of hematological cancers. However, its efficacy in GI cancers has been hindered by antigen heterogeneity and immunosuppressive cancer associated fibroblasts (CAFs) in the tumor microenvironment (TME.) Here we aimed to 1) study CART-CAF interactions and 2) engineer CART cells capable of eliminating CAFs to overcome their inhibitory role and enhance their antitumor activity in GI cancers. We first analyzed primary CAFs from 20 GI cancer patients. CAFs showed high fibroblast activation protein (FAP) expression and significantly suppressed T cells proliferation (26-77%, p≤0.0001). We next generated CART cells capable of safely eliminating both GI cancer cells and CAFs. To target cancer cells, we developed a novel nanobody (VHH)-based CAR against Claudin18.2, a GI cancer specific target. To avoid off-tumor toxicity in healthy FAP+ tissues, we incorporated an inducible FAP-CAR using the synNotch platform, enabling conditional targeting upon engagement with Claudin18.2+ cells, restricting FAP targeting within the TME. VHH Claudin18.2 CART cells showed superior specificity and efficacy vs scFv-based CART cells (hu8E85 or zolbetuximab). In Claudin18.2+ SNU-C2B xenografts, tumors were 92mm3 (VHH) vs 48mm3 (huE85) vs 211mm3 (zolbetuximab) on day 23 post-CART, median survival undefined vs 17 days vs 23 days respectively; p=0005. In Claudin18.2+ LS411N xenografts, tumors were 315mm3 vs 106mm3 vs 382 mm3 respectively at 23 days post-CART, median survival undefined vs 19 days vs undefined respectively; p=0004. VHH Claudin18.2 CART cells showed no off-tumor effects, while hu8E85 CART induced significant toxicity. We next established a co-engraftment model. NSG mice received 1x106 SNU-C2B and 6x106 CAFs from a GI cancer patient. At 500 mm3 tumors, mice received 5x106 VHH Claudin18.2 or inducible Claudin18.2/FAP CART. Dual-targeting CART cells showed greater antitumor response (728mm3 vs 320mm3, p=0.0011 at day 20 post-CART), expansion (61 vs 2,350 CART/µl of blood, p=0.0037), and tumor infiltration (8.71 x 105 vs 1.03 x 106 CART/g of tumor, p<0.0001). This was also observed in gastric, colon, and pancreatic PDX models, with no observable off-tumor cytotoxic effects with inducible dual-targeting Claudin18.2/FAP CART. Our findings demonstrate that our novel VHH-based Claudin18.2 CART cells are highly specific and effective against GI malignancies, and that inducible Claudin18.2/FAP dual-targeting CART cells safely remodel the TME into an immune-permissive state. This approach represents a mechanistically innovative and translationally relevant strategy to overcome CAF-mediated immunosuppression, enhance CART cell efficacy, and improve clinical outcomes in refractory solid tumors. Citation Format: Omar L. Gutierrez Ruiz, Jennifer Feigin, Brooke Kimball, R. Leo Sakemura, Elizabeth L. Siegler, Dominic A. Skeele, Ateka Saleh, Claudia Manriquez Roman, Ismail Can, Olivia Sirpilla, Kun Yun, Carli M. Stewart, Mehrdad Hefazi, Yoh Yamaguchi, Lionel Aurelien Kankeu Fonkoua, Long K. Mai, Truc Huynh, Sophia Y. Goldberg, Grace E. DeFranco, Sara Foote, Ekene Johnkennedy Newton Ogbodo, Hong Xia, Michael Redig, Matthew L. Matthew, Tamiel Nichole Turley, Skyeler Klinge, John Alton Copland, Saad J. Kenderian. Optimized CART cell therapy to remodel the tumor microenvironment and eliminate GI malignancies [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 1525.

Abstract Background: Patients with residual invasive disease after neoadjuvant chemotherapy (NAC) for HER2-positive breast cancer remain at high risk for recurrence even with standard adjuvant trastuzumab emtansine (T-DM1). H2NVAC, a multi-epitope HER2 peptide vaccine, has previously demonstrated the ability to elicit robust and durable immune responses. Here, we assessed the safety and early immunogenicity of H2NVAC in combination with T-DM1 in this high-risk population. Methods: Patients with stage II-III HER2-positive breast cancer and residual invasive disease following NAC and HER2-directed therapy received standard adjuvant T-DM1 alongside six priming doses of H2NVAC administered every 3 weeks. The primary endpoint was the occurrence of dose-limiting toxicity (DLT) within 21 days of the first vaccination, with adverse events graded per CTCAE v5.0. Immunogenicity assessments included HER2-specific T-cell responses measured by IFNγ ELISpot and HER-specific serum antibodies measured by ELISA at baseline and 30 days after the sixth priming dose. Results: 20 patients were enrolled. H2NVAC plus T-DM1 was well tolerated with no DLTs observed. All treatment-emergent AEs were grade 1-2, with fatigue (grade 1 60%, grade 2 5%), peripheral sensory neuropathy (grade 1 35%, grade 2 10%), elevated alkaline phosphatase (grade 1 40%), nausea (grade 1 35%, grade 2 5%), arthralgia (grade 1 30%), and injection-site reactions (grade 1 50%) being most common. Robust vaccine-specific T-cell responses by ELISpot were observed in 94% of evaluable patients, who mounted HER2-specific T-cell responses at 30 days after the sixth priming dose, with a mean 197-fold increase over baseline. Further CYTOF analysis showed a significant decrease in exhausted T-cell phenotypes, with reduced LAG3 and TIM3 expression on both CD4+ and CD8+ T cells. For antibody responses by ELISA, 26% of patients developed HER2 protein-specific antibody responses at this time point. There was no significant increase in antibodies to H2NVAC peptides or to unrelated tumor antigens (p53, IGFBP2, TERT, myoglobin) at 30 days after the sixth vaccine. Ongoing correlative analyses will assess the evolution of immune responses at later time points following completion of T-DM1 therapy. Conclusions: Concurrent administration of H2NVAC with T-DM1 was safe and did not increase toxicity beyond single-agent T-DM1. To our knowledge, this is the first study to demonstrate that a cancer vaccine can be delivered concurrently with an antibody-drug conjugate without increasing adverse effects while still eliciting robust T-cell responses. Ongoing analyses will further define the kinetics and durability of vaccine-induced antibody and T-cell responses at later time points, as well as evaluate the efficacy of H2NVAC in reducing recurrence risk in the randomized phase 2 trial. Citation Format: Saranya Chumsri, Sharmila Giri, Andy J. Ness, David W. Hillman, Nadine Norton, Davitte Cogen, Brian M. Necela, Aziza Nassar, Donald W. Northfelt, Pooja Advani, Rohit Rao, Alvaro Mareno-Aspitia, Brenda Ernst, Kathryn J. Ruddy, Matthew P. Goetz, Keith L. Knutson. Safety, tolerability, and early immune responses in the safety run-in of an HER2 multi-epitope vaccine combined with T-DM1 in residual HER2-positive breast cancer [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 6715.

Abstract Prostate cancer requires continual escape from immune surveillance, but mechanisms underlying this escape in the transition from androgen sensitivity to resistance and metastasis remain incompletely understood. Androgen-deprivation therapy (ADT)—the treatment backbone—induces tumor cell apoptosis but also remodels the tumor microenvironment (TME). We have assembled a single-cell RNA-seq atlas from >800,000 cells across 68 patients treated at Columbia University and Memorial Sloan Kettering. The resource includes 33 primary tumors (16 treatment-naïve, 17 post-ADT) annotated with time to PSA recurrence, and 31 metastases, including longitudinal pre- and post-ADT data in 14 castration-sensitive tumors (CSPC), 17 castration-resistant tumors (CRPC), and 4 neuroendocrine tumors (NEPC). This dataset spans the spectrum of prostate cancer disease progression.Using ARACNe and VIPER to infer activity of transcriptional regulatory and signaling proteins, we detect conserved cell state programs while minimizing batch effects. Several significant features emerged. Immune infiltration—both lymphoid and myeloid—is significantly higher in CSPC metastases than in primary or CRPC tumors and increases further after ADT. T cells constitute ∼15% of cells in primary tumors versus 30% in treatment-naive CSPC (40% post-ADT), falling to 13% in CRPC and <1% in NEPC (p<1e-32), consistent with an initially "hot" CSPC TME that progressively excludes T cells with disease progression.Among ∼350,000 epithelial/tumor cells, trajectory analysis reveals two differentiation pathways from normal epithelium. Pathway 1—dominant in primary tumors—shows androgen receptor (AR) activity with three sub-phenotypes (T1-T3): T1 exclusive to primary tumors; T2 shared by primary and CSPC; and T3 restricted to CRPC, with high ERG expression, and very strong AR activity. Pathway 2 (clusters T4-T6), present in NEPC and a subset of CRPC, lacks AR activity and displays SOX2, FOXA2, PRKD1. These aggressive states uniformly overexpress AURKA, RET, and TOP2A, suggesting druggable targets for combination therapy.We further define five cancer-associated fibroblast (CAF) subtypes. CAF1-3 track with advanced, treatment-resistant disease—enriched in NEPC versus CRPC (25% vs 18% of CAFs) and in CRPC versus CSPC (18% vs 11%)—while CAF4-5 predominate in primary tumors. We identify druggable proteins and receptor-ligand interactions linking CAF1-3 with T4-T6 tumor states; in vivo perturbations are underway.This atlas provides a resource profiling tumor and immune ecosystems across prostate cancer progression. It highlights MR programs and stromal-tumor-immune interactions that underlie immune evasion and ADT resistance, yielding actionable targets and combination-therapy hypotheses for aggressive disease. Citation Format: Aleksandar Obradovic, Kwangmin Yoo, Chang Liu, Patrick McCann, Casey R. Ager, Matthew C. Dallos, Samir Zaidi. Comprehensive profiling of primary and metastatic prostate tumors reveals distinct tumor and fibroblast cell states associated with androgen resistance [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 7419.

Abstract Background: Glioblastoma recurrence is inevitable despite aggressive surgical resection of contrast enhancing (CE) tumor. Unlike the deep molecular characterization of CE glioblastoma, unresected infiltrating tumor, representing minimal residual disease (MRD) has not been well profiled. Increasingly, clinical trials are being conducted in the MRD setting for glioblastoma patients, necessitating understanding of unresected infiltrating tumor for targeted therapeutic development. Methods: Visium HD spatial transcriptomics was used to profile a cohort of newly diagnosed, treatment-naïve glioblastoma patients that underwent supramaximal resection, with both core and infiltrating in situ components. Spatial profiles generated expression signatures of infiltrative and CE tumor. For in silico prediction of therapeutics targeting the MRD, expression signatures of infiltrative and CE tumor were compared with drug perturbation signatures from the NIH L1000 database using the sRGES method to predict reversal of disease expression patterns. IC50 data from human cell lines were aggregated from the Genomics of Drug Sensitivity in Cancer or individual study data while blood-brain barrier (BBB) penetrance was predicted using the CNS-MPO approach. Findings: Infiltrative tumor expression profiles were enriched for genes corresponding to neural and oligodendrocyte progenitor like cells. In contrast CE glioblastoma displayed mesenchymal and astrocytic programs, suggesting spatial concentration of glioblastoma states within MRD and CE tumor. For CE glioblastoma, proteasome inhibitors and glucocorticoid receptor agonists were predicted to have the highest antitumor activity. In contrast, histone deacetylase inhibitors (HDACi) were highly ranked for MRD while exhibiting low predicted efficacy in CE tumor. Synergy analysis of drug candidates provided multiple dual treatment options for both the CE and MRD contexts. Conclusions: Cumulatively, these data indicate that the resected tumor does not reflect the MRD state, providing clarity on why targeted therapeutics for glioblastoma have not been particularly successful. Spatial profiling of the MRD demonstrates vulnerability to unique drug targets, which if utilized in the correct setting could confer greater survival benefit. Citation Format: Harrshavasan Congivaram, Shashwat Tripathi, Mateo Gomez, Katy McCortney, Ching Man Wai, Ruochen Du, Thomas K. Sears, Jianzhong Zhang, Daniel J. Brat, Craig M. Horbinski, Mark W. Youngblood, Jared T. Ahrendsen, Adam M. Sonabend, Stephen T. Magill, Matthew C. Tate, Maciej S. Lesniak, Sean Sachdev, Timothy Sita, Priya Kumthekar, Karan Dixit, Robin Buerki, Ditte Primdahl, Mustafa Khasraw, John de Groot, David A. Reardon, Rimas V. Lukas, Roger Stupp, Amy B. Heimberger. In silico screening of therapeutics candidates targeting minimal residual disease in glioblastoma [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 2496.

Abstract Background: Deletions in MTAP occur in 10-15% of cancers and are associated with poor prognosis. MTAP deletion (MTAPdel) leads to accumulation of MTA which causes partial inhibition of PRMT5, creating a synthetic lethality approach with PRMT5 inhibitors. Multiple investigational MTA-cooperative PRMT5 inhibitors have demonstrated early clinical efficacy in MTAPdel NSCLC and other cancers. MTAPdel, PRMT5 expression, and clinical outcomes to standard of care immune checkpoint inhibitors (ICI) and chemotherapy (chemo) have not been well characterized. Methods: The Tempus Lens Platform (Tempus AI, Inc., Chicago, IL) was used to query the Tempus multimodal de-identified database and establish and subsequently analyze a cohort of patients (pts)with NSCLC and DNA (Tempus xT) and RNA (Tempus xR) testing, treated with front-line (1L) ICI with or without chemo (N=3,676). PRMT5 expression was normalized as transcripts per million (TPM) and reported as log2(TPM + 1). Pts were classified into high vs low groups based on median PRMT5 expression and into MTAPdel vs intact (MTAPwt) based on MTAP biallelic loss. Immune infiltration was estimated using Quantiseq. Real-world overall survival (rwOS) was defined as the time from 1L start to death or last known follow-up. Real-world progression-free survival (rwPFS) was defined as the time from 1L start to death, progression or last known follow-up. Hazard ratios (HR) were calculated using Cox proportional hazards models and adjusted for age, sex, histology, smoking status and PDL1 status. Results: PRMT5 expression was moderately increased in pts with MTAPdel (6.65 v 6.70, p=0.012). Alterations in MYC, SMARCA4 and KEAP1 were enriched in PRMT5-high tumors in both MTAPdel and MTAPwt pts (p<0.03). Pts with PRMT5-high tumors had decreased infiltrate of multiple immune cells, including B cells, macrophages (M1 and M2), NK cells, CD4 and CD8 T cells (p<0.001). Pts with MTAPdel/PRMT5-high expression had shorter rwOS compared to MTAPwt/PRMT5-low when treated with ICI alone (HR 1.58, 95% CI 1.11-2.24, p=0.011) or with ICI + chemo (HR 1.65, 95% CI 1.34-2.03, p<0.001). These associations persisted on multivariate analysis. Pts with MTAPdel/PRMT5-high expression were also associated with worse rwPFS when treated with ICI + chemo (HR 1.5, 95% CI 1.22-1.86, p<0.001) on univariate and multivariate analysis, but did not associate with rwPFS in pts treated with ICI alone (HR=0.97, 95% CI 0.66-1.42, p=0.9). Conclusion: PRMT5 expression defines an immune suppressed subpopulation of pts with NSCLC. Pts with elevated PRMT5 expression and MTAPdel have poor outcomes with ICI+/- chemo, indicating PRMT5 RNA expression levels can be used to inform clinical trial designs, particularly those that are testing novel combinatorial strategies. Citation Format: Natalie Vokes, Stamatina Fragkogianni, Pooja A. Shah, Santiago G. Trevino, Konstantinos Leventakos, Lei Deng, Matthew Lee, Tali Azenkot, KAYLA LAYNG, JACOB MERCER, Jyoti Patel, Jordi Rodon Ahnert. High PRMT5 expression is associated with decreased immune infiltrate and worse outcomes to immune checkpoint inhibitors in non-small cell lung cancer [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 2446.

Abstract Background: The increasing volume and diversity of clinical genomic reports pose a significant challenge across healthcare institutions for the interpretation and proper implementation of precision oncology research. Reports from multiple different vendors (e.g., Invitae, Ambry Genetics, Foundation Medicine) are typically PDFs and exhibit substantial heterogeneity in panel design, gene coverage, and reporting standards, which hinders efficient retrospective data mining and patient cohort identification within the electronic medical records. We sought to develop a framework to interrogate all of these reports. Methods: We extracted genomic reports from patients with breast cancer treated with neoadjuvant chemotherapy and developed MolHarmonizer, a novel, scalable framework leveraging Python and Gemini LLMs, designed to process and harmonize genomic data from disparate multi-vendor reports. Gemini LLMs are employed explicitly for robust information extraction, normalization, and structuring of key genomic features, transforming unstructured data into a unified, queryable dataset. Results: Our MolHarmonizer framework successfully processed 1,147 genomic reports from 1703 breast cancer patients (2006-2023) from 23 different companies, demonstrating robust capability to extract and standardize critical actionable biomarkers. Data sources included Invitae (n=554), Ambry Genetics (n=189), Natera (n=95), Mayo Clinic (n=88), Tempus (n=63), Guardant Health (n=47), and Foundation Medicine (n=37), with others contributing less than 20 reports. Of the samples, 827/1147 (72.1%) were germline (blood/saliva). A majority of the patients were tested using the panels due to a personal/family history (n=925). Overall, 413/1147 (36.0%) reports identified at least one mutation. For breast cancer, 75 reports showed BRCA1/2 mutations (37 BRCA1, 37 BRCA2, and one patient with both BRCA1 and 2). Other mutations identified included: PIK3CA (n=40), TP53 (n=96), PTEN (n=21), ESR1 (n=13) and AKT1/2 (n=8). Conclusion: MolHarmonizer, a powerful framework leveraging Gemini LLMs, effectively addresses genomic data heterogeneity by automating biomarker extraction and harmonization. This enables rapid cohort identification and deep retrospective analyses for clinical insights, biomarker discovery, understanding disease history, facilitating novel pattern discovery, e.g., predicting BRCA1 mutations from WSI, and accelerating research within our neoadjuvant BC cohort. Future plans include expanding to include over 20,000 breast cancer patients, developing a user-friendly chatbot, and ensuring inter-institutional adaptability for a variety of complex diseases. Citation Format: Krishna Rani Kalari, Xiaojia Tang, Thanmayee Boyapati, Tanya L. Hoskin, Sumathilatha Myla, Sumedha G. Penheiter, Richard M. Weinshilboum, Liewei Wang, Hamid R. Tizhoosh, Karthik Vikram Giridhar, Matthew P. Goetz, Judy C. Boughey. Development of an LLM framework for analysis of heterogeneous breast cancer patients genomic reports [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 2756.

Abstract Large Language Models (LLM) are being widely adopted into the medical field for their impressive ability to analyze and summarize large amounts of text data. These models enable clinicians and researchers to extract meaningful insights from complex datasets and may assist with decision making. Here we present our workflows and application of LLM for the interpretation and summarization of clinical data related to CAR-T cell therapy. Using an LLM (Gemini 2.5 pro) in the Google Cloud Computing (GCP) environment, two applications were developed to analyze CAR-T cell therapy clinical data: 1) extracting and summarizing CRS and ICANS event-related data to streamline the compliance team workflow, and 2) identifying features available at time of CAR-T infusion able to classify patients into high- or low-monitoring needs 14 days post CAR-T. Patient data (vitals, labs, hematology notes, and EKGs) were extracted from the electronic medical record (EMR) using Google BigQuery into SQL tables in GCP. For each application, relevant data fields were retrieved, formatted into JSON objects, and embedded in the LLM prompt for context-aware processing. Both applications have undergone iterative prompt engineering after analyzing the LLM output against ground truth data in the EMR. For the application extracting and summarizing CRS and ICANS events, the LLM was optimized on Mayo Clinic Rochester data. When compared to the IEC compliance database, the LLM achieved 100% accuracy and F1 score for CRS events, and 96% accuracy and 82% F1 score for ICANS events. The match rate for CRS and ICANS grades were 83% and 89% respectively. We applied the same LLM to Mayo Clinic Arizona (MCA) and Mayo Clinic Florida (MCF) cases, who document clinical notes and toxicity flowsheet differently, and achieved an accuracy (MCA: 93%, MCF:93%) and F1 score (MCA: 96%, MCF: 96%) for CRS and accuracy (MCA: 81%, MCF:82%) and F1 score (MCA: 76%, MCF: 75%) for ICANS. LLM was able to capture events missed by manual reviews. For most of the discrepancies where compliance team final adjudication is needed, LLM will be updated to flag discrepancies for review by the compliance team. For the application related to the monitoring needs 14 days post CAR-T, the LLM identified 5 categories predictive of high or low monitoring needs post CAR-T infusion (disease status, inflammatory and tumor burden markers, hematologic status, renal function and performance status). Our model achieved a sensitivity of 85.7%, specificity of 23.8%, and F1 score of 65.5% for our first cohort, compared with data extracted by the LLM from the EMR. For our second cohort, with demographics statistically similar to cohort 1 and using the same 5 categories, the LLM achieved a sensitivity of 83.3%, a specificity of 28.6%, and F1 score of 65.4%. Our workflow and applications of LLM’s provide examples and guidance to others interested in applying LLM’s for clinical and research applications. Citation Format: Emmanuel Contreras Guzman, Matthew Jankowski, Andre De Menezes Silva Corraes, Malvika Gupta, Monica L. Shaw, Madiha Iqbal, Talal Hilal, Saurabh Chhabra, Ricardo Daniel Parrondo, Jody K. Mclean, Kim R. Riester, Kayla Joseph, Melinda Tan, Holly Ross, Cleyonia Barnett, Sylvia Carter, Semy Girmay, Rachel Wolan, Milana Ramsey, Christian Downhour, Kristy Morgan, Shae Sibley, Erica Rushing, Lucy Holmes, Allison Burgstahler, Stephen M. Ansell, Hassan Alkhateeb, Matthew Hathcock, Ramona Bruno, Allison C. Rosenthal, Hemant Murthy, Patrick B. Johnston, Jonas Paludo, Yi Lin. Applications of large language models to CAR-T cell therapy clinical data using Google cloud computing [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 2740.

Abstract Introduction: Pathological complete response (pCR) is a strong prognostic marker, but survival outcomes comparing treatment to control do not reliably align with treatment-control differences in pCR rates in breast cancer. A novel Bayesian hierarchical framework models treatments within trials, allowing us to predict treatment effects on distant recurrence-free survival (DRFS) from pCR with greater accuracy. Methods: We analyzed 12 neoadjuvant breast cancer trials (6,000 patients, all HR/HER2 subtypes; including I-SPY2). The framework of Burzykowski, Molenberghs & Buyse (2005) is extended to a novel arm-based hierarchical structure providing a distribution of pCR and DRFS treatment effects controlling for subtype (HR/HER2), N and T stage, grade, and calendar year. Three held-out trials (877 patients; 26 regimens; med follow-up >4 years) validate predictions of DRFS treatment benefit from pCR. All data was used to estimate treatment-effect correlation and the surrogate threshold effect (STE). Analyses were repeated for Residual Cancer Burden Index (binary RCB01, continuous RCB). Results: Predicted probability of DRFS benefit closely matched actual DRFS follow-up (mean absolute error 0.06; Pearson r = 0.9). Across all trials, pCR showed moderate surrogacy (ρ = 0.82, R2 = 0.67). A 60% increase in pCR odds achieves ≥95% probability of DRFS benefit (STE = OR of 1.60). RCB outperforms pCR across all metrics, with 92% sensitivity and 93% specificity for detecting DRFS benefit in 26 validation regimens (Table 1). Conclusion: This novel Bayesian meta-analytic framework reveals that pCR, RCB01 and continuous RCB reliably predict survival benefit across heterogeneous trials, with RCB performing the best in an external validation. This provides a statistical foundation for accelerated approval decisions using robust early biomarkers in modern neoadjuvant trial designs by accurately predicting survival at the treatment arm level. Citation Format: Keli S. Santos-Parker, Jessica R. Santos-Parker, W. Fraser Symmans, Laura J. Esserman, Christina Yau, Angie DeMichele, Laura van't Veer, Doug Yee, Fabien Reyal, Helena Earl, Jean Abraham, David Cameron, Peter Hall, Judy Boughey, Matthew Goetz, Gabe Sonke, Miguel Martín, Sara López-Tarruella, Priyanka Sharma, Rachel Freiberg, Jane Perlmutter, Aditya Bardia, Martin Eklund, Rachel Freiberg, Lajos Pusztai. A novel statistical framework for surrogate endpoint prediction of survival in neoadjuvant breast cancer trials [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 1401.