Kalari, Krishna R.

Background Tumor-infiltrating lymphocytes (TILs) are established determinants of clinical outcomes in breast cancer. 1–6 While stromal TILs (sTILs) are commonly evaluated due to easier scoring and higher interobserver reproducibility, intraepithelial TILs may provide additional insights into anti-tumor immunity. 7 We aimed to comprehensively characterize immunotype-associated cellular landscapes in triple-negative breast cancer (TNBC) and identify features associated with clinical outcomes. Methods 212 TNBC tumors were classified into three immunotypes based on immune infiltrate location and density: immune desert (ID), immune excluded (IE), and immune inflamed (IN). 8 Single-cell spatial transcriptomics was performed on 75 tumors to define the cellular composition and functional states of the tumor microenvironment across immunotypes. Findings were validated and integrated with bulk RNA sequencing data and clinical outcome datasets from FinXX, CALGB-40603, I-SPY 2, and Real-World Clinico-Genomic Data (RWCGD) cohorts. Results Patients with IN tumors had significantly improved outcomes compared with ID tumors. Despite high sTILs, IE tumors showed poor outcomes similar to ID tumors. Single-cell spatial analysis revealed that ID and IE tumors exhibited reduced major histocompatibility complex I/II expression and fewer tumor-resident plasmacytoid dendritic cells (pDCs) compared with IN tumors, which were enriched for interferon-alpha (IFNα) and interferon-gamma (IFNγ) responses. High IFN response scores were associated with favorable outcomes across multiple therapy types in independent datasets. Deconvolution of RWCGD bulk RNA-seq data confirmed that pDC abundance correlated with improved outcomes specifically in hormone receptor–negative subtypes. Conclusions Our study highlights pDCs and IFN signaling as hallmarks of effective anti-tumor immunity in TNBC. Immunotype-based profiling underscores the limited prognostic value of sTILs alone in immune-excluded tumors and supports pDCs and IFN pathways as potential biomarkers for prognosis and therapeutic development in TNBC.

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 Background: DNA methyltransferase (DNMT) isoforms 1, 3A and 3B are enzymes that regulate DNA methylation, controlling gene expression, chromatin stability and genetic imprinting. We have previously shown that DNMT3A protein expression in early triple negative breast cancer (TNBC) is associated with more aggressive clinicopathologic features (higher tumor grade, tumor proliferation and frequency of nodal metastases) and worse clinical outcomes in the absence of chemotherapy (Leon-Ferre et al, SABCS 2024). Pre-clinically, we have shown that DNMT3A protein expression predicts sensitivity to DNMT inhibitors (Yu J et al., J Clin Invest 2018), and that DNMT inhibitor treatment leads to downregulation of DNMT3A, enhanced expression of immune-related pathways, and halted tumor growth, particularly in combination with immunotherapy and paclitaxel. Methods: This is a phase 1, open-label, multicenter study evaluating the safety and recommended phase 2 dose (RP2D) of the oral DNMT inhibitor ASTX727 + paclitaxel + pembrolizumab in metastatic TNBC. ASTX727 is an oral fixed dose combination (FDC) tablet of 35 mg decitabine (active agent) and 100 mg cedazuridine (a cytidine deaminase inhibitor that prevents decitabine degradation in the bowel and liver, allowing oral delivery). The study consists of two parts: Part 1 is a dose-finding cohort, using a 3+3 design, with 5 dose levels (Table), and its primary objective is to determine the RP2D. Part 2 is an expansion cohort (n=6), with the primary objective to confirm tolerability of the regimen at the RP2D. Secondary objectives are to record the antitumor activity of the regimen, describe the AE profile, explore the association of baseline gene expression profiles with clinical benefit, and to evaluate impact of treatment on tumor and ctDNA methylation. Exploratory objectives will evaluate the association of baseline DNMT3A protein expression and antitumor activity, and the impact of treatment on the tumor and peripheral blood immune phenotype. Eligible patients are adults 18 years or older with metastatic TNBC (ER/PR ≤10%, and HER2-negative per ASCO/CAP guidelines), independent of PD-L1 expression. Patients are allowed to have received 0-3 prior lines of systemic therapy in the metastatic setting, including prior immune checkpoint inhibitors in the early and/or metastatic settings. Patients must be eligible for taxanes and have adequate organ and bone marrow function. Patients with treated brain metastases are eligible if previously treated without evidence of progression after 4 weeks. Leptomeningeal carcinomatosis is not allowed. Patients must have an ECOG performance status of ≤2 and have no gastrointestinal disorders impacting absorption of oral medications. As of June 20, 2025, the study is enrolling patients, with 10 patients dosed in the dose-finding phase. Citation Format: R. Leon-Ferre, D. M. Zahrieh, P. P. Advani, D. Quiroga, K. V. Giridhar, P. J. Dizona, D. Stover, X. Wang, A. J. Tevaarwerk, G. M. Choong, S. Yasir, K. R. Kalari, J. Foldi, R. Parajuli, M. H. Hackney, H. K. Chew, R. Wajeeha, H. Bear, E. Sharon, B. Ko, G. I. Shapiro, S. Gore, L. Pelosof, L. Wang, M. P. Goetz. Trial in progress: Phase 1 Study Targeting DNA Methyltransferases in Metastatic Triple-Negative Breast Cancer (NCI Protocol #10546) [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 PS5-07-20.

Integrating artificial intelligence (AI) with healthcare data is rapidly transforming medical diagnostics and driving progress toward precision medicine. However, effectively leveraging multimodal data, particularly digital pathology whole slide images (WSIs) and genomic sequencing, remains a significant challenge due to the intrinsic heterogeneity of these modalities and the need for scalable and interpretable frameworks. Existing diagnostic models typically operate on unimodal data, overlooking critical cross-modal interactions that can yield richer clinical insights. We introduce MarbliX (Multimodal Association and Retrieval with Binary Latent Indexed matriX), a self-supervised framework that learns to embed WSIs and immunogenomic profiles into compact, scalable binary codes, termed “monogram.” By optimizing a triplet contrastive objective across modalities, MarbliX captures high-resolution patient similarity in a unified latent space, enabling efficient retrieval of clinically relevant cases and facilitating case-based reasoning. In lung cancer, MarbliX achieves 85%–89% across all evaluation metrics, outperforming histopathology (69%–71%) and immunogenomics (73%–76%). In kidney cancer, real-valued monograms yield the strongest performance (F1: 80%–83%, Accuracy: 87%–90%), with binary monograms slightly lower (F1: 78%–82%).

Abstract Cutaneous squamous cell carcinoma (cSCC) poses significant clinical challenges due to its rising incidence and potential for metastasis. Histopathologic risk stratification is further limited by substantial inter-observer variability. Unsupervised AI approaches based on content-based image retrieval offer scalable and interpretable decision support for diagnostic pathology. The objective of this study was to evaluate the use of image retrieval within histopathology atlases to stratify cSCC tumor differentiation from whole-slide images (WSIs), while comparing different patch selection and feature extraction strategies. This retrospective study included 552 archived WSIs comprising 385 well-differentiated, 102 moderately differentiated, and 66 poorly differentiated cases collected across Mayo Clinic sites in Arizona, Florida, and Minnesota. Image atlases were constructed using multiple patch aggregation strategies (Mosaic, Collage, and Montage) and deep learning encoders (KimiaNet, PathDino, and H-Optimus-0). A leave-one-WSI-out evaluation framework was used to assess differentiation classification performance using accuracy, specificity, sensitivity, and F1 score. Mosaic combined with KimiaNet achieved the highest Top-1 accuracy (74.9%) and specificity (92.6%), while Mosaic with H-Optimus-0 yielded the best Top-5 accuracy (79.0%) and macro-F1 score (62.6%). Collage combined with KimiaNet produced the highest Top-5 specificity (99.5%). The generalizability of the evaluated AI models varied across hospitals, reflecting differences in imaging protocols, staining practices, and patient populations. Overall, unsupervised image search and retrieval provides effective, annotation-free support for cSCC differentiation and has the potential to enhance dermatopathology workflows when appropriate combinations of patch selection and feature ex-traction methods are employed.

Abstract Background Plasma biomarkers offer improved accessibility compared to established diagnostic modalities for Alzheimer's disease (AD). However, current plasma biomarker candidates have not been tested comprehensively in diverse populations, and represent only a fraction of pathways dysregulated in AD. This study seeks to address these knowledge gaps through untargeted plasma biomarker discovery in African Americans (AA), an underrepresented group in AD research despite having a higher risk of developing dementia compared to non‐Hispanic Whites (NHW). Method Untargeted proteomics were generated using the SomaScan 7k platform on plasma samples from AA study participants clinically diagnosed as AD dementia ( n = 181) or cognitively unimpaired (CU; n = 142). Machine learning was implemented to identify a set of plasma proteins to distinguish between AD dementia and CU. Predictive models were constructed by elastic net regression optimized with nested cross‐validation on 6,853 proteins and evaluated by receiver operating characteristic analysis. The generalizability of plasma protein predictors was evaluated in an external Alzheimer's Disease Research Center cohort of 369 NHW and AA study participants diagnosed as AD dementia or CU. Result Thirty‐six proteins were selected as predictors. Proteins selected by machine learning were enriched for processes previously associated with AD including synaptic dysfunction, inflammation, vascular dysfunction, and amyloid processing. The 36 plasma proteins alone achieved an area under the curve (AUC) of 0.94 to classify AD dementia in the discovery cohort, representing a 22% improvement over a model including age and sex. The 36 proteins were validated in an external cohort also achieving AUC=0.94, a 35% improvement over age and sex. Conclusion Biomarker discovery combining untargeted proteomics and machine learning identified a set of plasma proteins that could serve as a novel and highly accurate diagnostic plasma biomarker panel for AD. Work is underway to assess the association of the predictive panel with p ‐tau217, medical comorbidities, and social/structural determinants of health.

Abstract Deconvolution models are a powerful tool for extracting cell-type-specific information from bulk gene expression profiles. Current methods leverage advanced machine learning models and high-resolution sequencing, like single-cell RNA-sequencing, showing promising results across diverse tissues and conditions. However, they still present important limitations: First, many depend on selecting a robust reference, which can strongly affect the deconvolution. Second, pseudobulk data used for training and real bulk RNA-seq samples often exhibit strong distribution shifts, which are currently unaccounted for. Finally, most deconvolution approaches behave as black boxes, which can compromise the reliability of the results. Here, we present Sweetwater, an adaptive and interpretable autoencoder that efficiently deconvolves bulk samples leveraging multiple classes of reference data. Moreover, we propose an improved way of generating training data from a mixture of FACS-sorted FASTQ files, reducing platform-specific biases and outperforming current single-cell-based references. Furthermore, we introduce a gold standard dataset to facilitate fair and accurate evaluation of deconvolution approaches. Finally, we demonstrate that Sweetwater adapts effectively to deconvolved samples during training, uncovering biologically meaningful patterns and enhancing result’s reliability. Sweetwater is available at https://doi.org/10.6084/m9.figshare.29609180, and we anticipate it will expedite the accurate examination of high-throughput clinical data across diverse applications.

Abstract INTRODUCTION African American (AA) individuals are underrepresented in biomarker studies for Alzheimer's disease (AD). Biomarkers that reflect the heterogeneity of AD and achieve the greatest accuracy across populations are sorely needed. METHODS Untargeted proteome measurements were obtained using the SomaScan 7k platform to identify novel plasma biomarkers for AD in AA participants with clinical diagnoses of AD dementia ( n = 181) and cognitively unimpaired (CU, n = 142). Machine learning was used to identify a set of plasma proteins that yielded the best classification accuracy. RESULTS A set of 36 proteins achieved an area under the curve (AUC) of 0.94 to classify AD dementia versus CU, a 16% improvement over age, sex, and apolipoprotein E ( APOE) . This finding was replicated in multiple plasma and brain datasets (AUCs 0.73–0.97). Our findings underscore the importance of matrisome and cerebrovascular dysfunction in AD pathophysiology. DISCUSSION This study demonstrates the potential of biomarker discovery through untargeted plasma proteomics and machine learning. Highlights Conducted large‐scale plasma proteomics in Alzheimer's disease (AD) versus cognitively unimpaired controls. Machine learning biomarker discovery was replicated in an independent cohort. Novel set of proteins distinguishes AD versus controls with high accuracy (area under the curve [AUC] = 0.94). Achieved reproducibility across multiple replication cohorts (AUC = 0.73–0.97). Network analyses implicates matrisome biology and cerebrovascular dysfunction.

e14545 Background: Next generation sequencing (NGS) identifies genomic alterations in DNA obtained from tissue biopsies and/or blood [circulating tumor (ct)DNA]. We anticipate a lower rate of copy number loss (CNL) in ctDNA compared to tissue, as lower ctDNA tumor fraction may limit the ability to detect a deletion. However, we observed several instances of a CNL in ctDNA that was not identified in a tissue specimen, prompting this systematic evaluation of an institutional dataset. We focused on BRCA1/2 pathogenic variants (PVs) given their therapeutic implications and coverage across testing platforms. Methods: Using an IRB-approved protocol, we evaluated an institutional repository of NGS results performed across the Mayo Clinic Enterprise. We focused on patients (pts) who had both tissue and blood NGS testing performed. We further narrowed this cohort to those with a BRCA1/2 PVs, somatic and/or germline, as determined by testing performed at Guardant Health, Tempus, or Foundation Medicine. We abstracted information about each variant, including type of alteration classification and functional impact. Demographic information, cancer characteristics, and germline testing information was collected. Fisher’s exact test was used to compare alteration frequencies in ctDNA vs tissue using GraphPad Prism v 10.1. Results: From Dec 2015-Sept 2024, we identified 158 unique pts with BRCA1/2 PV who had NGS results for both tissue and ctDNA. The median age at first reporting BRCA1/2 variant was 67 years (range: 29-98 years). 48% of pts were female, and the most common cancer type sites were lung (23%), colorectal (20%), and pancreaticobiliary (20%). We identified 198 unique pathogenic genomic alterations; 71 in BRCA1 & 127 in BRCA2 . Most common alteration was an SNV or indel (127/198), followed by CNL (n = 63), splice site variants (n = 5), fusions (n = 2), and rearrangements (n = 1). Of the 126 SNV or indels, 50 (39.7%) were identified in both ctDNA and tissue, 40 (31.7%) were detected in blood only, 36 (28.6%) identified only in tissue (p < 0.0001). Amongst the 63 CNL, only 3 (4.8%) were identified in both ctDNA and tissue; 41(65.1%) were only identified in ctDNA, and 19 (30.2%) in tissue only (p < 0.0001). In the 19 tissue-only CNL tests, 12/19 noted one or more additional pathogenic SNV/Indel in BRCA1/2. In the 41 ctDNA-only CNL tests, 1/41 noted one or more additional pathogenic SNV/Indel in BRCA 1/2 . Germline testing was available in 41/63 pts with copy number loss. 3/41 deletions were identified in germline (2 in ctDNA, 1 in tissue). Conclusions: We observed a low degree of concordance in CNL between tissue and ctDNA for BRCA1/2 . CNL detection in ctDNA that could not be confirmed in germline or with a second alteration warrants further attention. In particular, clinicians should be cautioned when interpreting ctDNA reports noting BRCA1/2 CNL without clarity if this represents biallelic loss or a synchronous loss of function.

548 Background: TNBC is associated with higher risk of recurrence and poorer survival rates than other breast cancer subtypes. For node-positive TNBC or tumors larger than 0.5 cm, chemotherapy is generally recommended. For stage I TNBC, the benefit from chemotherapy, the optimal regimen, and whether to administer it in the neoadjuvant versus adjuvant setting remain controversial. Here, we report the treatment patterns and clinical outcomes of patients with stage I TNBC treated at Mayo Clinic. Methods: Using the Mayo Clinic Tumor Registry data, we identified patients with stage I TNBC treated between 2010 and 2021. We used the pathologic tumor (T) category for patients treated with upfront surgery and the clinical T category for patients receiving neoadjuvant chemotherapy. We used the Kaplan-Meier method and log-rank test to compare recurrence-free survival (RFS) according to treatment groups, measured from the day of surgery. RFS was reported at a median follow-up of 3.9, 6.2, and 7.3 years for patients who received chemotherapy neoadjuvantly, adjuvantly, and surgery alone with no chemotherapy, respectively. Results: A total of 602 patients with Stage I TNBC were included, with a median age of 62 years. 290 (48%) underwent upfront surgery followed by adjuvant chemotherapy (pT1a: 2%, pT1b: 25%, pT1c: 74%), 127 (21%) received neoadjuvant chemotherapy followed by surgery (cT1a: 2%, cT1b: 11%, cT1c: 87%), and 185 (31%) underwent primary surgery without any chemotherapy (pT1mi/T1a: 33%, pT1b: 32%, pT1c: 35%). Most patients treated with neoadjuvant therapy received anthracycline/cyclophosphamide + taxane (70%), while most patients treated in the adjuvant setting received a non-anthracycline containing regimen (60%). The 5-year RFS was 96% (95% CI: 93-100%) for patients who received chemotherapy neoadjuvantly, 95% (95% CI: 92-98%) for those who received chemotherapy adjuvantly, and 85% (95% CI: 80-91%) for those treated with surgery alone and no chemotherapy ( P < 0.0001). 71 (56%) of the 127 patients who received neoadjuvant therapy achieved a pCR, with only 1 RFS event (at 14 months) in that group, compared to 4 RFS events among those not achieving pCR (at 12, 20, 28 and 75 months, respectively). Among patients not receiving chemotherapy (either adjuvantly or neoadjuvantly), the 5-year RFS rates were 94% for pT1mi/T1a, 92% for pT1b and 72% for pT1c (T1mi/a/b vs T1c, P = 0.009). Conclusions: In this large cohort of stage I TNBC, patients who received chemotherapy (adjuvantly or neoadjuvantly) had better 5-year RFS compared with those treated with locoregional therapy only. Among patients receiving chemotherapy, 5-year RFS was nearly identical regardless of whether chemotherapy was administered before or after surgery. Interestingly, patients undergoing upfront surgery were more likely to receive an anthracycline-sparing chemotherapy regimen.