Citations

Abstract Deep learning models enable the prediction of clinical endpoints from whole-slide images (WSIs), but many such models function as “black boxes”, lacking transparency about whether and which histomorphological patterns drive their predictions, hindering interpretability and clinical adoption. Here we propose a human-in-the-loop explanation framework, MorphoXAI, which provides both local and global interpretability for deep learning models by incorporating human-expert interpretations. At the global level, it reveals the histomorphological patterns on which the model consistently relies to distinguish between classes of WSIs, as well as the patterns associated with confusion between classes. At the local level, it indicates which of these patterns are used in the prediction of an individual WSI and which regions within the slide correspond to such patterns. We validated our method on a deep learning model trained for ovarian tumor histologic subtype prediction. The results show that our framework generates explanations that accurately reflect the histomorphology underlying the model’s predictions at both global and local levels. For interpretability and clinical utility in diagnostic contexts, human evaluation results showed that our explanations were easy to interpret, rich in diagnostic features, and directly helpful for diagnostic decision-making, thereby enhancing pathologist-AI collaboration. Our work highlights that unifying global and local explanations and grounding them in expert-interpreted morphology enhances the interpretability and verifiability of deep learning models, thereby facilitating the transparent deployment of such models in clinical practice.

Abstract Disclosure: A. Zhang: None. J. Kapoor: None. C. Deveney: None. A.G. Cope: None. M.M. Walther-Antonio: None. N. Chia: None. C.W. Pattanayak: None. M.J. Tetel: None. The vaginal microbiome (VM) consists of the microorganisms and their metabolites residing in the vaginal tract and is critical for maintaining female reproductive health. The Lactobacillus species predominate most VMs and produce lactic acid, which has been cited as a key protector of vaginal health. Alterations in the finely balanced vaginal microbiome can cause infections, such as bacterial vaginosis, and may contribute to the development of endometrial and ovarian cancer. While it is well-established that stress and the gut microbiome impact each other via the gut-brain axis, the influence of stress, hormones, and lifestyle factors on the VM composition remains poorly understood. We investigated the relationship between stress, hormones (menstruation, hormonal contraception), lifestyle factors (diet, exercise, sleep), and the VM composition. In this 10-week study, healthy, young female volunteers (n=62) provided daily vaginal and fecal swabs, weekly stress ratings, daily data on sleep and activity (via FitBit) and diet information. High-throughput 16S rRNA sequencing was used to identify the vaginal microbial composition of each sample. Types of hormonal birth control were none (n=30), local progestin (n=9), systemic progestin-only (n=7), or systemic estrogens and progestin combined (n=16). Consistent with our previous study, vaginal microbial diversity, as measured by the Shannon index, varied by type of hormonal birth control, with participants not on hormonal birth control having the highest average diversity. There was a positive relationship between average stress and vaginal microbial diversity, with preliminary results suggesting that this relationship depends on hormonal birth control status. Participants who menstruated during the 10-week study had a higher vaginal microbial diversity than those who did not. We are currently analyzing which microbes contribute to these differences in vaginal microbial diversity. Microbiota and fungal data from the daily vaginal and fecal swab samples will be analyzed to determine how stress, hormones and/or lifestyle factors relate to crosstalk between the vaginal and gut microbiomes. These findings may provide long-term, time-longitudinal insights into how these two important microbiomes are involved in important aspects of women’s health. Presentation: Sunday, July 13, 2025

Abstract In this work, we seek to characterize the learned feature space of cancer drug response prediction models in a drug-blind setting and quantify the limits of their generalizability. Drug-blind prediction failure describes the inability of models to predict cell line response to drugs in the test set that are not present during training. The experiments performed in this study utilize a two-arm multilayer perceptron model where embeddings for drugs and cell-lines are calculated separately and then concatenated to predict response. Drug structure is represented by Morgan fingerprint while cell lines utilize gene expression values. We first examine the connection between learned drug features and cell lines by permuting responses within both cell lines and drugs during model training. When permuting response values within cell lines, model performance was entirely depleted, but permutation of responses within drugs resulted in only a 10-15% decrease in performance. This displays the bulk of model performance is due to learned distributions of response for each drug. We also determine the impact of dataset size on drug-blind performance. Increasing the cell line examples per drug did not improve performance, but drug-blind performance exhibited higher variance than mixed set testing. From these experiments, we hypothesized that drug-blind performance was a function of the set of drugs in the training set. We trained a set of 208 models with varying training set and constant test set and fit an elastic net model using drugs in the training set as features and performance of each model as the target. We were able to accurately predict the drug-blind performance of a model, measured in Pearson correlation, based on the drugs in the training set with a mean absolute error of 0.049. We then trained a set of 1641 models where all drugs in the dataset were present in the test set but only 50% of unique drugs were present during training. Hierarchical clustering on the coefficients of an elastic net model fit to each unique drug’s performance across all models shows drugs cluster into groups that correspond to specific mechanisms of action. By examining the relationship between drug embeddings during model training, we see that decreasing validation loss corresponds to reinforcement of mechanistic relationships of drugs not captured by raw Morgan fingerprints. Finally, we show that training on a dataset confined to a single mechanism of action significantly improves overall mixed set performance on those drugs against training on the entire set of drugs. In this study, we identify that drug-blind performance in current large pharmacogenomic datasets is limited to generalization within drug mechanism of action classes. Therefore, global drug-blind prediction benchmarking is a poor indicator of model generalization as the data itself creates these limits, not model architecture. As more data is collected on novel anti-cancer compounds, we hope that the results presented here create a foundation on which to measure progress in cancer drug generalization. Citation Format: William G. Herbert, Paul A. Jensen, Nicholas Chia, Marina RS. Walther-Antonio. Monotherapy cancer drug-blind response prediction is limited to intraclass generalization [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A005.

Abstract DNMT3A and TET2 are epigenetic regulator genes commonly mutated in age-related clonal hematopoiesis (CH). Despite having opposed epigenetic functions, these mutations are associated with increased all-cause mortality and a low risk for progression to hematologic neoplasms. Although individual impacts on the epigenome have been described using different model systems, the phenotypic complexity in humans remains to be elucidated. Here, we make use of a natural inflammatory response occurring during coronavirus disease 2019 (COVID-19), to understand the association of these mutations with inflammatory morbidity (acute respiratory distress syndrome [ARDS]) and mortality. We demonstrate the age-independent, negative impact of DNMT3A mutant (DNMT3Amt) CH on COVID-19–related ARDS and mortality. Using single-cell proteogenomics we show that DNMT3A mutations involve myeloid and lymphoid lineage cells. Using single-cell multiomics sequencing, we identify cell-specific gene expression changes associated with DNMT3A mutations, along with significant epigenomic deregulation affecting enhancer accessibility, resulting in overexpression of interleukin-32 (IL-32), a proinflammatory cytokine that can result in inflammasome activation in monocytes and macrophages. Finally, we show with single-cell resolution that the loss of function of DNMT3A is directly associated with increased chromatin accessibility in mutant cells. Hence, we demonstrate the negative prognostic impact of DNMT3Amt CH on COVID-19–related ARDS and mortality. DNMT3Amt CH in the context of COVID-19, was associated with inflammatory transcriptional priming, resulting in overexpression of IL32. This overexpression was secondary to increased chromatic accessibility, specific to DNMT3Amt CH cells. DNMT3Amt CH can thus serve as a potential biomarker for adverse outcomes in COVID-19.

Abstract Background DNMT3A and TET2 are epigenetic regulator genes commonly mutated in age-related clonal hematopoiesis (CH). Despite having opposing epigenetic functions, these mutations are associated with increased all-cause mortality and a low risk for progression to hematological neoplasms. While individual impacts on the epigenome have been described using different model systems, the phenotypic complexity in humans remains to be elucidated. Results Here we make use of a natural inflammatory response occurring during coronavirus disease 2019 (COVID-19), to understand the association of these mutations with inflammatory morbidity and mortality. We demonstrate the age-independent, negative impact of DNMT3A mutant CH on COVID-19-related cytokine release severity and mortality. Using single cell proteogenomics we show that DNMT3A mutations involve myeloid and lymphoid cells. Using single cell multiomics sequencing, we identify cell-specific gene expression changes associated with DNMT3A mutations, along with significant epigenomic deregulation affecting enhancer accessibility, resulting in overexpression of IL32, a proinflammatory cytokine that can result in inflammasome activation in monocytes and macrophages. Finally, we show with single cell resolution that the loss of function of DNMT3A is directly associated with increased chromatin accessibility in mutant cells. Conclusions We demonstrate the negative prognostic impact of DNMT3Amt CH on COVID-19 related inflammatory morbidity and mortality. DNMT3Amt CH involves myeloid and lymphoid cells and in the context of COVID-19, was associated with inflammatory transcriptional priming, resulting in overexpression of IL32. This overexpression was secondary to increased chromatic accessibility, specific to DNMT3Amt CH cells. DNMT3Amt CH can serve as a potential biomarker for adverse inflammatory outcomes.

Objective The aim of this study was to assess the diagnostic utility of cerebrospinal fluid (CSF) myelin oligodendrocyte glycoprotein antibodies (MOG‐IgG) testing. Methods We retrospectively identified patients for CSF MOG‐IgG testing from January 1, 1996, to May 1, 2023, at Mayo Clinic and other medical centers that sent CSF MOG‐IgG for testing including: controls, 282; serum MOG‐IgG positive MOG antibody‐associated disease (MOGAD), 74; serum MOG‐IgG negative high‐risk phenotypes, 73; serum false positive MOG‐IgG with alternative diagnoses, 18. A live cell‐based assay assessed CSF MOG‐IgG positivity (IgG‐binding‐index [IBI], ≥2.5) using multiple anti‐human secondary antibodies and end‐titers were calculated if sufficient sample volume. Correlation of CSF MOG‐IgG IBI and titer was assessed. Results The pan‐IgG Fc‐specific secondary was optimal, yielding CSF MOG‐IgG sensitivity of 90% and specificity of 98% (Youden's index 0.88). CSF MOG‐IgG was positive in: 4/282 (1.4%) controls; 66/74 (89%) serum MOG‐IgG positive MOGAD patients; and 9/73 (12%) serum MOG‐IgG negative patients with high‐risk phenotypes. Serum negative but CSF positive MOG‐IgG accounted for 9/83 (11%) MOGAD patients, and all fulfilled 2023 MOGAD diagnostic criteria. Subgroup analysis of serum MOG‐IgG low‐positives revealed CSF MOG‐IgG positivity more in MOGAD (13/16[81%]) than other diseases with false positive serum MOG‐IgG (3/15[20%]) ( p = 0.01). CSF MOG‐IgG IBI and CSF MOG‐IgG titer (both available in 29 samples) were correlated (Spearman's r = 0.64, p < 0.001). Interpretation CSF MOG‐IgG testing has diagnostic utility in patients with a suspicious phenotype but negative serum MOG‐IgG, and those with low positive serum MOG‐IgG results and diagnostic uncertainty. These findings support a role for CSF MOG‐IgG testing in the appropriate clinical setting. ANN NEUROL 2024;96:34–45

Abstract Given the morbidity and mortality associated with colorectal cancer, novel biomarkers are clearly warranted, especially for early detection, particularly for advanced adenomas. An earlier 1317-plex aptamer-based proteomic screen of stool proteins revealed 92 proteins to be significantly elevated in CRC stool, with the most discriminatory proteins being MMP9, haptoglobin, myeloperoxidase, fibrinogen, and adiponectin (PMID: 34117903). Results from a recent 2000-plex proteomic screen of CRC stools are detailed below, while a 11,000-plex aptamer based proteomic screen is in progress. Comprehensive proteomic screening revealed 116 proteins to be differentially expressed in CRC stool compared to control stool, with 45 being elevated 2-fold or higher. 37 of these proteins were ELISA validated in 3 independent cohorts. Stool MMP-8, MMP-9, Hemoglobin, PGRP-S, Haptoglobin, Myeloperoxidase and Fibrinogen emerged as being most discriminatory for distinguishing CRC from healthy controls (AUC: 0.91-0.95), while Stool Fibrinogen, MMP-9, Hemoglobin, MMP-8, and PGRP-S best distinguished advanced adenoma from healthy control (HC) stool, with stool Fibrinogen topping the list with an ROC AUC value of 0.86, as detailed in Table 1. Functional pathway analysis revealed a significant over-representation of pathways related to anti-oxidant activity, integrin/receptor binding, cytokines, blood coagulation, and lipoprotein biosynthesis in patients with CRC compared to HC. Nuclear Factor IC (NFIC) and IKZF2 were identified as key regulators of the molecular cascades over-represented in CRC. Stool Fibrinogen, MMP-8, MMP-9, PGRP-S, Haptoglobin, and Myeloperoxidase emerge as promising biomarkers for distinguishing CRC/advanced adenomas/healthy stools, outperforming Hemoglobin. Thus, comprehensive proteomics and independent ELISA validation have uncovered novel stool biomarkers for identifying CRC and advanced adenomas, that warrant further head to head comparison against current diagnostic tests. Citation Format: Chandra Mohan, Kamala Vanarsa, Jessica Castillo, Robert Bresalier, Nicholas Chia. 11,000-plex aptamer based proteomics of colorectal cancer stools for the discovery of novel disease biomarkers [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 1845.

AbstractObjectiveSeizures are a common manifestation of paraneoplastic neurologic syndromes. The objective of this study was to describe the seizure characteristics and outcomes in patients with high‐risk paraneoplastic autoantibodies (>70% cancer association) and to determine factors associated with ongoing seizures.MethodsPatients from 2000 to 2020 with seizures and high‐risk paraneoplastic autoantibodies were retrospectively identified. Factors associated with ongoing seizures at last follow‐up were evaluated.ResultsSixty patients were identified (34 males, median age at presentation = 52 years). ANNA1‐IgG (Hu; n = 24, 39%), Ma2‐IgG (n = 14, 23%), and CRMP5‐IgG (CV2; n = 11, 18%) were the most common underlying antibodies. Seizures were the initial presenting symptom in 26 (43%), and malignancy was present in 38 (63%). Seizures persisted for >1 month in 83%, and 60% had ongoing seizures, with almost all patients (55/60, 92%) still being on antiseizure medications at last follow‐up a median of 25 months after seizure onset. Ongoing seizures at last follow‐up were associated with Ma2‐IgG or ANNA1‐IgG compared to other antibodies (p = .04), highest seizure frequency being at least daily (p = .0002), seizures on electroencephalogram (EEG; p = .03), and imaging evidence of limbic encephalitis (LE; p = .03). Death occurred in 48% throughout the course of follow‐up, with a higher mortality in patients with LE than in those without LE (p = .04). Of 31 surviving patients at last follow‐up, 55% continued to have intermittent seizures.SignificanceSeizures in the setting of high‐risk paraneoplastic antibodies are frequently resistant to treatment. Ongoing seizures are associated with ANNA1‐IgG and Ma2‐IgG, high seizure frequency, and EEG and imaging abnormalities. Although a subset of patients may respond to immunotherapy and achieve seizure freedom, poor outcomes are frequently encountered. Death was more common among patients with LE.

The snATAC + snRNA platform allows epigenomic profiling of open chromatin and gene expression with single-cell resolution. The most critical assay step is to isolate high-quality nuclei to proceed with droplet-base single nuclei isolation and barcoding. With the increasing popularity of multiomic profiling in various fields, there is a need for optimized and reliable nuclei isolation methods, mainly for human tissue samples. Herein we compared different nuclei isolation methods for cell suspensions, such as peripheral blood mononuclear cells (PBMC, n = 18) and a solid tumor type, ovarian cancer (OC, n = 18), derived from debulking surgery. Nuclei morphology and sequencing output parameters were used to evaluate the quality of preparation. Our results show that NP-40 detergent-based nuclei isolation yields better sequencing results than collagenase tissue dissociation for OC, significantly impacting cell type identification and analysis. Given the utility of applying such techniques to frozen samples, we also tested frozen preparation and digestion (n = 6). A paired comparison between frozen and fresh samples validated the quality of both specimens. Finally, we demonstrate the reproducibility of scRNA and snATAC + snRNA platform, by comparing the gene expression profiling of PBMC. Our results highlight how the choice of nuclei isolation methods is critical for obtaining quality data in multiomic assays. It also shows that the measurement of expression between scRNA and snRNA is comparable and effective for cell type identification.