2026 ASH®
Annual Meeting & Exposition December 12-15, 2026 ▪ New Orleans, Louisiana

Paper Details

Abstracts, authors, affiliations, and disclosures.

Sunday, December 13, 2026
06:00 p.m. - 08:00 p.m.
MCCNO - Hall F
Technology and Procedures, Assays, Imaging, Molecular Testing, Omics Technologies, Pathology, Bioinformatics, Clinical Procedures, Computational Biology, Emerging Technologies, Gene Editing, Profiling, Treatment Considerations, Drug Development, Therapy Sequence, Biological Therapies, Transplantation, (CAR)-T Cell Therapies, Workforce, Adverse Events, Artificial Intelligence (AI), Computer Vision, Deep Learning, Generative AI, Machine Learning, Natural Language Processing, Biological Processes, Pathogenesis, CHIP, Genomics, Hematopoiesis, Immune Mechanism, Metabolism, Microbiome, Molecular Biology, Multi-Systemic Interactions, Clinical Practice (Health Services and Quality), Quality Improvement, Systems-Based Hematology, Diseases, Bone Marrow Failure Syndromes, Lymphoid Malignancies, Lymphoid Leukemias, Lymphomas, Plasma Cell Disorders, Myeloid Malignancies, Acute Myeloid Malignancies, Chronic Myeloid Malignancies, Metabolic Disorders, Measurable Residual Disease, ALL, CLL, LGL, Hodgkin Lymphoma, Research, Fundamental Science, AML, APL, Translational Research, MDS, MPN, Clinical Research, Clinical Trials, Epidemiology, Health Outcomes Research, Real-World Evidence, Registries, Survivorship, Study Population, Human, Adult, Elderly
PAPER

4340 Bespoke hierarchical analyses of high-dimensional spectral flow cytometry data facilitated by AI agents

Gabriel Salzman, MD, PhD, Dan Corin2, Cameron S. Bader, PhD3, Crystal Zhou, BA4, Everett Meyer, MD, PhD5, Tian Zhang, MD PhD6
1 Division of Hematology, Stanford University School of Medicine, Stanford, California, United States, Division of Medical Oncology, Stanford University School of Medicine, Stanford, California, United States
2 wvlen, LLC, Brooklyn, New York, United States
3 Division of Blood and Bone Marrow Transplantation and Cellular Therapy, Stanford, California, United States
4 Division of Hematology, Stanford University School of Medicine, Stanford, California, United States, Stanford Cancer Center, Stanford, California, United States
5 Division of Blood and Bone Marrow Transplantation and Cellular Therapy, Stanford, California, United States
6 Division of Hematology, Stanford University School of Medicine, Stanford, California, United States

Spectral flow cytometry (SFC) has enabled investigators to inexpensively generate datasets comprising expression of >40 proteins (markers) across >107 cells. Despite their richness, SFC datasets are frequently analyzed with the same manual hypothesis-testing approaches used for classical flow cytometry, which involve comparing known cell populations across samples. These methods rely on the traditionally subjective and manual assignment of fluorescence thresholds, known as gates, that define cells positive or negative for each marker. Importantly, bioinformaticians have developed numerous distinct computational techniques to explore the large hypothesis space in SFC and mass cytometry datasets, many of which utilize similar hierarchical clustering approaches to those used in single-cell transcriptomics analysis. However, with conventional computational tools, the time and effort required to effectively implement even one of these approaches with SFC data is nontrivial, requiring tuning of multiple dataset-specific parameters like cluster number or method-specific resolution. Furthermore, each clustering method imposes a unique set of assumptions about the shape, scale, and resolution of the underlying data. Thus, even with a single well-tuned hierarchical clustering analysis method, only one representation of the dataset is elucidated, potentially failing to identify certain cell populations as candidates for hypothesis generation. Here, we outline an analysis paradigm that utilizes artificial intelligence (AI) agents to identify, tune, and implement a broad set of data analysis approaches to surface candidates for hypothesis generation, which we then manually reproduce. We go on to incorporate machine-learning (ML) models trained on each marker to automatically determine gates such that cell populations could be labeled and compared across analysis approaches.

We implemented our approach in four steps. First, we determined that upon cell population identification, we would compute differential abundance (DA) across sample groups (i.e. a particular cluster is enriched in a remission sample relative to a diagnostic sample). Critically, DA of any clearly identified target population can be quickly repeated manually using FlowJo software and the same raw data. Second, we instructed AI agents to broadly search the literature for hierarchical clustering and population identification approaches that would be amenable to SFC data. Third, for each approach, AI agents performed a parameter-tuning sweep to optimize cluster stability and consistency across variably down-sampled data. Fourth, DA analysis was performed and results from selected clustering approaches were aggregated and visualized using an AI agent-built customizable dashboard that enabled investigators to switch among complementary population representations, rank candidates by effect size and statistical support, and drill down into phenotype and marker expression.

We used two SFC datasets with the same 28-color panel for analysis pipeline development and ML model training and validation. The first included 1 million cells across n=12 patient samples with manually drawn gates for all markers. The second included 11 million cells across n=50 patient samples plus n=3 healthy donor controls. Eleven population-discovery approaches were incorporated including FlowSOM clustering (Van Gassen, et al., Cytometry A. 2015) as well as other graph-based and topological methods. We also employed a multiresolution hierarchical clustering algorithm, inspired by treekoR (Chan, et al., Genome Biol. 2021), in which FlowSOM nodes were aggregated and DA was evaluated at multiple resolutions. This algorithm provided the most precise marker-based definition of a population (later confirmed by manual DA analysis in FlowJo) that was also identified by other approaches, albeit less precisely.

Investigator-directed AI made it feasible to search for, implement, and tune multiple clustering modalities which together generated more compelling and explicit hypothesis candidates than any single modality. AI agents also enabled the training of a customized supervised neural network ML model to automatically draw gates. Once prohibitively expensive with respect to time and effort, such a broad strategy is now readily implementable for analysis of SFC data. Finally, by integrating mandatory manual reproducibility of each result, we allay genuine concerns about trustworthiness of AI-written code.

Gabriel Salzman, MD, PhD: No relevant conflicts of interest to declare.