Transcriptomic profiling of T cell immune states before and after donor lymphocyte infusion in a patient with acute myeloid leukemia

Abstract Background: One of the biggest challenges in optimizing immunotherapy for acute myeloid leukemia (AML) is the inability to predict patient response to allogeneic targets after stem cell transplantation or donor lymphocyte infusion (DLI). To address this, we developed a machine-learning predictive platform that leverages transcriptomic data obtained directly from functionally active immune cells. Our approach uses a novel droplet-based microfluidics system that enables functional profiling of immune cells at the single-cell level, followed by transcriptomic profiling of these immune cells by RNA sequencing. This study integrates high-throughput transcriptomic data with machine learning classification to predict both the functional capacity of healthy donor T cells and the patient's response post-treatment. Methods: Peripheral blood samples were collected from a fully donor chimeric AML patient before, and one month after donor lymphocyte infusion (DLI), as well as from their healthy HLA-matched sibling donor. CD8⁺ T cells were isolated from each sample and co-encapsulated with CD33⁺ Kasumi-1 AML target cells and analyzed using our function-based, droplet microfluidic Fluorescent Activated Droplet Sorting (FADS) system to distinguish T cells based on their cytotoxic activity, isolating subpopulations of cancer killers and non-killers. In addition, CD8⁺ T cells from 28 healthy donors treated with CD3/CD19 bispecific antibody were similarly co-encapsulated with target cells in microfluidic droplets. The functionally defined T cell subsets were subjected to RNA sequencing to generate immune transcriptomic profiles for downstream analysis. Results: Principal component analysis revealed clear separation between the 28 healthy donors showing high cytotoxic activation (“effector positive”) and those with lower functional profiles (“effector negative”) based on the first two principal components (PC1: 30.5%, PC2: 15.7%). K-means clustering (k=2) further confirmed distinct transcriptional clusters corresponding to functional phenotypes, defined by differential expression of cytolytic, cytokine, activation, metabolic, and synaptic genes. After applying the K-mean centroid model to the AML case study by projecting the clinically selected donor's T cell transcriptomic profile onto the same PCA space, we found that the donor falls in the “effector positive” cluster, predicting strong cytolytic potential. Following DLI, we found that the AML patient's T cells showed marked transcriptional changes. Compared to the transcriptomic profile before DLI, the patient had increased expression of cytolytic, pro-inflammatory cytokines, and chemokines after DLI. In addition, RNA-seq data also showed transcriptomic shifts in synapse formation, metabolic activities, and T cell fitness. These transcriptomic shifts closely aligned with the “effector positive” classification predicted for the donor, supporting the utility of our AI model for predicting donor performance and patient response. Conclusion:This study presents a novel, function-to-omics platform that combines immune cell cytotoxicity with transcriptomic profiling and machine learning to predict T cell performance before clinical use. In addition, it provides high-resolution immune profiling of patient T cells before and after treatment, which may inform optimal donor selection to personalize immunotherapy strategies in AML and other hematological malignancies.