Abstract 6267: Training a machine learning model to predict T cell cytokine secretion function using a novel single-cell flow cytometry method that measures kinetic phenotypes

Advances in high-dimensional single-cell analysis have transformed our understanding of cellular heterogeneity and its role in complex pathologies. Most datasets are acquired through single-cell RNA sequencing, which provides expression of thousands of genes on each cell at a given timepoint. The typical workflow involves identifying potential new cell types via clustering, generation of hypotheses, and followed by post-hoc experiments to test these hypotheses. Technical bottlenecks include the sparse and noisy nature of sequencing data, and the inability to link observations directly to function due to the destructive nature of sequencing. Supervised machine learning (ML) is an increasingly useful approach to generate powerful predictive models of complex processes. This approach requires both high quality and high quantity of training data. An ideal method to acquire this data would be able to follow individual cells over time while measuring changes in phenotype and function. Here, we describe a novel single-cell method using non-destructive optical barcodes that enables acquisition of dynamic high-dimensional data at scale. We apply this method to generate unprecedented predictive models of single-cell function using a human T-cell activation model. Our barcoding approach relies on laser particles (Kwok et al., Nat. Biomed. Eng. 2024) to track individual cells over time. T-cells are barcoded, activated, and measured repeatedly over time using a flow cytometer to capture kinetic changes in expression of protein biomarkers (“kinetic phenotyping”). We link these kinetic phenotypes to cytokine secretion function through a stimulation step and cytokine secretion assay. The single-cell phenotype-to-function datasets are then used to train ML models to predict cytokine secretion prior to stimulation. We observed increased expression of CD25, PD-1 and CD69 in T cells post-activation from 2h to 48h, as expected. However, distinct kinetic patterns were observed. For example, while CD25 increased largely monotonically, there was significant transient expression of PD-1. These observations were consistent across multiple donors. We employed supervised ML (Random Forrest) to train a classification model to predict cytokine secretion from single-cell kinetic phenotypes. The model was validated using 5-fold cross-validation and achieved ROC-AUC values of 0.74 to 0.89 for predicting TNFa, IFNg and IL2 secretion. For each cytokine, we identified the most predictive biomarkers (and timepoint). Finally, we used our classification model to describe T-cell heterogeneity by probability of cytokine secretion and validated our method using conventional analyses. Our approach is expected to be of great utility for predicting single-cell function in complex systems, from immunotherapy development to cancer pathogenesis. Citation Format: Sheldon J.J. Kwok, Yulia Shulga, Emane Rose Assita, Sarah Forward, Trevor Brown, Pratip Chattopadhyay. Training a machine learning model to predict T cell cytokine secretion function using a novel single-cell flow cytometry method that measures kinetic phenotypes [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 6267.