{
  "abstract": "Background Predictive modeling of T-cell function is needed to accelerate development of more effective immunotherapies. However, training a predictive model remains challenging due to limitations of current single-cell analysis methods. Traditional endpoint flow cytometry provides static snapshots but cannot track single-cell evolution over time. Standard single-cell OMICs tools do not capture the time dimension of single-cell behavior and thus are not ideal training sets for predicting single-cell behavior. Time-lapse flow cytometry—enabled by optical barcodes—addresses this gap by non-destructively tracking individual cells across multiple timepoints. This generates kinetic profiles of activation markers (e.g., CD25, CD69), exhaustion markers (PD-1, LAG-3, TIM-3) and connects them to functional assays like cytokine secretion and proliferation. Such longitudinal datasets (‘before, during, after’) are ideal for training supervised machine learning (ML) models to predict T-cell behavior.Methods We measured the kinetic phenotypes of T cells using Time-lapse flow cytometry with TCR-dependent, TCR-independent and antigen specific activations across 0h, 12h, 36h and 60h timepoints. At each time point we measured activation markers (CD69, CD25) and exhaustion markers (PD-1, LAG-3, TIM-3) with releasable antibodies. At the last time point we also measured cytokine release using intra-cellular cytokine staining (ICS). Single-cell data for all markers at all timepoints were extracted and used to train a supervised machine learning model (MATLAB and Python).ML models were trained with at least 100k cells, 5-fold cross-validated and 10% of cells were reserved for model testing.Results Single-cell marker expression trajectories (12h-60h) revealed heterogeneous marker dynamics of both activation markers (e.g., sustained CD25 vs. transient PD-1 expression) and exhaustion markers (eg. sustained co-expression of Tim-3 and Lag-3), These kinetic data were used to train neural networks for:Classify mode of activation based on dynamic marker expression.AUC: 0.82–0.92, reducing error rates by >30% compared to endpoint data.Predict single-cell cytokine secretion (IL-2, TNFα, IFNγ) 12h before measurement (AUC: 0.84–0.88; sensitivity: 75–82%, specificity: 74–87%).Kinetic data identified critical predictive windows (e.g., CD69 at 36h) and outperformed endpoint models in both tasks. For cytokine prediction, ML leveraged early-phase marker dynamics to forecast secretion before functional assays.Conclusions Longitudinal single-cell data (‘before, during, after’) significantly enhances ML prediction of T-cell function. Our models achieved high accuracy in classifying activation pathways and forecasting cytokine secretion, demonstrating that kinetic heterogeneity is a critical predictor of functional outcomes. This approach enables in silico screening of immunotherapies and accelerates the development of virtual cell models for precision immunotherapy.",
  "authors": [
    {
      "affiliations": [
        "LASE Innovation, Waltham, MA, USA"
      ],
      "name": "Sheldon Kwok"
    },
    {
      "affiliations": [
        "LASE Innovation, Waltham, MA, USA"
      ],
      "name": "Yulia Shulga"
    },
    {
      "affiliations": [
        "LASE Innovation, Waltham, MA, USA"
      ],
      "name": "Sarah Forward"
    },
    {
      "affiliations": [
        "LASE Innovation, Waltham, MA, USA"
      ],
      "name": "Emane Rose Assita"
    },
    {
      "affiliations": [
        "Talon Biomarkers, Whippany, NJ, USA"
      ],
      "name": "Alexis Hernandez"
    },
    {
      "affiliations": [
        "Talon Biomarkers, Whippany, NJ, USA"
      ],
      "name": "Pratip Chattopadhyay"
    },
    {
      "affiliations": [
        "LASE Innovation, Waltham, MA, USA"
      ],
      "name": "Trevor Brown"
    }
  ],
  "title": "1100 Predicting T cell activation and exhaustion using a supervised machine learning model trained on novel live T cell dynamic phenotype and function single-cell data",
  "uid": "c2eb12ef-e2f3-5108-929e-9ee48c704937"
}