Predicting Cytotoxic T-cell Age from Multivariate Analysis of Static and Dynamic Biomarkers

Abstract

Adoptive T-cell transfer therapy relies upon in vitro expansion of autologous cytotoxic T cells that are capable of tumor recognition. The success of this cell-based therapy depends on the specificity and responsiveness of the T cell clones before transfer. During ex vivo expansion, CD8؉ T cells present signs of replicative senescence and loss of function. The transfer of nonresponsive senescent T cells is a major bottleneck for the success of adoptive T-cell transfer therapy. Quantitative methods for assessing cellular age and responsiveness will facilitate the development of appropriate cell expansion and selection protocols. Although several biomarkers of lymphocyte senescence have been identified, these proteins in isolation are not sufficient to determine the age-dependent responsiveness of T cells. We have developed a multivariate model capable of extracting combinations of markers that are the most informative to predict cellular age. To acquire signaling information with high temporal resolution, we designed a microfluidic chip enabling parallel lysis and fixation of stimulated cell samples on-chip. The acquisition of 25 static biomarkers and 48 dynamic signaling measurements at different days in culture, integrating single-cell and population based information, allowed the multivariate regression model to accurately predict CD8؉ T-cell age. From surface marker expression and early phosphorylation events following T-cell receptor stimulation, the model successfully predicts days in culture and number of population doublings with R 2 ‫؍‬ 0.91 and 0.98, respectively. Furthermore, we found that impairment of early signaling events following T cell receptor stimulation because of long term culture allows prediction of costimulatory molecules CD28 and CD27 expression levels and the number of population divisions in culture from a limited subset of signaling proteins. The multivariate analysis highlights the information content of both averaged biomarker values and heterogeneity metrics for prediction of cellular age within a T cell population.

Three-Dimensional In Vitro Tri-Culture Platform to Investigate Effects of Crosstalk Between Mesenchymal Stem Cells, Osteoblasts, and Adipocytes

“…32,33 To optimize the quality of PCA and PLS-DA models, several pruning procedures were performed to remove outlying observations (outside 95% confidence interval) and noninfluential variables (weight *0 in both components) and enable statistical significance testing of the model and the variables used to generate it. 32,34 The quality of each model was summarized by two nondimensional statistical parameters: (1) R 2 X (for PCA) or R 2 Y (for PLS-DA), which quantitatively measures the extent to which the model explains the variation in the data matrices and dictates a goodness of fit, and (2) Q 2 , which quantitatively measures the extent to which the variation of a future experimental data set may be predicted by the model (goodness of prediction). 34 Both of these parameters are analogous to regression statistics, with a value ranging from 0 (poor) to 1 (perfect) fit or predictive capability.…”

“…32,34 The quality of each model was summarized by two nondimensional statistical parameters: (1) R 2 X (for PCA) or R 2 Y (for PLS-DA), which quantitatively measures the extent to which the model explains the variation in the data matrices and dictates a goodness of fit, and (2) Q 2 , which quantitatively measures the extent to which the variation of a future experimental data set may be predicted by the model (goodness of prediction). 34 Both of these parameters are analogous to regression statistics, with a value ranging from 0 (poor) to 1 (perfect) fit or predictive capability. The appropriate number of PCs or latent variables was determined by cross-validation.…”

“…The appropriate number of PCs or latent variables was determined by cross-validation. 32,34 The results of this procedure were fed into a jack-knifing analysis 32 to calculate the standard errors of the regression coefficients (weights), which were then converted into 95% confidence intervals via the t-distribution to determine which X-variables (genes) of high weight have a statistically significant influence ( p < 0.05) on each class of observations (Y-variables) in each of the PLS-DA models generated (Figs. 5D and 6D, F).…”

Insights into Proteomic Immune Cell Signaling and Communication via Data-Driven Modeling

“…This is especially useful in clinical settings where it is necessary to rapidly and efficiently differentiate between different immune states or cell types, such as the examples we review here, including differentiation between different types of infections (Prilutsky et al 2011), or screening high quality cells for use in cell therapy (Rivet et al 2011). In these settings, distinguishing between biological states is of greater importance than knowledge of specific mechanisms governing these differences.…”

“…One major difficulty of the approach is loss of in vivo tumor specificity during clonal expansion of the T cells, and the transition of expanded cells to senescent states. A recent study (Rivet et al 2011) employed multivariate analysis to determine T cell senescence based on cell surface markers and intracellular protein signaling events in four healthy donors. A newly developed microfluidic device allowed for flow cytometric measurement of CD28, CD27, cell shape, and cell size in parallel with the dynamic phosphorylation of six proteins (CD3, CREB, ERK, LAT, Lck, and Zap70) downstream of T cell activation signaling at eight time points, from 0.5 to 7 min.…”