Pipeline detailing the workflow of maRQup

a, A schematic of the in-vivo experimental setup used to evaluate the efficacy of CAR-T cell immunotherapy experiments. NALM-6 leukemia cells expressing luciferase were injected via the tail vein of mice followed by the adoptive transfer of CAR-T cells (IV) and the evaluation of tumor growth using a Xenogen IVIS Lumina, generating the output raw images for the maRQup pipeline. b, Processing steps to quantify tumor radiance at each timepoint for all mice in an experiment. IVIS output images are grouped by experimental condition (columns) and across multiple timepoints (rows) as indicated. Each output image contains a Brightfield (top) and Luminescent (bottom) image. Individual mice are then separated from the group image, and all mice images are aligned via image processing. Tumor burdens in different anatomical regions are quantified, and the tumor burden over time is calculated. c, Modeling, parametrization, and analysis of tumor behavior. Different phases of the tumor response dynamics — growth, decay, and relapse — are identified for each mouse. Each phase is modeled using differential equations, allowing for the rates of tumor growth in each phase to be compared across mice and experimental conditions.

Impact of CAR-T cell constructs and doses on in-vivo dynamics.

a, Schematic showing the extracellular and intracellular components of T cells harboring the CD19.4-1BB (left, blue) and the CD19.CD28 (right, orange) CAR constructs. b, Overall tumor response curves for all mice treated with CD19.4-1BB CAR-T cells (top) and CD19.CD28 CAR-T cells (bottom). c, The percentages of mice treated with CD19.4-1BB (blue) or CD19.CD28 (orange) CAR-T cells that exhibit an initial tumor growth (left) vs. those that relapse (right). d, Tree diagrams showing tumor behavior in mice treated with increasing doses of CD19.4-1BB CAR-T cells. e, The percentages of mice demonstrating an initial growth (top, green) as compared to a decay (bottom, blue) of NALM6 leukemia cells presented as a function of the dose of CD19.4-1BB CAR-T cells (<1, 1, 2, or >2 million). f, The percentages of leukemia-bearing mice that exhibit a final phase of relapse (top, red) as compared to decay (bottom, blue) presented as a function of the dose of CD19.4-1BB CAR-T cells.

Parametrization of tumor response to CAR-T cell therapy.

a, Tumor dynamics for all 1,060 mice in the dataset. Model calculations are shown as curves for all mice. A representative example is highlighted, with the experimental data shown as points. The growth, decay, and relapse phases are labeled in green, blue, and red, respectively. b, Residuals in log10-scale are shown for each timepoint in each phase. The percent of modeled points within one order of magnitude in either the positive or negative direction of the experimental points (dotted grey lines; 10−1 to 101) are labeled for each phase. c, Diagram displaying the parameters used to model the tumor dynamics. d, Comparison of the growth rate of NALM6 leukemia in the growth phase (kGrowth) and in the relapse phase (kRelapse) in mice that exhibited tumor growth, decay, and relapse. e, Tumor growth rates can be divided into slow growing (gold) and fast growing populations (purple). The vertical dotted line at 0.85 days−1 represents the division of the two populations. f, Tree diagram for the slow growing (left) and fast growing tumor populations (right). g, (left) Percentage of mice harboring slow or fast growing tumor populations as a function of their last phase being growth (NR), decay (CR/PR), or relapse (PD). (right) The percentages of mice that exhibit tumor decay and then relapse under conditions where the tumors exhibit slow (gold) or fast (purple) initial growth. h, Kernel density estimate plots of the distribution of the growth rate of tumors in the growth phase (kGrowth) (left), the rate of tumor killing by CAR-T cells (k − kGrowth) (middle), and the growth rate of tumors in the relapse phase (kRelapse) (right). Mice were classified into two groups of Effector to Target Ratios (E:T) depending on if the mice were treated with more (red) or fewer/equal (blue) CAR-T cells than tumor cells. The division of the slow (gold) and fast (purple) growing tumor populations, as determined from e is shown. i, Violin plots showing the growth rate of tumors during the relapse phase (kRelapse) as stratified by the initial behavior of the tumor: initial decay (kGrowth = 0), slow initial growth (0 < kGrowth ≤ 0.85), or rapid initial growth (kGrowth > 0.85)). All rates have units of days−1.

Characterizing spatial differences in tumor progression dynamics.

a, Zeros of four principal components measured across row-averaged and column-averaged mouse image data were used to demarcate nine distinct spatial regions. b, Images (top) and quantified radiance dynamics (bottom) both globally (all) and per region are shown for a representative mouse. c, Characterization of tumor growth dynamics in all mice engrafted with NALM6 leukemia (left) and the percentages of mice where tumor dynamics in a given region diverge from the dynamics evaluated from the overall total body radiance (right). The ‘bmR’ region is defined as the region containing the bone marrow of the mouse’s right leg; the ‘bmL’ region is defined as the region containing the bone marrow of the mouse’s left leg. d, Scatter plots of mice with a given overall tumor behavior as compared to the classified tumor behavior within each specific region. Residuals from the identity line (dashed diagonal line) with magnitudes outside the 95% confidence intervals (grey bands) indicate a significant divergence of a region specific growth pattern as compared to the overall growth pattern. e, Correlation coefficients for all pairwise combinations of overall tumor growth and region specific tumor growth.

maRQup classification of tumor response dynamics.

a, Representative example of the tumor dynamics showing growth (green), decay (blue), and relapse (red) phases for a single mouse. Tumor radiance images are shown below each corresponding timepoint. b, Equations used to determine the timepoint for the end of each phase. c, Experimental tumor dynamics for all 1,060 mice in the dataset showing the three different phases. d, Tree diagram detailing the tumor behavior for all 1,060 mice. Each row corresponds to a different phase (i.e., growth, decay, or relapse), and each node (circle) indicates the number of mice with tumor in each phase. Each endpoint (square) indicates the number of mice with the respective tumor dynamics. The thickness of each line denotes the overall percentage of mice that pass through that path, while the darkness of each line denotes the relative percentage of mice that move from the previous node to the next node or endpoint. e, Classification of tumor dynamics into three categories: No Response (NR); Complete/Partial Response (CR/PR); and Progressive Disease (PD). Tumor growth curves for all mice in the respective categories are shown. A representative tumor curve is shown on each plot with the corresponding phases identified.

Summary of different experimental conditions in the dataset.

a, Different tumors in the dataset. b, Initial dose of all tumors injected into the mice. c, Initial dose of NALM6 tumors injected into the mice. d, Different binding domains for CAR-T cells. e, Different costimulatory domains for CAR-T cells. f, Different CAR-T cell constructs used to target all tumors in the dataset. g, Different CAR-T cell constructs used to target NALM6 tumors only. h, Different CAR-T cell constructs used to target non-NALM6 tumors. i, Doses of CAR-T cells (not including mock/control conditions). j, Doses of the CD19.4-1BB CAR-T cells. k, Doses of the CD19.CD28 CAR-T cells. l, Initial dose of NALM6 tumors that were treated with CD19.4-1BB CAR-T cells. m, Initial dose of NALM6 tumors that were treated with CD19.CD28 CAR-T cells.

Bayesian Priors methods.

a, Distributions of the rate of tumor growth in the growth phase (green), the rate of tumor decay in the decay phase (blue), and the rate of tumor growth in the relapse phase (red) prior to Bayesian Priors correction. b, Total number of data points per phase as a function of the tumor growth/decay rate before Bayesian Priors correction. Outlier data points are shown in red in the dotted box for each phase. c, Distribution of rates before Bayesian Priors correction for each phase with outliers identified in the dotted box for each phase. d, Distributions of the rate of tumor growth in the growth phase (green), the rate of tumor decay in the decay phase (blue), and the rate of tumor growth in the relapse phase (red) after Bayesian Priors correction. e, Total number of data points per phase as a function of the tumor growth rate in the growth and relapse phases after Bayesian Priors correction. Outlier data points are shown in red in the dotted box for each phase. The decay phase is not shown since Bayesian Priors was not applied to this phase. f, Distribution of rates after Bayesian Priors correction for the growth and relapse phases with remaining outliers identified in the dotted box for each phase.

Image processing methods.

a, Each composite IVIS image consists of a radiance image (left), a binary mouse pixel image (middle), and a brightfield image (right). b, Tail cropping method. Dotted green line denotes the row where the start of the tail is identified via a Savitzky-Golay filter (SG, orange curve) and KneeLocator algorithm. c, Representative mouse image following tail cropping. d, Representative image of a slanted mouse. The ordinary least squares regression line of the (x,y) pixel coordinates is shown in blue. e, Distribution of the angle of slant for all 7,633 images in the dataset. Each red line represents one image. Images were removed from analysis if the angle of slant was greater than 6 degrees. f, Representative example of one of eight images that were removed due to incorrect image cropping (e.g., multiple mice in one image). g, Composite images (i.e., standard deviation of the merged mouse pixel boolean images, average of the merged radiance images, average of the merged mouse pixel images, and average of the merged brightfield images) after all 7,534 images were merged following image processing.

Spatial partitioning.

a, (left) Cumulative variance of principal components for the principal component analysis performed in the row-dimension (solid line) or column-dimension. The principal components that cumulatively explain greater than 90% of the variance are shown in different colors. (middle) Values across all rows of the first four principal components from the row-wise principal component analysis. Principal component 2 (orange) and principal component 3 (green) are shown in bold and were used to identify anatomical regions of interest. (right) Values across all columns of the first eight principal components from the column-wise principal component analysis. Principal component 4 (red) and principal component 5 (purple) are shown in bold and were used to identify anatomical regions of interest. b, Number of mice with region-specific behavior compared to overall tumor behavior for the nine previously identified anatomical regions. Residuals from the identity line (dashed diagonal line) with magnitudes outside of 95% confidence intervals (grey bands) indicate a significant divergence of a region specific growth pattern from the overall growth pattern. c, Pairwise comparisons of region-specific tumor behavior. d, Heatmaps showing the number of mice that have different tumor behavior across each region shown for each of the five classified tumor behaviors (i.e., growth only (G; green); decay only (D; blue); growth and decay (G,D; teal); decay and relapse (D,R; purple); and growth, decay, and relapse (G,D,R; grey)).