Apparatus and method for a knowledge processing system that applies inference techniques to cell-based analyses to predict clinical outcomes - Patents.com

Abstract

In some implementations, a machine learning model (e.g., a knowledge processing system) can apply inference techniques (e.g., using a genetic algorithm) to the cell-related knowledge representation to predict a clinical outcome (e.g., for a patient). In some implementations, a method includes generating a cell population. Gene expression data for the cell population can be received. The gene expression data can be analyzed to identify a set of gene signatures. The set of gene signatures can be provided as input to a machine learning classifier. A predicted clinical outcome can be generated.

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to and the benefit of U.S. Provisional Patent Application No. 63 / 319,547, filed March 14, 2022, and entitled "APPARATUS AND METHODS FOR A KNOWLEDGE PROCESSING SYSTEM THAT APPLIES A REASONING TECHNIQUE FOR CELL-BASED ANALYSIS TO PREDICT A CLINICAL OUTCOME," and U.S. Provisional Patent Application No. 63 / 405,128, filed September 9, 2022, and entitled "APPARATUS AND METHODS FOR A KNOWLEDGE PROCESSING SYSTEM THAT APPLIES A REASONING TECHNIQUE FOR CELL-BASED ANALYSIS TO PREDICT A CLINICAL OUTCOME," each of which is incorporated by reference in its entirety herein.

[0002] The present disclosure relates to apparatus and methods for machine learning models (e.g., knowledge processing systems) that can apply inference techniques to cell-based analyses and predict clinical outcomes (e.g., for patients). [Background technology]

[0003] Chimeric antigen receptor T cells (CAR-T) have shown considerable activity in the treatment of B-cell malignancies. However, T cells may pose unique challenges for therapeutic development. These so-called "living drugs" can proliferate, differentiate, actively transport between tissues, and / or engage in bidirectional communication with the patient's immune system while performing their functions. The resulting pharmacology may differ from that of small molecules or biologics, for example, in terms of the relationship between administered dose and exposure.

[0004] The pharmacokinetics ("cell kinetics") of circulating CAR-T can be characterized by three distinct phases: initial proliferation, followed by rapid contraction, then slow and prolonged decay. The degree of cell proliferation (Cmax) and prolonged exposure (AUC) can vary between patients (e.g., by about three orders of magnitude) and can predict efficacy (reduction in tumor size) and / or toxicity. However, the product- and host-specific factors that mediate this pharmacology remain suboptimal. Thus, the facts and / or relationships (i.e., knowledge representations) related to CAR-T may not be optimally known, understood, and / or exploited (e.g., by knowledge processing systems). Several known empirical nonlinear mixed-effect models for quantifying the pharmacokinetics of Kymriah® (Tisagenlecleul, CTL019) are provided as part of the Biologics License Application (BLA). These known formulations have been applied to other CAR-T therapies in various indications and have also been adopted by the FDA for benchmarking. While empirical equations are potentially applicable to quantifying clinical data, they do not account for the underlying biology and are therefore of limited value in simulating the effects of alternative CAR-T designs or treatment regimens. In that context, mathematical models capable of quantitatively describing clinical data while being based on sound biological mechanisms would be desirable in the development of novel CAR-T products, similar to the use of mechanism-based pharmacokinetic-pharmacodynamic (PKPD) modeling (quantitative systems pharmacology) during the exploratory phase of drug development.

[0005] Some known models are derived from a predator-prey equation in which CAR-T cells are endowed with unlimited proliferation capacity. However, such known models do not address the clinical observation that CAR-T proliferation capacity varies dramatically between patients, nor the mechanisms underlying such clinical variability.

[0006] It would be desirable to eliminate or mitigate one or more of the above deficiencies. Summary of the Invention

[0007] In some implementations, a machine learning model (e.g., a knowledge processing system) can apply inference techniques (e.g., using a genetic algorithm) to the knowledge representation related to the cells to predict clinical outcomes (e.g., for a patient). In some implementations, the method includes receiving gene expression data for a plurality of cells. A differential gene expression analysis can be performed based on the gene expression data to generate differential gene expression data. Gene signature enrichment for each sample for a plurality of identified biological pathways is estimated based on the differential gene expression data. Gene signatures having a certain differential enrichment pattern can be filtered based on a predetermined threshold between a group of responders and a group of non-responders to the treatment. The filtered gene signatures can be grouped into a plurality of groups based on pairwise correlations between gene signature enrichment scores. A set of gene signatures can be determined from the plurality of sets of gene signatures by randomly selecting a gene signature from each group from the plurality of groups by performing a predetermined number of iterations to determine the plurality of sets of gene signatures. The multiple sets of gene signatures may be provided as input to a feature selection algorithm to define a feature vector that includes a reduced set of gene signatures identified from the multiple sets of gene signatures. The feature vector may be used to train a machine learning classifier. [Brief description of the drawings]

[0008] [Figure 1-1] FIG. 1 shows an antigen toggle switch model of T cell regulation that quantitatively describes the PKPD behavior for the response of patient populations that are complete responders (CR), partial responders (PR), and non-responders (NR) to Kymriah (registered trademark) in chronic lymphocytic leukemia (CLL), according to one embodiment. [Figure 1-2] Same as above.

[0009] [Diagram 2]FIG. 2 shows single sample gene set enrichment analysis (ssGSEA) estimates of signaling pathway activity and cell population enrichment in CAR-T, broken down by response, according to one embodiment.

[0010] [Figure 3-1] FIG. 3 shows single-cell RNA sequencing of twelve pre-infusion CAR-T products stratified by response, according to one embodiment. [Figure 3-2] Same as above.

[0011] [Figure 4] FIG. 4 shows Kymriah® response prediction using a ssGSEA-based classifier, according to one embodiment.

[0012] [Figure 5-1] FIG. 5 shows the clinical variability of dose, tumor burden, and CR / PR / NR pharmacological archetypes with respect to population distribution of exposure to Kymriah®, predicting clinical covariates of response to Yescarta®, according to one embodiment. [Figure 5-2] Same as above.

[0013] [Figure 6-1] FIG. 6 shows the training, analysis, and testing of a model relating to the dose response of Abecma®, according to one embodiment. [Figure 6-2] Same as above.

[0014] [Figure 7A] FIG. 7A shows a flowchart of a response classifier workflow according to one embodiment.

[0015] [Figure 7B] FIG. 7B shows an illustration of a response classifier workflow showing example data at each step in the pipeline, using a process similar to that discussed with respect to FIG. 7A, according to one embodiment.

[0016] [Figure 7C] FIG. 7C shows a flowchart of a response classifier workflow according to one embodiment.

[0017] [Figure 7D] FIG. 7D shows an illustration of a response classifier workflow showing example data at each step in the pipeline, using a process similar to that discussed with respect to FIG. 7C, according to one embodiment.

[0018] [Figure 8] FIG. 8 illustrates a PKPD model workflow according to one embodiment.

[0019] [Figure 9] FIG. 9 shows a goodness of fit plot of the Kymriah® model fitting of FIG. 1, according to one embodiment.

[0020] [Figure 10] FIG. 10 shows a Local Parameter Sensitivity Analysis of the CR / PR / NR population, according to one embodiment.

[0021] [Figure 11] FIG. 11 shows a volcano plot of differentially expressed genes between CR versus NR groups, according to one embodiment.

[0022] [Figure 12] FIG. 12 shows a display of select gene sets differentially enriched between CR versus NR groups, according to one embodiment.

[0023] [Figure 13]FIG. 13 shows (A) the average receiver operating characteristic (ROC) curves of 2500 trained models for the Bai 2022 and Fraietta data, and (B) the classification of PR and PRTD samples (not used in training the model) mixed between 0 and 1, according to one embodiment.

[0024] [Figure 14] Figure 14 shows PKPD response dependent on initial tumor burden and CAR-T dose in CR, according to one embodiment.

[0025] [Figure 15] FIG. 15 shows a goodness of fit plot for Abecma®, according to one embodiment.

[0026] [Figure 16A] FIG. 16A shows a flowchart of a method for training a machine learning classifier, according to one embodiment.

[0027] [Figure 16B] FIG. 16B shows a flowchart of a method for training a machine learning classifier, according to one embodiment.

[0028] [Figure 17] FIG. 17 shows a flowchart of a method for predicting a patient's clinical outcome in response to a cell therapy treatment, according to one embodiment.

[0029] [Figure 18] FIG. 18 shows a flowchart of a method for generating a predicted clinical outcome, according to one embodiment.

[0030] [Figure 19] FIG. 19 shows a flowchart of a method for predicting a patient's clinical outcome in response to a cell therapy, according to one embodiment.

[0031] [Figure 20]FIG. 20 shows a flowchart of a method for administering cell therapy to a patient, according to one embodiment.

[0032] [Figure 21] FIG. 21 shows a flowchart of a method for generating a cell therapy product, according to one embodiment.

[0033] [Figure 22] FIG. 22 shows a flow chart of a method for determining a patient-specific dosage of a cell therapy to be administered, according to one embodiment.

[0034] [Figure 23] FIG. 23 illustrates a block diagram of a system for implementing one or more of the concepts / methods discussed herein, according to one embodiment.

[0035] [Figure 24A] 24A-24B show a list of model parameters, units, and lower and upper bounds used in the particle swarm optimization algorithm, according to one embodiment. [Figure 24B] Same as above.

[0036] [Diagram 25] FIG. 25 shows 12 example parameter sets for CR (complete responder) for the parameters of FIGS. 24A-24B, according to one embodiment.

[0037] [Figure 26] FIG. 26 shows 12 example parameter sets for PR (partial responders) for the parameters of FIGS. 24A-24B, according to one embodiment.

[0038] [Figure 27] FIG. 27 shows 12 example parameter sets for NR (non-responder) versus the parameters of FIGS. 24A-24B, according to one embodiment.

[0039] [Figure 28]FIG. 28 shows a representation of the structure of the original model and possible variations, according to one embodiment.

[0040] [Figure 29-1] FIG. 29 shows the fitting accuracy of the model for five structural variants. [Figure 29-2] Same as above.

[0041] [Diagram 30] Figure 30 shows model fitting to preclinical CD19-CAR-T data, according to one embodiment.

[0042] [Diagram 31] FIG. 31 shows model fitting to preclinical BCMA-CAR-T data, according to one embodiment.

[0043] [Diagram 32] FIG. 32 shows the simulated frequencies of memory (TM), effector (TE), and exhausted (TX) CAR-T cells for the CR, PR, and NR patient groups shown in FIG. 1, according to one embodiment.

[0044] [Diagram 33] FIG. 33 shows pairwise Pearson correlation coefficients between Thymic Atlas cell population gene signatures calculated using ssGSEA scores from the RNAseq data of Fraietta et al., according to one embodiment.

[0045] [Figure 34-1] FIG. 34 shows single-cell RNA sequencing of twelve pre-infusion CAR-T products stratified by response, according to one embodiment. [Figure 34-2] Same as above.

[0046] [Diagram 35]FIG. 35 shows a comparison of the pharmacokinetics of Kymriah® in B-ALL, a model described herein of Kymriah® in CLL, and Yescarta® in LCBCL, according to one embodiment.

[0047] [Diagram 36] FIG. 36 shows a simulation of cmax / tumor burden versus tumor response to Abecma®, according to one embodiment.

[0048] [Figure 37-1] Figure 37 shows single-cell RNA sequencing of two pre-infusion CAR-T products separated by response, according to one embodiment. [Figure 37-2] Same as above.

[0049] [Figure 38] FIG. 38 shows Kymriah® response prediction in CLL using a ssGSEA-based classifier, according to one embodiment.

[0050] [Figure 39] FIG. 39 shows (A) cluster size and (B) average intra-cluster correlation distributions based on a variable number of distinct clusters calculated for 864 gene signatures differentially enriched between CR and NR groups, according to one embodiment.

[0051] [Figure 40-1] FIG. 40 illustrates cell-intrinsic defects associated with non-sustained responses, according to one embodiment. [Figure 40-2] Same as above.

[0052] [Figure 41-1] FIG. 41 shows CD19-CART response prediction, according to one embodiment. [Figure 41-2] Same as above.

[0053] [Diagram 42] Figure 42 shows model fitting results based on the hypothesis that the only distinguishing feature between the CR, PR, and NR populations is the fraction of memory and exhausted T cells in the CAR-T infusion product, according to one embodiment.

[0054] [Diagram 43] FIG. 43 shows simulated pharmacokinetic and tumor dynamic responses to increasing cell doses of pure memory cell populations from the CR, PR, and NR population models, according to one embodiment.

[0055] [Figure 44-1] Figure 44 shows T cell phenotyping of CAR-T infusion products, according to one embodiment. [Figure 44-2] Same as above.

[0056] [Diagram 45] FIG. 45 shows the performance of the transcriptome classifier compared to the null model and the random pathway model, according to one embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0057] In some implementations, cell therapy, as used herein, may include any suitable cell therapy, such as adoptive cell therapy and cellular immunotherapy. Cellular immunotherapy is a form of treatment that uses cells of our immune system to eliminate cancer cells or other unwanted cells. Some of these approaches involve directly isolating and simply expanding the numbers of our own immune cells, while other approaches involve genetically engineering our immune cells (via gene therapy) to enhance their ability to fight cancer. Our immune system can recognize and eliminate infected or damaged cells, as well as cells that have become cancerous. In the case of cancer, immune cells known as killer T cells can be powerful against cancer due to their ability to bind to markers known as antigens on the surface of cancer cells. Cellular immunotherapy can be deployed in a variety of ways, taking advantage of this natural ability. Exemplary types of cell therapy include, for example, chimeric antigen receptor T cell therapy (CAR-T cell therapy), tumor infiltrating lymphocyte (TIL) therapy, engineered T cell receptor (TCR) therapy, and natural killer (NK) cell therapy. The devices and methods of the present disclosure are specifically illustrated with respect to CAR-T cell therapy. However, it will be appreciated that the present disclosure may be readily applied to other cell therapies.

[0058] As used herein, clinical outcomes may include, for example, complete response, partial response, non-response. In some implementations, the definition of response categories for clinical outcomes is disease indication specific, for example, referring to tumor burden measured at a defined time point after administration of therapy.

[0059] Cmax (maximum concentration) is the maximum observed abundance of CAR-T cells in the circulation after dosing, which may occur 1 week to 1 month after dosing. CAR-T cells can be quantified, for example, using bioluminescence imaging (BLI), PCR for the CAR transgene (counts per ug of genomic DNA), or by flow cytometry of CAR expression (% of circulating T cells). AUC is the area under the curve calculated from pharmacokinetics over time.

[0060] Exemplary forms of gene expression data that may be obtained and used in accordance with the present disclosure include, for example, RNAseq, RNA sequencing data, scRNAseq, single cell RNA sequencing data, GSEA, gene set enrichment analysis, and ssGSEA, single sample gene set enrichment analysis.

[0061] The machine learning classifier may be, for example, a classifier used to automatically classify data and may include, for example, logistic regression, multinomial logistic regression, decision trees, perceptrons, support vector machines, K-nearest neighbors, naive Bayes, random forests, etc.

[0062] As used herein, a pathway or biological pathway can be a set of genes known to be involved in a defined intracellular biochemical mechanism.

[0063] A pharmacokinetic / pharmacodynamic model, in some implementations, can integrate components of pharmacokinetic and pharmacodynamic models into a set of mathematical equations to describe the dynamics of a drug and / or physiological components in response to an administered dose of the drug.

[0064] Mathematical models of T cell-tumor interaction can be adapted to describe various aspects of CAR-T pharmacology, namely antigen binding, cell-cell signaling cytokine release, tissue distribution, and / or competition with host T cells for immune system reconstitution.However, some known models do not adequately define what limits cell proliferation or what accounts for the wide variation in exposure and tumor response observed between patients.In addition, these known models are slow and inefficient due to the use of large and non-optimal data sets.

[0065] Insights can be gleaned by examining the dynamics of T cells in response to viral infection. When viral antigens are encountered, antigen-specific T cells clonally expand and differentiate into cytotoxic effectors that clear infected cells. After clearance of the pathogen, effector cells undergo a rapid contraction phase, during which a small proportion survive to form long-term memory T cells capable of self-renewal, restoring the response. However, if the infection does not resolve, chronic antigen stimulation leads to T cell exhaustion, where residual T cells lose the ability to produce cytokines, kill target cells, or proliferate in response to antigen.

[0066] In one embodiment, a similar process underlies the pharmacology of CAR-T. Some concepts discussed herein involve the use of a mathematical model of T cell differentiation control, where an antigen-driven toggle switch regulates cell fate transitions between memory, effector, and exhausted T cells. The model can quantitatively describe pharmacokinetic and tumor dynamics data from multiple clinical trials, deconvoluting cell-intrinsic and host-intrinsic sources of interpatient variability. These mathematical results can be confirmed via bulk and single-cell RNAseq profiling of CAR-T products, finding that pre-infusion transcriptome predicts response. The model novel predicts clinical variance in exposure, response covariates, and underlying biological mechanisms.

[0067] FIG. 1 shows an antigen toggle switch model of T cell regulation that quantitatively describes the PKPD behavior of patient population responses that are complete responders (CR), partial responders (PR), and non-responders (NR) to Kymriah® in chronic lymphocytic leukemia (CLL), according to one embodiment. A Three populations of T cells are identified: T memory cells (T M ), T effector cells (T E1 and T E2 ), and exhausted T cells (T X Depiction of a model structure including a B cell tumor (B), and a B cell tumor (B). The tumor cells express B cell antigens (B), which stimulate the proliferation and differentiation of T cells and inhibit the formation of T memory cells.A ). B Fit of model structure to known PKPD profiles separated by response categories (CR / PR / NR) using particle swarm optimization (PSO). The model fits (curves are averages of 12 parameter sets, dark shaded areas are middle 90%) are consistent with both CAR-T and B-cell tumor dynamics over time (dots are average data, light shaded areas are range of data) for each of the three archetypal populations. C PCA plot of the logarithm of the best fitting parameters colored by population. Principal component 1 (PC-1) captures 35.3% of the variability and principal component 2 (PC-2) captures 21.7% of the variability. D Coefficients of sorted PC-1 are TK50 (bars at 103) and d E2 , μ M And d M (101 bars) suggest that this is the largest source of variation between the CR and NR populations. These parameters correspond to cytotoxic potential, effector cell mortality, memory cell proliferation and mortality, respectively.

[0068] FIG. 2 shows single sample gene set enrichment analysis (ssGSEA) estimates of signaling pathway activity and cell population enrichment in CAR-Ts, split by response, according to one embodiment. A, C-F ssGSEA reveals population differences in cell populations and signaling pathways for selected cell signatures and signaling pathways (figure titles). Differences between populations were assessed using unequal variance t-tests (p-values ​​shown). B The percentage of day 60 T cell populations that are non-exhausted was calculated using the 12 best fitting parameter sets for each population and model simulation. The median day 60 non-exhausted T cell population (across the 12 parameter sets) is nearly 100% for both CR and PR populations, while the median is about 50% for NR populations. Differences between populations were assessed using unequal variance t-tests (p-values ​​shown).

[0069] FIG. 3 shows single-cell RNA sequencing of twelve pre-infusion CAR-T products stratified by response, according to one embodiment. A Uniform Manifold Approximation and Projection (UMAP) projections of cells annotated by response category: CR; complete response, RL; relapse, NR; non-response. B UMAP projections of cells annotated as exhausted using ProjectTILs (see, e.g., Andreatta, M. et al. Interpretation of T cell states from single-cell transcriptomics data using reference atlases. Nat Commun 12, 2965 (2021), the contents of which are incorporated herein by reference in their entirety). C UMAP projection of cells annotated as high (above the mean) or low (below the mean) CAR-T cell dysfunction signature from Good et al. (See, e.g., Good, CR et al. An NK-like CAR T cell transition in CAR T cell dysfunction. Cell (2021) doi:10.1016 / j.cell.2021.11.016, the contents of which are incorporated herein by reference in their entirety). D Cell type frequencies derived using ProjectTILs for CR, RL, and NR groups. Only cell types with at least 5% representation in the samples are shown. Data are shown as mean +- standard error, with individual sample frequencies overlaid as dots. Significance is indicated for significant comparisons between CR and NR according to Wilcoxon rank sum test (p≦0.05). E GSEA by cell type for selected pathways that show significant enrichment in at least one cell type (p≦0.05). A positive normalized enrichment score (NES) indicates higher enrichment in CR / non-exhausted cells. em is effector memory T cell, Th1 is type 1 helper T cell, T ex are exhausted T cells.

[0070] FIG. 4 shows that an ssGSEA-based classifier can be used to accurately predict Kymriah® response, according to one embodiment. A, C Trained classifier predictions on Fraietta's data (see, e.g., Fraietta, JA et al. Determinants of response and resistance to CD19 chimeric antigen receptor (CAR) T cell therapy of chronic lymphocytic leukemia. Nat Med 24, 563-571 (2018), the contents of which are incorporated by reference in their entirety) (A) and Bai 2022 data by pseudo-bulk analysis (see, e.g., Bai, Z. et al. Single-cell antigen-specific landscape of CAR T infusion product identifies determinants of CD19-positive relapse in patients with ALL. Sci Adv 8, (2022), the contents of which are incorporated by reference in their entirety) (C) from 2500 replicates in the test split of the data. The predicted probability of CR (p(CR)) is plotted as the mean ± standard error across the 2500 models. B, D Distribution of model accuracy for models trained on Fraietta data (B) and Bai 2022 data by pseudobulk analysis (D) compared to two test cases: a null model predicting response or non-response according to the proportion of each label in the dataset, and a random model choosing random pathways to use as features rather than our derived set of 30 pathways. Significance is p ≤ 10 -8 Comparisons using the Wilcoxon rank sum test are shown.

[0071] FIG. 5 shows that clinical variation in dose, tumor burden, and CR / PR / NR pharmacological archetypes explain population variance in exposure to Kymriah® and predict clinical covariates of response to Yescarta®, according to one embodiment. A Shaded areas indicate clinical variation in exposure to Kymriah®, with median model simulations overlaid for CR, PR, and NR populations. B AUC distribution of CAR-T. The first box plot labeled "Kymriah" shows the AUC distribution obtained from 1000 simulations of the clinical PK model (each black dot corresponds to a percentile of the AUC distribution). The group of box plots labeled "Model" shows the AUC distribution obtained for the 12 best fitting parameter sets for each population including CR, PR, and NR with shaded backgrounds that are the range (±std) of AUC obtained from clinical PK data. The group of box plots labeled "+Dose" shows the AUC distribution for each population when dose was randomized within the range reported in the hypothetical population, "+B0" shows the distribution when initial tumor volume was randomized, and "+Dose / B0" shows the distribution when both dose and initial tumor volume were randomized. Cmax distributions plotted as in CA. Response to DF treatment was defined as tumor AUC below 10,000 cells*day / μL, and each patient in the hypothetical CR population with the randomized dose and tumor volume (+Dose / B0) was assessed for response (black binary data points). Logistic regressions (shown as solid lines with 95% confidence intervals labeled "+Dose / B0") on tumor volume (D), Cmax (E), or the quotient of Cmax and tumor volume (F) reveal how each predicts response. In contrast, a uniform random sampling of parameter space (1000 parameter sets) does not show these response relationships (dashed lines and confidence intervals labeled "random parameters"). The clinical covariates of response calculated using the hypothetical population have the same trend as the published covariates of response for Yescarta® (dotted line labeled "Yescarta").Note that the response covariates for Yescarta® were increased linearly to match the range of the hypothetical population for the plots.

[0072] FIG. 6 shows the extension of the model to Abecma® dose response, according to one embodiment. Model training (A and B): Fit of the toggle switch model to the Phase I dose-response data and the observed desired fit, with Pearson's linear correlation coefficient from the goodness-of-fit plot (FIG. 13) being 0.59 for CAR-T cells and 0.74 for tumors. Model analysis (C, D, and E): Comparison of the percentage of the total T cell population across doses for memory, effector, and exhausted groups by plotting the mean values ​​against the best fitting parameter set from the simulation. At the low dose, the T cell population is mostly exhausted, while at the high dose, the memory and effector cell populations persist for longer. Model testing (F and G): Comparison of predicted simulations at two doses with data reported in known studies (150-450M cell doses). Tumor dynamics out to one year fall within the range predicted for 150-450M cell doses.

[0073] Figure 7A shows a flow chart of the response classifier workflow according to one embodiment. Using a database of gene signatures (e.g., DAVID, PROGENy, Hallmark, etc.) and a reference gene expression dataset, GSEA can be performed to find a subset of gene signatures that are statistically significantly enriched in either NR or CR groups. Then, ssGSEA scores can be calculated for the subset of signatures, hierarchically clustered into 26 modules, and seeded with one term per module into a machine learning classifier to predict clinical response. Each classification model can be trained using a genetic algorithm until convergence, and multiple models are fitted by seeding with different starting terms to ensure robustness.

[0074] Figure 7B shows an illustration of a response classifier workflow, showing example data at each step in the pipeline, using a process similar to that discussed with respect to Figure 7A, according to one embodiment. Interpretation of differential expression analysis, shown as a volcano plot, via GSEA allows for the selection of a subset of gene signatures. ssGSEA scores are then calculated for this subset, clustered on terms to maximize and / or increase correlation between clusters while simultaneously minimizing and / or reducing average cluster size (depicted as heatmaps and bar plots). Machine learning models can be trained and evaluated, and example results are shown showing cross-validated accuracy and distribution of response predictions.

[0075] Figure 7C shows a flow chart of the response classifier workflow according to one embodiment. Using a database of gene signatures (e.g., DAVID, PROGENy, Hallmark, etc.) and a reference gene expression dataset, GSEA can be performed to find a subset of gene signatures that are statistically significantly enriched in either the NR or CR group. The top X (e.g., 10, 20, 30, 40, etc.) significantly enriched gene signatures from the subset of gene signatures are selected / identified for use as feature vectors and seeded into a machine learning classifier with one term per module to predict clinical response. Each classification model can be trained using a genetic algorithm until convergence, and multiple models were fitted by seeding with different starting terms to ensure robustness. In some implementations, unlike Figures 7A and 7B, clustering is not performed. In some implementations, pathways are ranked by false discovery rate (FDR) adjusted p-value (p) based on the difference between the CR population vs. the NR population. For example, a p-value cutoff threshold can be set at p<0.05. In at least one example, a p-value cutoff threshold of p<0.05 resulted in a set of the top 30 pathways. However, it should be noted that any suitable p-value threshold (e.g., 0.01, 0.02, etc.) can be used.

[0076] FIG. 7D shows an illustration of a response classifier workflow, showing example data at each step in the pipeline, using a process similar to that discussed with respect to FIG. 7C, according to one embodiment. Interpretation of the differential expression analysis, shown as a volcano plot via GSEA, allows for the selection of a subset of gene signatures.) The top X (e.g., 10, 20, 30, 40, etc.) significantly enriched gene signatures from the subset of gene signatures are selected / identified for use as feature vectors and input into a genetic algorithm and / or model. The feature vector (e.g., the top 30 pathways) is input into a genetic algorithm and / or model to perform feature selection. This output pathway (e.g., 2-6 pathways) is then fed into a multivariate logistic regression for model fitting. If the genetic algorithm and / or model is probabilistic, the features it selects may change each time it is run (hence the 2,500 iterations to create a distribution of models). Machine learning models can be trained and evaluated, and example results showing cross-validated accuracy and distribution of response predictions are presented.

[0077] Figure 8 shows a PKPD model workflow according to one embodiment. A mechanism-based kinetic model that integrates features of the cell therapy is trained on the PKPD data (e.g., CAR-T cells and tumor cells) to accurately predict kinetics. Given the patient's tumor burden, the trained model can be used to select a dose of the cell therapy that optimizes the clinical response.

[0078] FIG. 9 shows a goodness of fit plot of the Kymriah® model fitting of FIG. 1, according to one embodiment.

[0079] 10 shows a Local Parameter Sensitivity Analysis of the CR / PR / NR populations according to one embodiment. Local Parameter Sensitivity Coefficients (LPSCs) were calculated for the CR, PR, and NR populations using the AUC for T cells (aucT) and the AUC for B cells (aucB) as outputs, and the samples and outputs were organized by agglomerative hierarchical clustering.

[0080] FIG. 11 shows a volcano plot of differentially expressed genes between CR versus NR groups, according to one embodiment.

[0081] Figure 12 shows the selected gene sets that were differentially enriched between the CR group and the NR group according to one embodiment. The gene sets were derived from BioCarta, Reactome, DAVID, Fraitta et al., Thymic Cell Atlas, and PROGENy and expressed as signed log10(P value).

[0082] 13 shows the average receiver operating characteristic (ROC) curves of 2500 trained models on Bai 2022 and Fraietta data, according to one embodiment. Lines 1301A, 1301B represent the average performance of trained models using selected pathways, while lines 1303A, 1303B represent the average performance of trained models using randomly selected pathways. (B) PR and PR TD The classification of the samples (not used in model training) is set to a value between 0 and 1.

[0083] Figure 14 shows that PKPD response depends on initial tumor burden and CAR-T dose in CR according to one embodiment. Model simulations were performed across a grid of CAR-T doses, initial tumor burdens, and parameter sets in the CR population to determine (A) the mean tumor AUC and (B) the mean CAR-T Cmax. Tumor AUC increases with initial tumor burden and decreases with initial CAR-T dose for parameters in CR. Cmax shows a more complex relationship, peaking at intermediate tumor burdens and generally increasing with initial CAR-T dose.

[0084] Figure 15 shows a goodness of fit plot for Abecma®, according to one embodiment. The data was fitted simultaneously, with Pearson's linear correlation coefficients of 0.59 for CAR-T and 0.75 for tumor.

[0085] FIG. 16A shows a flowchart of a method for training a machine learning classifier, according to one embodiment. FIG. 16B shows a flowchart of a method for training a machine learning classifier, according to one embodiment. FIG. 17 shows a flowchart of a method for predicting a patient's clinical outcome in response to a cell therapy treatment, according to one embodiment. FIG. 18 shows a flowchart of a method for generating a predicted clinical outcome, according to one embodiment. FIG. 19 shows a flowchart of a method for predicting a patient's clinical outcome in response to a cell therapy, according to one embodiment. FIG. 20 shows a flowchart of a method for administering cell therapy to a patient, according to one embodiment. FIG. 21 shows a flowchart of a method for generating a cell therapy product, according to one embodiment. FIG. 22 shows a flowchart of a method for determining a patient-specific dosage of a cell therapy to be administered, according to one embodiment. FIG. 23 shows a block diagram of a system for implementing one or more concepts / methods discussed herein, according to one embodiment. Further details related to FIGS. 16A-23 are discussed below.

[0086] 24A-24B show a list of model parameters, units, and lower and upper bounds used in the particle swarm optimization algorithm, according to one embodiment.

[0087] FIG. 25 shows 12 example parameter sets for CR (complete responder) for the parameters of FIGS. 24A-24B, according to one embodiment.

[0088] FIG. 26 shows 12 example parameter sets for PR (partial responders) for the parameters of FIGS. 24A-24B, according to one embodiment.

[0089] FIG. 27 shows 12 example parameter sets for NR (non-responder) versus the parameters of FIGS. 24A-24B, according to one embodiment.

[0090] FIG. 28 shows a representation of the structure of the original model and possible variations, according to one embodiment.

[0091] Figure 29 shows the fitting accuracy of the model for five structural variants. A Mean Square Error (MSE) of the original (full) model and the five variant models fit to the training data (from Fraietta) and a random sampling of the parameter search space for the original model. The MSE plots are separated by fit to pharmacokinetics and tumor dynamics and ranked by overall fit. B Model simulations are overlaid with the training data for the original (full) model and the five variants.

[0092] Figure 30 shows model fitting to preclinical CD19-CAR-T data, according to one embodiment. NALM-6 xenograft-bearing mice (injected with 106 tumor cells on day -6) were treated with increasing doses of Kymriah® and tumor size was measured by fluorescence imaging (e.g., Reporter Enzyme Fluorescence (REF) imaging). For simplicity, a 1:1 scaling relationship between photons / second and tumor cell count can be assumed. To fit mice for comparison with clinical data, the tumor-related parameter B 最大Boundaries for TC (maximum tumor size) and TK50 (T cell EC50 that drives tumor cell killing) were lowered, and tumor growth rates (uB) were floated between 0.1 and 1 per day. Data and model simulations of (A) pharmacokinetics and (B) tumor kinetics for CAR-T doses of 0, 1 million, 3 million, and 5 million cells. Goodness-of-fit plots of CAR-T pharmacokinetics (C) and tumor kinetics (D), with Pearson correlation coefficients of 0.88 and 0.84, respectively.

[0093] Figure 31 shows model fitting to preclinical BCMA-CAR-T data, according to one embodiment. MM1.s xenograft-bearing mice (5x10 on days -14 to -8) 6 In order to compare and match mice with clinical data, we used the tumor-associated parameter B. 最大 Boundaries for TC (maximum tumor size) and TK50 (T cell EC50 that drives tumor cell killing) were lowered, and tumor growth rates (uB) were floated between 0.1 and 1 per day. Data and model simulations of (A) pharmacokinetics and (B) tumor kinetics for CAR-T doses of 0, 1 million, 3 million, and 10 million cells. Goodness-of-fit plots of CAR-T pharmacokinetics (C) and tumor kinetics (D), with Pearson correlation coefficients of 0.92 and 0.99, respectively.

[0094] FIG. 32 shows the memory (T M ) cells, effector (T E ) cells, and exhaustion (T X ) Shows the simulated frequency of CAR-T cells.

[0095] Figure 33 shows pairwise Pearson correlation coefficients between Thymic Atlas cell population gene signatures calculated using ssGSEA scores from the RNAseq data of Fraietta et al. M cells, CD4+ T M Note that the CD4+ T cell signatures are closely correlated.

[0096] Figure 34 shows single-cell RNA sequencing of twelve pre-infusion CAR-T products classified by response. A UMAP projection of cells annotated by response category: CR; complete response, RL; relapse, NR; non-response. B UMAP projection of cells annotated as TEMRA using SingleR58 with a reference dataset of healthy donor-derived T cells (Bai et al., 2021). C Cell type frequency for CR, RL, and NR groups. Only cell types with at least 5% representation in the samples are shown. Data are shown as mean +- standard error, with individual sample frequencies overlaid as dots. Significance is shown for all significant comparisons between CR and NR according to Wilcoxon rank sum test (p ≤ 0.05). D GSEA by cell type for selected pathways showing significant enrichment in at least one cell type (p ≤ 0.05). A positive normalized enrichment score (NES) indicates higher enrichment in CR / non-exhausted cells.

[0097] Figure 35 shows the comparative pharmacokinetics of Kymriah® in B-ALL, the model described herein for Kymriah® in CLL, and Yescarta® in LCBCL. The distribution of Cmax and C(t=3 months) of Kymriah® in B-ALL and CLL was obtained via simulation of the model of Stein et al. (see, e.g., Stein, A. M. et al. Tisagenlecleucel Model-Based Cellular Kinetic Analysis of Chimeric Antigen Receptor-T Cells. Cpt Pharmacometrics Syst Pharmacol 8, 285-295 (2019), the contents of which are incorporated herein by reference in their entirety), and the model described herein, respectively. Data for Yescarta® was digitized from Locke et al. (See, e.g., Locke, FL et al. Tumor burden, inflammation, and product attributes determine outcomes of axicabtagene ciloleucel in large B-cell lymphoma. Blood Adv 4, 4898-4911 (2020), the contents of which are incorporated herein by reference in their entirety). Distributions shown for Kymriah® and the models described herein are as in FIG. 5C for the left panel (Cmax). The labeled model groups in the box plot show the Cmax or C(t=3 months) for each of the three populations (CR, PR, and NR) with a background of ranges of Cmax or C(t=3 months) obtained from clinical PK data. Yescarta® bar pairs 3501A, 3501B, 3503A, 3505B, 3505A, 3505B show the range of hand estimates of Cmax and C(t=3 months), while the labels (DR for sustained response, RL for relapse, and NR for no response) are plotted at the median hand estimate.

[0098] FIG. 36 shows simulated Cmax / tumor burden vs. tumor response for Abecma®. Model fits to Abecma® Phase I data were simulated at CAR-T doses ranging from 100,000 to 1 billion cells. Tumor shrinkage was calculated at day 60 compared to untreated controls. Response covariates follow the same trends observed with Yescarta® and predicted with Kymriah® in diffuse large B-cell lymphoma (DLBCL).

[0099] FIG. 37 shows single-cell RNA sequencing of two pre-infusion CAR-T products separated by response, according to one embodiment. A Single cells were labeled using SingleR and formed distinct clusters in UMAP space. B Proportions of cell types in CAR-T products separated by patient response status reveal differential frequencies of several canonical cell types. C Gene set enrichment analysis within memory T cell subtypes (for cell types with more than 1% of cells per group annotated). Cells are displayed according to their normalized enrichment statistics, with CR indicating high CR and NR indicating high NR (except for the last row, which indicates high CD8+ EMRA subtypes compared to other CD8+ subtypes). EM: effector memory, CM: central memory, EMRA: effector memory re-expressing CD45RA, CFU-G: colony forming units of granulocytes.

[0100] FIG. 38 shows that Kymriah® response in CLL can be accurately predicted using an ssGSEA-based classifier, according to one embodiment. A. Using a probabilistic workflow, a family of 12,000 multivariate logistic regression classifiers was parameterized using the ssGSEA scores of samples as input features. Accuracy distribution from leave-one-out cross-validation (LOOCV) of the 12,000 parameterized classifiers using a threshold of 0.5. B. Individual predictions split by response category. CR and NR samples are accurately classified, while PR and PR TD The samples of (not used in model training) are classified as intermediate between the CR and NR groups.

[0101] FIG. 39 shows the (A) cluster size and (B) mean intra-cluster correlation distribution based on a variable number of separate clusters calculated for 864 gene signatures differentially enriched between CR and NR groups.

[0102] FIG. 40 shows that single-cell RNA sequencing of pre-infusion CAR-T products reveals cell-intrinsic defects associated with non-sustained responses, according to one embodiment. Uniform Manifold Approximation and Projection (UMAP) projections of three datasets representing Kymriah® in acute lymphoblastic leukemia (ALL) (A-C), Kymriah® in large B-cell lymphoma (LBCL) (D-F), and Yescarta® in LBCL (GI). A,D,G UMAP projections annotated by response category: CR; complete response, RL; relapse, PR; partial response, NR; non-response. B,C,H UMAP projections annotated as exhausted using ProjectTILs ("Tex" is exhausted T cells, "Non-Tex" is non-exhausted T cells). C, F, I Projection of UMAP annotated as high (above the mean) (denoted as “Dys. High”) or low (below the mean) (denoted as “Dys. Low”) CAR-T cell dysfunction signature from Good et al. E GSEA by cell type on selection pathways comparing both exhausted vs. non-exhausted and CR vs. PR / RL / NR categories in cells annotated as T effector memory (Tem by ProjecTILs) or early memory (Tmem by CITESeq, CD8+CD45RO-CD27+). In some implementations, CITESeq refers to a sequencing-based method that simultaneously quantifies cell surface proteins and transcriptional data, e.g., mRNA, within a single-cell readout. A positive normalized enrichment score (NES) indicates higher enrichment in CR / non-exhausted cells. *NR=NR / RL or NR / PR.

[0103] Figure 41 shows that CD19-CART responses can be predicted from infusion products using a ssGSEA-based transcriptome classifier with better accuracy than T cell immunophenotypes, according to one embodiment. The distribution of prediction accuracy is shown for 2500 iterations using a 60:40 training:test split cross-validation. Results of the transcriptome-based ssGSEA classifier are compared against (A) a reported classifier based on the frequency of T memory (CD8+CD45RO-CD27+) and T exhausted (CD8+PD1+) cells from Fraietta et al. (B) a bivariate classifier based on the frequency of T memory (CD8+CD45RO-CD27+) and T exhausted (CD8+PD1+) cells calculated from Bai et al. 2022 (C, D) a bivariate classifier based on the frequency of T effector memory and T exhausted cells from annotations in ProjecTILs from Haradhvala et al. (see, e.g., Haradhvala, NJ et al. Distinct cellular dynamics associated with response to CAR-T therapy for refractory B cell lymphoma. Nat Med 1-12 (2022), the contents of which are incorporated herein by reference in their entirety). As a control, the accuracy distribution obtained from the null model (random classification) is shown. *** indicates Wilcoxon rank sum test with p<10-15. E CART response scorecard, representing the 28 gene signatures provided to the transcriptome classifier ranked by differential GSEA in Fraietta et al. The bubble size indicates the frequency of inclusion in the 2,500 trained models after feature selection, and the shading indicates differential enrichment between response groups by dataset based on GSEA (score=-1×sign(NES)×log10-P value) by pseudobulk analysis. Reference 4506 shows three examples of bubble sizes and the frequencies they represent, with the larger the circle, the higher the frequency. More specifically, reference 4506 shows a bubble size representing a frequency of 10, a bubble size representing a frequency of 30, and a bubble size representing a frequency of 70.Shaded scores in score range 4502 are enriched for NR / PR / RL and shaded scores in score range 4504 are enriched for CR, in other words, circles in solid boxes are enriched for CR and circles in dashed boxes are enriched for NR / PR / RL. Gene signatures are annotated by source.

[0104] Figure 42 shows the results of model fitting based on the hypothesis that the only distinguishing feature between CR, PR, and NR populations is the fraction of memory and exhausted T cells in the CAR-T infusion product. The fraction of memory cells (f_Tm) and fraction of exhausted cells (f_Tx) can be estimated to be 1-50% independently for CR, PR, and NR populations, while other model parameters can be estimated simultaneously using a single vector for CR, PR, and NR populations. Simulations of best fit models from 12 optimization runs for (A) CAR-T pharmacokinetics and (B) tumor dynamics (e.g., estimated by MSE minimization). (C) Estimated fraction of memory and exhausted cells (f_Tm, f_Tx) in the CAR-T infusion product for CR, PR, and NR populations. Bars represent median ± 25th percentile intervals from n=12 model fits.

[0105] Figure 43 shows simulated pharmacokinetic and tumor dynamic responses to increasing cell doses of pure memory cell populations from the CR, PR, and NR population models. Using parameter sets estimated for the CR (A), PR (B), and NR (C) populations, simulations were performed at doses of 1 million (labeled "1M"), 3 million (labeled "3M"), 10 million (labeled "10M"), 30 million (labeled "30M"), 100 million (labeled "100M"), 300 million (labeled "300M"), and 1 billion (labeled "1000M"). For direct comparison, the percentage of memory cells was set to 100% for each.

[0106] Figure 44 shows T cell phenotyping of CAR-T infusion products. Cell frequencies annotated in ProjecTILs by response category for (A) Kymriah® in ALL, n=12, digitized from Bai et al. 2022; (B) Kymriah® in LBCL, n=11, digitized from Haradhvala et al.; (C) Yescarta® in LBCL, n=19, digitized from Haradhvala et al. Cell types annotated with frequencies below 5% were excluded: CD4+ naive, CD8+ naive, CD8+ T precursor exhausted (Tpex), and CD4+ follicular helper (Tfh). Frequency of immunophenotyped T early memory cells (CD8+CD45RO-CD27+) and exhausted cells (CD8+PD1+) by response category for Kymriah® in CLL, n=38, digitized from D Fraietta et al. Frequency of immunophenotyped early memory and exhausted cells by response category for Kymriah® in ALL, n=12, calculated from CITEseq antibody tags from B Bai et al. 2022. Box plots represent median ± 25th percentile, whiskers are minimum / maximum or additional 1.5x interquartile distance.

[0107] Figure 45 shows the performance of the transcriptome classifier compared to the null model and the random pathway model. The distribution of prediction accuracy is shown for 2500 iterations using a 60:40 train:test split cross-validation. The results of the transcriptome-based ssGSEA classifier of 28 signatures ("Transcriptome") are compared to the null model (random classification; "Null") and to an ssGSEA classifier trained on a randomized selection of pathways from the compendium ("Random").

[0108] Example 1: Method

[0109] Clinical Data: Kymriah® (tisagenlecleucel, CTL019)

[0110] Mean pharmacokinetic and tumor dynamics profiles were digitized from clinical trials of chronic lymphocytic leukemia (CLL) patients treated with Kymriah® (tisagenlecleucel, CTL019) and divided into complete responders (CR, n=8), partial responders (PR, n=5), and non-responders (NR, n=25). Samples annotated as PRTD (late relapse to B-cell lymphoma) were excluded. Patients had a CR of 0.14–11 × 10 8 Patients were treated with a range of CAR-T doses, from 10 to 400 mg / kg / day, consistent with the median dose used in this and other clinical trials of Kymriah®, for parameter estimation. 8 A fixed dose of 10 cells was assumed. Tumor size data were reported as B cells / μL and were therefore used directly in the model fitting (a total of 10 10 (Assuming an initial tumor burden of 100–200 cells.) Pharmacokinetics were reported as CD19-CAR transgene copy numbers in peripheral blood (copies per μg of genomic DNA) and converted to cell numbers for mechanistic modeling (see below).

[0111] A nonlinear mixed-effects model of Kymriah® cell kinetics was used to simulate population pharmacokinetics in refractory B-cell acute lymphoblastic leukemia (B-ALL). 8 The parameters were parameterized using data compiled from two clinical trials in which 1000 patients (n=91) were treated with a median dose of 1000 cells. In some cases, the pharmacokinetic profile of Kymriah in CLL patients is not substantially different from B-ALL patients. To calculate the distribution of exposure (AUC and Cmax), the PK profiles of 1000 virtual patients were simulated. At each time step (0.1 days over a year), the 1st to 99th percentiles were calculated, and the AUC and Cmax were calculated from these percentiles.

[0112] Clinical Data: Abecma® (Idecbutagenvir ubiquinol, BB2121)

[0113] Mean pharmacokinetic and tumor dynamics profiles were digitized from a Phase 1 dose-escalation study (Locke et al.) of refractory multiple myeloma (MM) patients (n=33) treated with Abecma® across dose groups (50, 150, 450, and 800×10 6 Tumor size data were reported as % change in serum BCMA levels. For model fitting, initial tumor volume was adjusted to 10 10 A linear scaling was performed between tumor burden and reported soluble BCMA, assuming 100 cells. PK data are reported as transgene copies per μg of DNA and converted to CAR-T cell numbers by applying the same scaling factor as above. Mean PK and tumor kinetic profiles ± standard deviation are calculated using a 150 × 10 6 and 450×10 6 Digitized from a Phase 2 study in the same patient population (N=128) treated with cell doses of 100-fold (data not separated by dose) (see, e.g., Munshi, NC et al. Idecabtagene Vicleucel in Relapsed and Refractory Multiple Myeloma. New Engl J Med 384, 705-716 (2021), the contents of which are incorporated herein by reference in their entirety). Tumor kinetics data in this study were reported as serum BCMA (ng / mL). Data were converted to % change from baseline and rounded to 10 for comparison to model simulations. 10 An initial tumor burden of 10 cells was assumed.

[0114] Scaling factors and virtual populations

[0115] To estimate the scaling factor between transgene counts and cell numbers, data from Kalos et al. (See, e.g., Kalos, M. et al. T Cells with Chimeric Antigen Receptors Have Potent Antitumor Effects and Can Establish Memory in Patients with Advanced Leukemia. Sci Transl Med 3, 95ra73-95ra73 (2011), the contents of which are incorporated herein by reference in their entirety) were used, with both counts / ug and total circulating CD19+ cells reported, approximately 10 4 For conversion between total cell counts and cells / μL for plots, a total blood volume of 2 L for humans and 2 μL for mice was assumed.

[0116] Model Structure and Assumptions

[0117] These may encode three functionally distinct T cell populations: T memory cells capable of long-term renewal (self-renewal) and differentiation (T M ), T effector cells (Ts) that arise from memory populations and are directly involved in tumor cell killing. E ), and exhausted T cells that lack effector function and proliferation capacity (T X ). T effectors can proliferate for N population doublings but may lack the ability to self-renew. One aspect of the mechanism-based description of T cell differentiation control is the toggle switch sensor of tumor antigens, encoded as the Hill equation. This toggle switch regulates the rate of self-renewal versus differentiation of T memory cells, the rate of T effector proliferation, the rate of T effector exhaustion, and the regeneration of T memory cells from T effectors.

[0118] This control of T cell fate can be described via a system of nonlinear ordinary differential equations (ODEs), for example as follows:

number

[0119] Here, the self-renewal and differentiation of memory cells occurs at a rate μ M and B cell antigen B A The parameter f is adjusted via a Hill-like switch that depends on 最大 describes the proportion of memory cells that undergo self-renewal as opposed to differentiation into effector cells. E2 It is reproduced from the population (speed parameter r M The effector population is divided into two subgroups: E1 and T E2 This division can be for mathematical simplicity, i.e., the non-tumoricidal subgroup is differentiated from memory cells and cytotoxic effector cells (T E2 ) is further differentiated into the rate parameter μ E For the parameter estimation routine, instead of using a hierarchy of ODEs that track the number of cells that have each undergone n rounds of differentiation, we use a E2 We can encode the N population doublings in a single source term in Eq. T effector cells are expressed by a rate parameter k ex and the T cell population is exhausted with the corresponding kinetic parameter d M , d E1 ,d E2 , d X This toggle switch is activated by the B cell antigen B A Note that the function is encoded as the Hill function in and has the same half-maximal parameter B50 across all T cell populations, but different exponents (km, kr, km, ke, and kx) that account for the putative differential dose-response relationships.

[0120] The dynamics of B-cell tumors are characterized by the rate μ B and carrying capacity B 最大 Logistic growth with rate k 死滅 Nonlinear tumor killing via effectors bearing B cell antigen BA can be modeled with the production and decay of, for example,

number

[0121] Tumor antigen (B) A ) to encode proliferation / differentiation as driven by the production rate degradation rate (k B1 , k B2 ) generates an alternative transient compartment, which accounts for the time delay between changes in tumor burden and T cell fate responsiveness. Transient compartments can be employed in PK / PD modeling to link drug concentrations to measured pharmacokinetic responses.

[0122] Two empirical, rapid reactions can be implemented to map cell dose to initial conditions. First, a portion of the infused cell dose is lost quickly to account for the discrepancy between cell dose and initial conditions observed in both clinical and preclinical models when cell counts / μL are reported. Second, because the initial cell dose is rapidly converted into four T cell subpopulations, this composition does not need to be specified in advance and can be determined through parameter estimation. This reaction accounts for the fact that the CAR-T product consists of a mixed population of T cells (memory, effector, and exhausted), and this composition may vary and therefore does not need to be specified in the clinical data. Rather than pre-specifying the composition via initial conditions, the rapid conversion reaction allows each proportion to be estimated as a model parameter. This can be accomplished via the following set of equations, where Dose is the dose of CAR-T administered and Dose X is the remaining dose fractionated into T cell subpopulations, for example:

number

[0123] A zero bound can be applied to cell populations to limit artificial regrowth; i.e., in some implementations, if any cell population has a fraction (less than 1), the cell population can be set to 0. Particle swarm optimization (PSO) can be used to estimate model parameters based on minimization of the log-mean-square error between the model simulation and the data. The model structure can be coded in Matlab® SimBiology® (R2021a) and particle swarm optimization (PSO) can be used to run the particle swarm function with 100 particles × 100 iterations, and all CAR-T cells are swarmed with 100 particles × 100 iterations. 6 Using LLQ set to 1, model parameters based on minimizing the log-mean squared error (MSE) between the model simulation and the data can be estimated. Models can be fitted separately to the CR, PR, and NR populations by applying the PSO algorithm 12 times for each population, generating a total of 36 parameter sets for analysis. See Figures 24A-24B for a list of model parameters, units, and lower and upper limits used in the PSO algorithm. See Figure 25 for the 12 CR parameter sets, Figure 26 for the 12 PR parameter sets, and Figure 27 for the 12 NR parameter sets.

[0124] Variant models based on alternative T cell population structures were also evaluated for their ability to fit the data, but none were superior to the formulation above (see below and Figures 28, 29).

[0125] Local parameter sensitivity analysis

[0126] Local parameter sensitivity coefficients (LPSCs) can be calculated by simulating the model and calculating the CAR-T AUC and tumor AUC in response to a 10% increase in parameter values ​​estimated across the 36 parameter sets characterizing the CR / PR / NR populations. The coefficients can be calculated based on the median change in AUC for each population according to the following formula:

number

[0127] Virtual population

[0128] The virtual population samples a baseline set of parameters, but with log-uniform sampling within the reported range of CAR-T dose (10 7 ~10 9 cells) and initial tumor burden (8.5 × 10 8 ~2.7×10 10 It can be created by fitting CR / PR / NR populations using a Monte Carlo method that varies the number of cells.

[0129] Modeling Workflow

[0130] One strategy for model-based integration of heterogeneous data sets is to: 1. fit a PKPD model to the CR, PR, and NR profiles from Fraietta et al. independently; 2. create a virtual population from this model and compare the predicted population PK variance to the Kymriah® data from Stein et al. and the response covariates to the Yescarta® data from Locke et al.; 3. fit the PKPD model to the Abecma® dose-response data from Raje et al. to understand the mechanisms underlying the response covariates.

[0131] RNASeq analysis

[0132] Bulk RNASeq data analysis was performed in R version 4.1.1. Read count data were downloaded from the supplement provided by Fraietta et al. (See, e.g., Fraietta, JA et al. Determinants of response and resistance to CD19 chimeric antigen receptor (CAR) T cell therapy of chronic lymphocytic leukemia. Nat Med 24, 563-571 (2018), the contents of which are incorporated herein by reference in their entirety), normalized with TMM, and converted to log (counts per million) by applying the Voom transformation. Differential gene expression was performed using limma (see, e.g., Ritchie, ME et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res 43, e47-e47 (2015), the contents of which are incorporated herein by reference in their entirety), and gene signature enrichment was estimated by single-sample GSEA (see, e.g., Barbie, DA et al. Systematic RNA interference reveals that oncogenic KRAS-driven cancers require TBK1. Nature 462, 108-112 (2009), the contents of which are incorporated herein by reference in their entirety). The normalized ssGSEA score for each pair of sample and gene signature can be calculated, for example, as follows:

number

[0133] where A is the matrix of ssGSEA signature scores (i) by sample (j).

[0134] Gene signatures of cell signaling pathways can be generated using a variety of tools, including PROGENy (10 gene set) (see, e.g., Schubert, M. et al. Perturbation-response genes reveal signaling footprints in cancer gene expression. Nat Commun 9, 20 (2018), the contents of which are incorporated herein by reference in their entirety), BioCarta (217 gene set) (see, e.g., BioCarta Pathways. https: / / maayanlab.cloud / Harmonizome / dataset / Biocarta+Pathways, the contents of which are incorporated herein by reference in their entirety), Reactome (674 gene set) (see, e.g., Gillespie, M. et al. The reactome pathway knowledgebase 2022. Nucleic Acids Res 50, gkab1028- (2021), the contents of which are incorporated herein by reference in their entirety), Hallmark (50 gene set) (see, e.g., Liberzon A, Birger C, Thorvaldsdottir H, et al (2015) The Molecular Signatures Database Hallmark Gene Set Collection. Cell Syst 1: 417-425. https: / / doi.org / 10.1016 / j.cels.2015.12.004, the contents of which are incorporated herein by reference in their entirety), and DAVID (6577 gene set) (see, e.g., Huang, DW, Sherman, BT & Lempicki, RA Systematic and integrative analysis of large gene lists using DAVID bioinformatics resources. Nat Protoc 4, 44-57 (2009).The cell population signatures were compiled from those published in Fraietta et al. (7 sets of cell population signatures), as well as the single cell atlas of thymic development (13 sets of cell population gene signatures) (see, e.g., Park, J.-E. et al. A cell atlas of human thymic development defines T cell repertoire formation. Science 367, eaay3224 (2020), the contents of which are incorporated herein by reference in their entireties), and individual signatures of CAR-T dysfunction and CD28z tonic signaling (Tables 1, 2). [Table 1] [Table 2]

[0135] In addition, two supplementary datasets are provided herewith (Supplementary Datasets A, B). Supplementary Dataset A shows the top 30 pathways scored for the number of times they were included in the final model ("Model Counts") (for both Fraietta and Bai versions). Supplementary Dataset B shows the top 30 pathways scored for the number of times they were included in the final model ("Model Counts") (for Fraietta, Bai, Haradhvala (Yescarta®) and Haradhvala (Kymriah®) versions).

[0136] ssGSEA-based response classifier

[0137] Single-sample GSEA scores (30, based on an FDR-adjusted p-value of 0.05) corresponding to gene signatures differentially enriched between the CR and NR groups can be used to construct a logistic regression-based classifier of the following response states:

number

[0138] where p(CR) is the probability of complete response (vs. non-response) and β i are the regression coefficients. A genetic algorithm, implemented in R using the glmulti package (1.0.8), was used for feature selection with a training split of 60% of the data using Akaike Information Criterion (AIC) as the objective function. For example, a population size of 100 can be used with a mutation rate of 0.001, a transition rate of 0.3, and a reproduction rate of 0.1. This was repeated 12,000 times due to the stochastic nature of the genetic algorithm. In other implementations, other population sizes, mutation rates, transition rates, and / or reproduction rates can be used. Furthermore, in other implementations, the genetic algorithm can be repeated any number of times. Such processing (e.g., grouping of gene signatures) can reduce computational requirements and allow for more effective selection of feature vectors. A computing device can then select a feature vector, use the feature vector to train a classifier, and / or use the classifier and / or feature vector to generate an output much more quickly and / or with reduced computational requirements. An example of this kind of workflow is illustrated in the workflow (Figure 7C and Figure 7D). For the randomized control model (used as a comparison), 2-6 pathways were randomly selected from a compendium of pathways as input features based on the feature frequencies observed in the trained model (in other implementations, any other number of pathways can be randomly selected).

[0139] In an alternative embodiment, single-sample GSEA scores (860, based on an FDR-adjusted p-value of 0.05) corresponding to gene signatures differentially enriched between the CR and NR groups can be used to construct a logistic regression-based classifier of response status, for example:

number

[0140] where p(CR) is the probability of complete response (vs. non-response) and β i are the regression coefficients. To limit the number of input features, Pearson correlation-based hierarchical clustering can first be used to group the gene signatures into 25 clusters. This number can be chosen to simultaneously maximize and / or improve the correlation between clusters and the cluster size (Figure 39). In other implementations, other numbers of clusters can be used. The model can then be randomly seeded with the 25 signatures by drawing one per cluster. A genetic algorithm, implemented in R using the glmulti package, was then used for feature selection using the Akaike Information Criterion (AIC) with model accuracy as the objective function. Model accuracy can be defined, for example, as follows: Accuracy=(TP+TN) / (TP+TN+FP+FN) In the formula, TP, TN, FP, FN refer to true positive, true negative, false positive, and false negative, respectively. For example, a population size of 100 can be used with a mutation rate of 0.001, a transition rate of 0.3, and a reproduction rate of 0.1. Due to the stochastic nature of the genetic algorithm, this was repeated 12,000 times. In other implementations, other population sizes, mutation rates, transition rates, and / or reproduction rates can be used. Furthermore, in other implementations, the genetic algorithm can be repeated any number of times. Such processing (e.g., grouping of gene signatures) can reduce computational requirements and allow for more effective selection of feature vectors. A computing device can then select a feature vector, use the feature vector to train a classifier, and / or use the classifier and / or feature vector to generate an output much more quickly and / or with reduced computational requirements. An example of this type of workflow is illustrated as workflow (Figure 7A). Figure 7A is similar to Figure 7C, but includes selecting the top 30 most significant pathways from the extracted significant signatures to seed a machine learning classifier (rather than computing per-sample signature enrichment from the extracted significant signatures and clustering the signatures into functional modules). Figure 7B is similar to Figure 7D, but includes selecting the top 30 pathways based on significance with respect to input to a genetic algorithm (rather than performing clustering of per-sample signature enrichment and selecting clusters to seed a machine learning classifier).

[0141] In some implementations, prediction accuracy is assessed using a 40% test split of the data, and model accuracy distributions are compared via Wilcoxon rank sum tests and visualized as kernel density estimates with manually selected bandwidth. Immunophenotype classifiers using an identical workflow except for feature selection can be developed and / or used with input features that are either reported cell frequencies from Fraietta et al., calculated cell frequencies from Bai et al. 2022 CITESeq data, or calculated cell frequencies from the ProjecTILs annotation of Haradhvala et al. data.

[0142] A binomial test can be used to assess GSEA overlap in the CR vs. NR / PR / RL comparison between datasets. Starting with the top 28 gene signatures identified as differentially expressed in Fraietta et al. and used to seed the transcriptome classifier, 13 of 28, 13 of 28, and 15 of 28 are significant at the p<0.05 level in the Bai et al., Haradhvala et al., Kymriah®, and Yescarta® datasets, respectively. Of the 7548 signatures in the compendium, 1123, 742, and 751 met this level of significance, with 6x10 -5 , 7x10 -7 , and 10 -8 This corresponds to the p-value of

[0143] scRNAseq (single-cell RNA sequencing) and CITEseq analyses

[0144] Single-cell RNA sequencing count data (and associated metadata) were provided by Bai and collaborators (see, e.g., Bai, Z. et al. Single-cell multiomics dissection of basal and antigen-specific activation states of CD19-targeted CAR T cells. J Immunother Cancer 9, e002328 (2021), the contents of which are incorporated by reference in their entirety). Single-cell RNA sequencing counts and associated metadata from Bai et al. 2022 and Haradhvala et al. were obtained from GEO (GSE197215 and GSE197268, respectively). Seurat (i.e., a software package for quality control, analysis, and exploration of single-cell RNA-seq data; Seurat allows users to identify and interpret sources of heterogeneity from single-cell transcriptome measurements and integrate diverse types of single-cell data) was used to normalize gene counts, and then cell type labels were assigned by comparison with flow-sorted reference data using SingleR (see, e.g., Aran, D. et al. Reference-based analysis of lung single-cell sequencing reveals a transitional profibrotic macrophage. Nat Immunol 20, 163-172 (2019), the contents of which are incorporated herein by reference in their entirety). Although multiple reference datasets were used to confirm the robustness of cell labels, for final analysis, subpopulations were assigned using the same method as Novershtern et al. (see, e.g., Novershtern, N. et al. Densely Interconnected Transcriptional Circuits Control Cell States in Human Hematopoiesis. Cell 144, 296-309 (2011)).The results were used as specified by the GSVA (1.40.1), the contents of which are incorporated herein by reference in their entirety. Differential expression analysis was performed in Seurat using Wilcoxon rank sum test for each cell type separately, and GSEA was performed on differentially expressed genes using a p-value threshold of 0.05. It is also contemplated that the p-value threshold may be adjusted and GSEA performed on more or less differentially expressed genes (e.g., p-value threshold of 0.1, 0.01, 0.001, etc.). Single-sample GSEA scores were calculated using GSVA (1.40.1) and used without normalization as input features to the classifier.

[0145] In an alternative embodiment, the differential expression analysis may employ a median difference test.

[0146] For CITEseq-based immunophenotyping, each cell was designated positive / negative based on reference to the relevant control antibody tag (Bai et al. 2022).

[0147] software

[0148] Model simulations and analyses were performed using Matlab® R2021a and the SimBiology® toolbox (6.1). Bioinformatics analyses were performed on Ubuntu 20.04.3 LTS running R4.1.1 ("Kick Things"). Some packages used in this work include GSVA (1.40.1) for ssGSEA, fgsea (1.21.2) for GSEA, celldex (1.2.0) to obtain reference datasets for SingleR (1.6.1) and Seurat (4.1.0), data.table (1.14.2), limma (3.50.3), edgeR (3.34.1), Matrix (1.4.3), and ggplot2 (3.3.6) for data wrangling and visualization.

[0149] Example 2:

[0150] Model Structure

[0151] In this example, we investigated T cells (and CAR-T products) that consist of three functionally distinct cell populations: T memory cells (T M ) are capable of long-term self-renewal and immunological memory. T effector (T E ) are involved in target-mediated cell killing and can proliferate by undergoing N cell doublings, but lack the ability to self-renew. T effectors can regenerate T memory cells after antigen clearance. T effectors are capable of regenerating exhausted T cells (T) after carrying out N cell divisions. X ) that lack both killing and proliferation capabilities. The antigen-sensing switch coordinately regulates the decision of memory cells to self-renew versus differentiate, the rate of effector proliferation, exhaustion, and the rate of memory cell renewal from effectors (see Example 1). This represents a biologically sound description of T cell function and regulatory control in response to immunological needs determined by systemic antigen challenge (Figure 1A).

[0152] Model parameterization: CLL patients treated with Kymriah® and grouped by response

[0153] We analyzed whether the mathematical description of T cell regulatory control can quantitatively capture the characteristic profiles of CAR-T pharmacokinetics and tumor dynamics, and whether the parameter estimates reveal something about the biological underpinnings of clinical variability. A known study reported the average PK and tumor dynamics profiles of chronic lymphocytic lymphoma (CLL) patients treated with Kymriah (CTL019; CD19-targeted CAR-T), grouped by complete responders (CR), partial responders (PR), and non-responders (NR). The data (mean ± standard deviation) were digitized and particle swarm optimization was used to estimate model parameters characterizing the three population archetypes (Figure 1B). The parameters can be estimated 12 times per patient group. Although the parameters are non-identifiable (see Tables 25-27), clinical data were captured with good accuracy (Figure 9). Accuracy was quantified by calculating the Pearson correlation of the goodness of fit (log simulation vs. log data), separated by CAR-T and tumor kinetics for the CR, PR, and NR populations. For CAR-T pharmacokinetics, the median correlation coefficients for the CR, PR, and NR populations are 0.89, 0.88, and 0.76, respectively. Thus, the model captures most of the variance in the PK data for all three groups. For tumor kinetics, the correlations are 0.78, 0.77, and -0.23, respectively. The model accurately captures the CR tumor kinetics. For the PR group, the model describes the kinetics of the initial tumor size reduction. For the NR group, the model captured the lack of tumor response. Accuracy was further quantified by comparing the mean squared error resulting from a set of estimated model parameters to the mean squared error obtained by random sampling (n=100) of the parameter search space. All P values ​​for the three groups were less than 10 (rank sum test), indicating that the improved parameters represent a small segment of the parameter space. In some implementations, the tumor growth portion of the model may be minimal (logistic equation) and does not capture aspects of the dynamics, while still capturing overall trends and differences between populations.In particular, in some implementations, both PR and NR tumor dynamics appear to fluctuate up to day 100 and then decline.

[0154] Biological mechanisms that distinguish CR, PR, and NR populations

[0155] To decipher the biological mechanisms underlying the different patient response profiles, the parameter estimates from the three patient populations can first be decomposed into principal components (Figure 1C). Note that the three populations form relatively clear clusters in the parameter space, with the x-axis representing PC1 (accounting for 35.3% of the variance) separating the hypothetical patients by response, and the y-axis representing PC2 (accounting for 21.7% of the variance) separating the CR and NR groups from the PR. Examining the coefficients of principal component 1 (Figure 1D), the lowest value (associated with NR) is for TK50 (the cytotoxic potential of effectors), and the largest positive contribution (associated with CR) is for memory and effector cell turnover (proliferation and mortality rates; μ M , d M , and d E2 ), i.e., in responding patients, CAR-T effectors lyse target tumor cells much more efficiently, and both memory and effector cells circulate at a higher rate. These findings were confirmed by local parameter sensitivity analysis (Figure 10).

[0156] The frequency of memory cells in the CAR-T infusion product, when assessed by standard T cell immunophenotyping, can predict clinical response (see, e.g., (1) Locke et al., (2) Xu et al. Closely related T-memory stem cells correlate with in vivo expansion of CAR. CD19-T cells and are preserved by IL-7 and IL-15. Blood 123, 3750-3759 (2014), the contents of which are incorporated by reference in their entireties). This was one conclusion of Fraietta et al. However, PC1 loading (Figure 1D) is a predictor of memory cell function (μTm) rather than frequency (fTm). M , d M This suggests that cell-intrinsic differences in memory cell frequency (Tm, Te, and Tx cell frequencies) may be a more important determinant of response. To identify the importance of memory cell frequency to function, two experiments were performed. First, the data were fitted under the hypothesis that the only difference between the CR / PR / NR populations was the product composition (frequencies of Tm, Te, and Tx cells), while cell-intrinsic kinetic parameters were preserved (Figure 42). This model captured the pharmacokinetic and tumor dynamics differences between the populations, and the estimated CAR-T product composition was consistent with that reported by Fraietta et al. However, the magnitude of the differences between the populations cannot be fully explained by Fraietta's hypothesis. That is, CAR-T cell composition determined solely by memory and exhausted cell frequencies may be insufficient to explain the variance in clinical activity.

[0157] To directly compare putative differences in memory cell function between the CR / PR / NR groups, a dose-ranging study was simulated using purified memory cell populations from the CR / PR / NR archetype (Figure 43). While CR memory cells produced robust and dose-dependent CAR-T proliferation, persistence, and tumor reduction, NR cells showed very little proliferation or antitumor activity, and PR memory cells displayed somewhat intermediate functionality. These results imply that while memory cell frequency in the CAR-T infusion product contributes to exposure and response, cell-intrinsic characteristics such as proliferation potential may be used to explain variance in clinical outcomes.

[0158] Molecular and cellular features distinguishing CR, PR, and NR populations

[0159] To explore the molecular and cellular features underlying these differences in kinetic parameters, RNA sequencing data can be used and pre-infusion CAR-T products can be sequenced and annotated by response category. Differential expression analysis in CR vs. NR populations revealed biological features (gene signatures) consistent with model predictions (Figure 11, Figure 12). Findings from the original report can be confirmed, and furthermore, the CR population was found to be enriched in gene signatures of CD4+ and CD8+ memory cells (as defined by single cell sequencing of thymic tissue) and may show upregulation of signatures characterizing T cell proliferation, effector cytokine (interferon) signaling, IL2RB, IL7 and JAK / STAT signaling (as defined by curated pathway databases (e.g., (1) BioCarta Pathways. https: / / maayanlab.cloud / Harmonizome / dataset / Biocarta+Pathways, (2) Gillespie, M. et al. The reactome pathway knowledgebase 2022. Nucleic Acids Res 50, gkab1028-(2021), and (3) Liberzon, A. et al. The Molecular Signatures Database Hallmark Gene Set Collection. Cell Syst 1, 417-425 (2015), the contents of each of which are incorporated herein by reference in their entireties. CAR-T cells from NR patients show enhanced p53 and DNA damage signaling pathways that may attenuate the proliferation defect.

[0160] We then used single sample gene set enrichment analysis (ssGSEA) to examine the distribution of pathways and cellular signatures in individual samples, separated by response. The CR population is significantly enriched in the "non-exhausted T cell" signature (Figure 2A). This is broadly consistent with model predictions, as exhausted T cells, by definition, exhibit reduced cytolytic and proliferative capacity. Furthermore, the simulated proportion of non-exhausted cells at day 60 (peak of antitumor efficacy) is significantly higher in the CR group (Figure 2B), whereas CAR-Ts from NR patients rapidly progress to exhaustion (Figure 32). This simulation is also consistent with other clinical analyses reporting that CAR-T products that do not proliferate in vivo exhibit upregulation of exhaustion markers LAG3 and PD1.

[0161] CRs are differentially enriched in both CD8+ and CD4+ memory T cell signatures (Fig. 2C,D). CR populations also show enhanced IL2RB and IL7R signaling (Fig. 2E,F). These pathways converge on the canonical JAK / STAT cascade, indicating that CR cell products may display increased sensitivity to these important cytokines. IL2 and IL7 are known components of CAR-T growth media, and peak serum IL-7 concentrations may predict CD19 CAR-T exposure and progression-free survival (PFS). Although the results shown in Fig. 2 are statistically significant, the ssGSEA distributions overlap between response categories. Thus, in some implementations, none of the evaluated gene signatures are used as univariate classifiers of patient response. Also, note that enrichment of gene sets in bulk sequencing data does not necessarily elucidate the frequency of cell populations or distinguish between transcriptionally similar and covarying cell types (Figure 33). For example, a CR product may have a higher frequency of CD4+ and CD8+ memory cells, or may contain comparable cell frequencies but have a more "memory-like" transcriptome.

[0162] To evaluate the model predictions at deeper resolution, a known single-cell RNA sequencing study of pre-infusion autologous CD19 CAR-T products (with 4-1BB costimulatory domain and similar to Kymriah®) from 12 patients with acute lymphoblastic leukemia was used, including five complete responders (CR), two non-responders (NR), and five relapsed patients (RL). Cell type labels can be assigned by mapping the expression profiles of 31,105 individual cells against tumor-infiltrating lymphocytes and T cell subpopulations from healthy donors using two complementary methods (Figure 3A and Figure 34A,B). Cells from the two products form clearly distinct clusters in UMAP (Uniform Manifold Approximation and Projection) space. UMAP is a dimensionality reduction algorithm based on manifold learning techniques and topological properties of the data that can be used in some implementations to build a high-dimensional graph representation of the data and then optimize the low-dimensional graph to be as structurally similar as possible, creating the UMAP space (the two-dimensional space between UMAP_1 (x-axis) and UMAP_2 (y-axis)). Consistent with the bulk sequencing data, the frequency of both CD8+ and CD4+ central memory cells, the most primitive annotated cell types, can be enriched in CR compared to NR patients (Figure 3B). The majority of cells are T EM Although annotated as NR (effector memory), NR and RL patients showed significant enrichment for exhausted T cells (Figure 3D), as well as CD8+EMRA cells (effector memory cells that re-express CD45RA, Figure 34C) (p=0.04, one-tailed t-test). Thus, regardless of cell annotation method, CAR-T products resulting in non-responsive or non-sustained responses show a higher frequency of a subpopulation of terminally differentiated CD8+ effectors. We annotated the cells using a "CART dysfunction" signature, characteristic of functionally exhausted CAR-T cells, which is accompanied by reduced proliferative and cytotoxic performance. Visually, the dysfunction signature separated into NR / RL categories, consistent with the functional differences predicted by the model (Figure 3C).

[0163] To investigate cell-intrinsic differences underlying the response, differential gene expression analysis was performed on T sub-cell populations, followed by pathway enrichment analysis for selected functional gene signatures (Figure 3E). As a control, we first assessed differences between cells annotated as exhausted versus non-exhausted. Exhausted cells are enriched in the signatures of "exhausted T cells" and "CART dysfunction", as well as "P53 signaling" and "negative regulation of T cell proliferation". Conversely, non-exhausted cells are enriched for the signatures of "T cell proliferation" and "early memory". Analysis of CR versus NR / RL populations reveals similar consistent patterns. Focusing on either effector memory or non-specific CD8+ T cells, NR / RL populations display characteristic features of exhaustion, enhanced CART dysfunction and P53 signaling, and reduced T proliferation and early memory signatures. Comparable analyses performed using cell annotations derived from healthy donors produced similar results (Figure 34D).

[0164] In summary, the single-cell data confirm predictions from the model in a different indication (ALL): infusion products associated with non-sustained responses show elevated frequencies of exhausted cells and effector memory populations display inherent defects in proliferative and functional capacity.

[0165] In additional or alternative embodiments, known single-cell RNA sequencing studies of pre-infusion autologous CD19 CAR-T products (with 4-1BB costimulatory domain and similar to Kymriah®) from two patients with acute lymphoblastic leukemia, one responder and one non-responder, can be used. Cell type labels can be assigned by mapping the expression profiles of 4879 individual cells to T cell subpopulations sorted from healthy donors (Figure 37A). Cells from the two products form clearly distinct clusters in UMAP space. Consistent with the bulk sequencing data, the frequency of both CD8+ and CD4+ central memory cells (the most primitive annotated cell types) can be enriched in CR compared to NR patients (Figure 37B). NR patients showed significant enrichment for CD8+EMRA cells (effector memory cells that re-express CD45RA). This is a differentiated subpopulation with severely reduced proliferative potential, consistent with the model-predicted reduced turnover rates of both memory and effector cells in the NR class.

[0166] Continuing with the alternative embodiment, to investigate the cell-specific differences underlying the response, differential gene expression analysis was performed on matched T cell populations between CR and NR patients, followed by pathway enrichment analysis for selected pathway signatures of exhaustion, proliferation, and inflammation (Figure 37C). Consistent with the bulk sequencing data and model predictions, cell populations from CR patients appeared significantly less exhausted, whereas cells from NR patients appeared exhausted. Cell populations from CR patients were also consistently enriched for T cell proliferation signatures. Although differences in interferon signaling were not significant, inflammatory signaling pathways TNFa and NFκB were elevated in CR cells. Differences between CD8+EM and EMRA subpopulations can also be compared. CD45RA+ cells showed increased exhaustion, decreased proliferation, and decreased inflammatory signaling (i.e., cytotoxicity potential).

[0167] Some implementations leveraged two recently published clinical trials containing single-cell RNA sequencing data of pre-infusion autologous CD19 CAR-T products matched with clinical outcomes to deconvolute the role of cell frequency versus function in mediating response. Bai et al. 2021 reports data on 12 patients with acute lymphoblastic leukemia (ALL) treated with a CD19 CAR-T product similar to Kymriah®, with 5 complete responders (CR), 2 non-responders (NR), and 5 relapsed (RL). Haradhvala et al. report data on 32 patients with large B-cell lymphoma (LBCL) treated with either Kymriah® (n=13) or Yescarta® (n=19). In the Kymriah arm, six patients had a CR and seven had a NR; in the Yescarta arm, eleven patients had a CR, one had a PR, and seven had a NR.

[0168] Examination of the UMAP projections of the three datasets (Kymriah® in ALL, Kymriah® in LBCL, and Yescarta® in LBCL) reveals some separation of response categories in transcriptome space, especially in ALL (Figure 40A,D,G). To assess whether the separation of responses is due to differences in T cell composition, cell type labels were assigned by mapping expression profiles of individual cells to the annotated tumor-infiltrating lymphocyte populations via ProjecTILs. The majority of CD8+ cells in all three datasets are classified as T effector memory (Tem) or T exhausted (Tex), with little to no consistent differences in composition per response category (Figure 44A-C). For example, the frequency of cells annotated as exhausted is higher in the NR / RL category compared to CR in the ALL data (p<0.05, mean 4.4% vs. 8.7%, respectively, Figure 40B,E,H). However, this pattern does not necessarily hold for the LBCL data, and the modest effect sizes may be insufficient to account for differences in clinical outcomes. Using CITESeq antibody tag data provided by Bai et al. 2022, we assigned annotations of early memory (Tmem:CD8+CD45RO-CD27+) and exhausted (CD8+PD1+) cells by immunophenotypes (Fraietta et al.), which have been reported to predict response in CLL. Although annotation of exhausted cells by ProjecTILs and immunophenotypes was concordant (6.7% vs. 5.9% of total cells), cell frequencies did not substantially differ between response categories in ALL (Figure 44D,E).

[0169] To explore cell-intrinsic functions, cells were annotated using a "CAR-T dysfunction" signature, a hallmark of functionally exhausted CAR-T cells with reduced proliferative and cytotoxic capabilities (Good et al.). Visually, the dysfunction signature is distributed across response categories and is not restricted to exhausted regions (Figure 40G,H,I). To explore cell-intrinsic functional differences at deeper resolution, differential gene expression analysis was performed on T sub-cell populations (annotated by both transcriptome and immunophenotype), followed by pathway enrichment analysis on selected gene signatures (Figure 40J). As a control, we first assessed differences between cells annotated as exhausted versus non-exhausted. Exhausted cells were consistently enriched in the CAR-T dysfunction signature across the entire dataset, while the signatures of "exhausted T cells" and "P53 signaling" appear specific for exhausted cells in ALL. Conversely, non-exhausted cells show differential enrichment for "early memory T cell" signatures as well as signatures of cytokine production and inflammatory responses that are characteristic of T cell functional efficacy.

[0170] Comparing cell populations from the CR and NR / PR / RL categories reveals consistent patterns across the data set. Focusing on either effector memory or early memory (CD8+CD45RA-CD27+) subsets, the NR / PR / RL group displays characteristic features of exhaustion. In particular, the CAR-T dysfunction signature is consistently elevated. CR cell populations, conversely, display increased expression of early memory and / or T cell functionality signatures (cytokine production and inflammatory response). That is, memory and effector cell populations from CAR-T products resulting in CR appear more functional or "memory-like," while the same cell populations from the NR / PR / RL category appear more exhausted. Thus, the single-cell data confirms inferences from the model in distinct indications (ALL and LBCL). CAR-T infusion products associated with non-sustained responses display deficits in the proliferative and functional capacities inherent to memory and effector cell populations.

[0171] Cell-intrinsic attributes predictive of CAR-T response can be inferred from the pre-infusion product transcriptome

[0172] If CAR-T response is product-specific rather than host-specific, it can be inferred that differences in the pre-infusion product transcriptome may be predictive of response. Furthermore, a comparison of response classifiers based on cell-specific functions (transcriptome) with cell composition-based response classifiers (T cell phenotype) may help elucidate which product-specific features are more clinically relevant. Although sometimes none of the individual gene signatures evaluated accurately classify the response, bulk RNAseq data can be used to develop a multivariate transcriptome classifier. For example, in one implementation, starting with 30 differentially expressed pathways between CR vs. NR groups (FDR-adjusted p-value <0.05), a logistic regression-based classifier using a genetic algorithm for feature selection can be used (see, e.g., Figure 13).

[0173] The resulting model was able to predictively distinguish CAR-T products from CR and NR patients, with an average accuracy of 83% based on a 60:40 training / testing split (Figure 4A, Figure 41A). We compared this result to two control scenarios: a naive classifier (randomly calling R / NR based on sample frequency in the dataset) and one in which input features are randomly selected from the remaining compendium of the pathway (FDR-adjusted p-values ​​>0.05). The trained model performed significantly better than each of these control cases (p<10 for each). -15 ). The PR samples (not used in training) were called intermediate between CR and NR, which may be a real hypothesis or simply a failure of the classifier for these separate samples (Figure 13).

[0174] To confirm that these findings were replicable, we applied the framework discussed to single-cell data from Bai. Of the 30 pathways identified as significantly different in the CART product transcriptome between CR and NR in CLL, six are coordinately differentially enriched between CR and NR / RL patients in ALL: E2F targets (Hallmark), delayed estrogen response (Hallmark), Myc targets V1 (Hallmark), cell cycle (Reactome), cell cycle mitosis (Reactome), and DNA replication (Reactome) (see Supplementary Dataset A). This degree of overlap was highly significant (p=3×10 -8 , binomial test), indicating a shared biological mechanism associated with robust efficacy between these diseases. Thus, the same 30 pathways can be used as input features to a logistic regression classifier (see, for example, Figures 7C and 7D) using the ssGSEA scores from the pseudobulk analysis. Again, the resulting accuracy was compared to a random pathway control case and a naive classifier. The trained classifier significantly outperformed these cases with a mean accuracy of 75% (both p<10 -8), which demonstrates the cross-dataset predictive accuracy. Note, however, that in other implementations, more or less than 30 pathways can be used as input features.

[0175] In summary, this indicates that response is primarily predetermined by the cellular composition of the CAR-T product, and that response can be accurately predicted from the pre-infusion transcriptome. For example, in one embodiment, response to Kymriah® in two different indications (CLL and ALL) is at least partially predetermined by the cellular composition of the CAR-T product, because response can be accurately predicted from the pre-infusion transcriptome, and because the predicted transcriptional signature of response is shared across the two disease indications.

[0176] In an additional or alternative implementation, an alternative logistic regression-based classifier using a genetic algorithm for feature selection can be used, starting with the 857 differentially expressed pathways between the CR vs. NR groups (see, e.g., Figures 7A, 7B, 39).

[0177] Continuing with additional or alternative implementations, the resulting model was able to predictively distinguish CAR-T products from CR versus NR patients with an average cross-validation accuracy of 94% (Figure 38A). This classifier can be trained on CR versus NR groups, allowing us to see how the model predicts PR samples and PR TD Samples were queried to classify (Figure 38B), and were inconsistently annotated as essentially intermediate between the R and NR classes, which recapitulated the clinical annotations.

[0178] In some implementations, classifiers were trained and evaluated using the frequencies of early memory (CD8+CD45RO-CD27+) and exhausted (CD8+PD1+LAG3+) cells as reported by Fraietta et al. (Figure 44D). The resulting accuracies (80% and 83%, respectively) are significantly better than chance, but are less accurate than the accuracies achieved using functional transcriptomes (p<10, respectively). -15 and p=6x10 -11 Thus, gene signature panels reveal clinical functionality to a degree not evident from immunotyping, suggesting that transcriptomics may provide more value as a CAR-T product characterization assay than current best practice flow cytometry panels.

[0179] To assess whether these findings transfer across datasets and indications, the workflow was applied to pseudo-bulk single-cell data from Bai et al. 2022 (Kymriah® in ALL) and Haradhvala et al. (Kymriah® and Yescarta® in LBCL). For the Bai et al. 2022 data (Kymriah® in ALL), the accuracy of classifying CR vs. NR / RL groups using a 28-gene signature panel was compared to a bivariate classifier trained using the frequencies of early memory (CD8+CD45RO-CD27+) and exhaustion (CD8+PD1+) immunophenotypes calculated from CITEseq antibody tags (Figure 44D). The median accuracy of the transcriptomic classifier was 80%, lower than previously achieved (as expected), but better than that achieved by T cell immunophenotyping (47%, p<10 -15, Fig. 41B). Similarly, predictive accuracy was assessed separately for Kymriah® and Yescarta® using the LBCL data from Haradhvala et al. As immunophenotype data were not provided, the transcriptomic classifier was compared to a bivariate classifier based on the frequency of T effector memory (Tem) and exhausted cells (Tex) as estimated from annotations in ProjecTILs (Andreatta et al.) (Fig. 44B,C). The median predictive accuracy of the transcriptomic classifier was 80% and 71% for Kymriah® and Yescarta®, respectively, outperforming T cell phenotype-based classification in both cases (60% and 67%, p<10 -15 , Fig. 41C and D). As an additional control, the classifier was seeded with a "random" pathway by extracting from a compendium of gene signatures that were not differentially expressed between CR vs. NR groups in the CLL data (FDR adjusted p-value>0.05, see Example 1 and Fig. 45). The resulting accuracies were slightly better or indistinguishable from chance ("null" model), and all significantly less accurate than predictions arising from the 28-gene signature panel.

[0180] To condense the inner workings of the transcriptome classifier into interpretable patterns, a CAR-T response scorecard was created (Figure 41E). It summarizes the GSEA for the 28 selected pathways and frequencies encompassed in the 2,500 trained models across each of the four datasets. As expected, there is variance in the directionality and statistical significance of the signatures between the datasets. These represent different diseases, CAR-T products, and platforms, and the data were generated for each independent group. However, the overlap is greater than would be expected by chance (p<10 for all). -5(See Example 1.) For example, the Yescarta®-LBCL scorecard is visually different from the three Kymriah® scorecards, and the resulting model predictions are less accurate, suggesting that the biology underlying the response between the two products still clearly overlaps.

[0181] In summary, response to two distinct CD19 CAR-T therapy products (Kymriah® and Yescarta®) in three indications (CLL, ALL, and LBCL) is at least partially predetermined by the functional attributes of the CAR-T infusion product. These functional attributes are shared to various degrees across the four datasets and are manifested via gene signatures that are not fully revealed from T cell immunophenotyping.

[0182] Example 3

[0183] Accounting for interpatient variability in Kymriah(R) pharmacokinetics

[0184] The pharmacokinetics of Kymriah® and other CAR-T products tested in clinical trials show high interpatient variability (e.g., with AUCs spanning three orders of magnitude). We tested whether the mechanism-based model described herein explains this variability. In particular, a mixed triple patient archetype (CR / PR / NR), combined with reported variability in administered dose and initial tumor burden, is sufficient to quantitatively explain the observed variance in exposure.

[0185] Simulations of the CR / PR / NR pharmacokinetic profiles were overlaid with a Kymriah simulation. Although these were different patient populations (CLL vs. B-ALL), pharmacokinetics are preserved between these two indications. Visually, the CR / PR / NR profiles roughly correspond to the top quartile, median, and bottom 5% of exposure (Figure 5A). Thus, while the CR / PR / NR population archetypes cover most of the pharmacokinetic variability, they do not fully explain the variability of individual patients because they fit the population mean.

[0186] The effect of dose and tumor burden variation was evaluated using a virtual population approach: a virtual population (n=1000) was defined by Monte Carlo sampling across the entire parameter set, while doses and tumor burdens within the reported ranges were randomized either alone or in combination by log-uniform sampling.

[0187] The simulated area under exposure (AUC) for these hypothetical populations spans the inter-individual variability of Kymriah (10 1 ~10 4 The variance in either dose or tumor volume is sufficient to cover and nearly match the variance in reported exposures within the CR / PR / NR populations. That is, the model was fitted to population-mean data assuming fixed tumor volume and dose, but relaxing either of these inputs is sufficient to explain the reported variance. Similar results arise by considering Cmax (Figure 5C). We used grid simulations to evaluate how tumor volume and dose drive exposure and tumor response (Figure 14), revealing a nonlinear relationship that may contribute to clinical variance. Tumor AUC was found to increase with initial tumor volume and decrease with initial CAR-T dose (Figure 14). Cmax shows a more complex relationship, increasing with initial CAR-T dose but peaking for intermediate tumor volumes. This nonlinear interaction between tumor volume and CAR-T dose may contribute to the variability observed clinically.

[0188] Predictive response covariates: Cmax and tumor burden

[0189] We investigated whether virtual populations could predict a priori the reported statistical relationship between cell proliferation and tumor burden on clinical response. Analysis of response covariates to Yescarta® in large B-cell lymphoma (LCBCL) identified the ratio of CAR-T proliferation to initial tumor burden (i.e., Cmax / B0) as the strongest correlate of durable response. The same results were reported for overall survival in B-ALL, indicating that this was a preserved feature across indications. The pharmacokinetics and median population variance of Yescarta® are similar to Kymriah® (Figure 35).

[0190] Focusing on the hypothetical CR population, responses were defined by the B cell AUC, 10 4 The covariate covariate ratio (Cmax) was set at cells*days / μL (minimum observed for the hypothetical PR population). Logistic regression models relating response to initial tumor burden (B0), Cmax, or the ratio as predictors were used (Figure 5D-F). Equivalent logistic curves from Yescarta® were digitized and overlaid by normalizing the x-axis. The results are in qualitative agreement with the clinical data in that these covariates predict response.

[0191] To assess whether these predictions emerge directly from the model structure or require model training, a "control" virtual population can be defined by random sampling of parameter space (N=1000). This control population did not replicate the same findings, emphasizing the need for adequate training data to make accurate predictions.

[0192] Implications of dose response: Multiple myeloma patients treated with Abecma® (BCMA-CART)

[0193] To better understand the relationship between dose, Cmax and tumor response, the modeling framework can be applied to a Phase I / II dose escalation study of Abecma® (BB2121, idecbutagenvir ucel), a BCMA-targeted CAR-T approved for the treatment of multiple myeloma. Using particle swarm optimization, model parameters characterizing pharmacokinetics and tumor dynamics can be estimated (Figure 6A,B). Although the parameters are indistinguishable (see Table 3 for parameters), both are obtained with good accuracy (Figure 15), and the simulation reproduces the relationship between Cmax / B0 and tumor response identified in Figure 5F for Kymriah® and Yescarta® (Figure 36). [Table 3]

[0194] Simulations provide insight into the impact of CAR-T dose on T cell population dynamics (Figure 6C-E). The lowest dose (50 million cells) resulted in no tumor reduction, a predominance of exhausted T cells, and a gradual loss of memory cells. The highest dose, where the greatest degree of tumor reduction was observed, resulted in the opposite response, with minimal exhausted cells and a high proportion of memory cells. This is similar to the changes in T cell composition following acute versus chronic infection, providing a mechanistic basis for the covariates identified above. That is, if the Cmax:tumor burden ratio is insufficient, either due to low dose or proliferative capacity, the infused CAR-T population will become exhausted before clearing the tumor.

[0195] To evaluate the predictability of the model, the simulations were compared to data from a phase 2 study where patients were treated with doses of 150, 300, and 450 million cells and tumor dynamics (BCMA levels) were monitored for one year (Figure 6F,G). Both PK and tumor dynamics are moderately underpredicted, but profiles are obtained with reasonable accuracy. That is, the phase 2 data (150-450 million cell doses) correspond to the simulated 150-450 million cell doses, with similar dynamics. This is noteworthy for tumor dynamics, considering that the model was trained on data up to 2 months, while predictions are extrapolated up to one year.

[0196] Some known clinical trials have noted that robust cell proliferation after CAR-T infusion is a prerequisite for clinical efficacy; the inability to predictively control CAR-T pharmacology may limit their clinical utility. Mechanism-based mathematical models represent the way forward. When trained with appropriate data sets, such models allow both inference of the underlying biological principles governing the response and the ability to generate quantitative predictions, ultimately guiding therapeutic design. We hypothesized that the principles governing T-cell dynamics during infection also govern CAR-T pharmacology, and tested this using a conceptually simple mathematical model of T-cell regulatory control, based on an analogy to a toggle switch. The model was trained using available clinical pharmacokinetic and tumor dynamics data that have yielded biological insights and clinical predictions, some of which have been confirmed.

[0197] First, CAR-T proliferation, persistence, and anti-tumor responses may be driven by the cell-intrinsic turnover rate of memory T cell populations and the cytotoxic potential of effectors. Using bulk gene expression data, enrichment of memory cell signatures, enhanced proliferative and inflammatory signaling, and lack of exhaustion markers in pre-infusion CAR-T products were found to correlate with response. Single-cell sequencing data from two additional disease indications and additional CD19 CART products confirmed that these differences between CR and NR archetypes are intrinsic to the function, not the frequency, of memory cells in infused products. CAR-T products that result in non-sustained responses display deficits in proliferative and functional capabilities that are characteristic of T cell exhaustion and terminal differentiation, even within immunophenotypically indistinguishable memory and effector cell populations. These functional differences were inferred from the models and / or methods described herein and confirmed via expression of a "CAR-T dysfunction" gene signature. CAR-T expansion (e.g., Cmax) following infusion can represent an in vivo readout of memory T cell expansion potential.

[0198] Response categories could be accurately predicted using the pre-infusion product transcriptome in three indications (CLL, ALL, and LBCL) and two CD19-targeted products (Kymriah® and Yescarta®). Furthermore, transcriptomic profiles reveal functional attributes not revealed from standard immunophenotyping, and these attributes are shared to different degrees among the data sets considered. In some implementations, memory / exhaustion phenotypes identified as predictive of response in CLL did not transfer to ALL, whereas gene signature panels did. Furthermore, if the pre-infusion product transcriptome predicts response, this may suggest that these pharmacological archetypes are unique to the infusion product and thus CAR-T efficacy could be improved through product design.

[0199] A simple and easily implemented molecular signature of an effective (CR-like) CAR-T product may be desirable to guide optimization studies. The CAR-T response scorecard (Figure 41E) reveals transcriptional features shared to various degrees among the four datasets. While statistically significant similarities exist, different molecular mechanisms appear to cooperatively mediate clinical outcomes among the three datasets, especially between two products (Yescarta® vs. Kymriah®). In some implementations, this scorecard may serve as a useful tool for CAR-T product optimization. However, the pathways selected are derived from the first dataset considered (Kymriah® in CLL). That is, this is a visual representation of the workflow, rather than a comprehensive map of features consistently shared across datasets. Furthermore, the classifier was trained on single-sample GSEA scores, while the different projections represent group-level differential pathway enrichment. This compression loses information about variance within sample groups that may be important for multivariate classification. That is, the algorithm and method selects signatures that do not vary significantly at the group level (i.e., intermediate scores (e.g., between 4502 and 4504 in Figure 41E), high frequency (e.g., at least 70 in Figure 41E)), but may still contain information. Furthermore, many of these signatures are biologically meaningful (e.g., JAK / STAT signaling, exhausted T cells), while others are less meaningful (e.g., EMT, xenobiotic metabolism). This is an expected outcome of comparison of gene lists to pathway databases, i.e., many of these signatures are manually curated with inconsistent degrees of validation, and gene lists have overlaps between biological processes.

[0200] Pharmacological archetypes, combined with variability in CAR-T cell dose and initial tumor burden, explain the patient-to-patient variability in exposure observed in Kymriah® clinical trials. The ratio of CAR-T proliferation (Cmax) to initial tumor burden (B0) quantifies whether the infused cell product is able to eliminate tumors, a novel prediction from models observed in known studies of Yescarta®. Mechanistically, cell doses insufficient to eliminate tumors result in CAR-T exhaustion, while sufficient doses lead to regeneration of memory populations.

[0201] It may be desirable to control clinical variability in cell dose and initial tumor burden. Cell dose is predefined resulting from any manufacturing process, and initial tumor burden, residual cancer cells after lymphodepleting chemotherapy, are both highly variable between patients. In some implementations, given a consistent quality CAR-T product (e.g., one that exhibits a CR class transcriptional signature), model simulations can be used to define patient-specific doses based on tumor burden (B-cell counts) to achieve an optimal balance between maximizing tumor reduction and minimizing Cmax-related toxicity (Figure 14). Figure 8 provides an overview of an exemplary workflow using model simulations to optimize treatment.

[0202] It should be noted that although variation in CAR-T dose and tumor burden may be sufficient to explain the observed variance in exposure, additional host-specific factors may be involved (e.g., response to lymphodepletion).

[0203] The CR vs. NR archetype may be a product-specific property. In some cases, product-specific may mean that the clinical response is fully predictable by the properties of the infused product. These properties (e.g., memory vs. exhaustion phenotype) may then be predetermined by the immunological state of the patient, i.e., host-specific properties. Note that some of the model parameters may incorporate some aspects of two or more properties. Cytotoxicity performance (TK50), for example, appears to be a cell-specific parameter. However, it brackets multiple cellular processes: CAR and antigen expression, CAR-antigen binding kinetics, intracellular signaling, and involvement of cytotoxic mechanisms. These processes are also regulated by systemic cytokines and cell-cell interactions. Similar cases may be made for other model parameters. Thus, while the variation in CAR-T dose and tumor burden is sufficient to explain the observed variance in exposure, the inclusion of additional host-specific factors can extend the usefulness of the model. Tumor-specific signaling and response to lymphocyte depletion are two examples. Both have been shown to mediate CAR-T proliferation and tumor responses, but this is because cytokine-mediated interactions between CAR-T, host T cells, and tumors likely mediate cell-intrinsic differences.

[0204] The model described addresses four necessary characteristics and has at least some variances compared to known predator-prey structures to incorporate basic T cell biology. Borrowing from the stem cell field, each T memory (T M ) cell division, tumor antigen (B A ) is encoded as a fate choice between self-renewal and differentiation driven by the differentiation and proliferation of CAR-T. That is, differentiation and proliferation of CAR-T occurs on a scale that depletes the pool of memory cells. In some implementations, effector cells (T E) do not self-renew but rather undergo a fixed number (N) of divisions. This addresses the unlimited CAR-T proliferation capacity seen in some known predator-prey models. Accounting for self-renewal versus memory cell differentiation also provides a mechanism by which chronic antigen stimulation (or alternatively, insufficient CAR-T dose relative to tumor size) drives exhaustion. If tumor cells cannot be sufficiently cleared to reduce systemic antigen load below a defined threshold (B50), Tm cells will continually differentiate until the pool of long-term memory cells is exhausted.

[0205] Also included is an exhausted T cell state, which in some implementations allows for differentiation between CAR-T pharmacokinetics and cytotoxic function, particularly in partial responders and non-responders (discussed in more detail below).

[0206] Example 5

[0207] Model structure evaluation

[0208] To systematically evaluate the topology of the model, we defined a series of variants with alternative T cell population structures, for example: 1.T E Population only (effector state) 2.T M and T E (Memory and effector status), T X none 3.T E and T X (Effector and exhaustion states) 4.T M , T E , and T X However, there is no differentiation from effector to memory (r M =0) 5. Add additional naive (T N ) Include state

[0209] FIG. 28 illustrates the structure of these models.

[0210] Note that the original (full) model describes four subpopulations that are modulated by antigen exposure via ODEs, e.g.:

number

[0211] For variant model 1, a single effector compartment is described, where proliferation / self-renewal is antigen driven, for example as follows:

number

[0212] For the modified model 2, the full model is adopted, but k is set to prevent exhausted T cells from being generated. ex =0 is set.

[0213] For variant 3, a version of variant 1 is adopted, in which effectors both proliferate / self-renew and migrate towards exhausted cells in an antigen-dependent manner.

number

[0214] For variant model 4, the original set of model equations is adopted, but with the addition of r such that memory cells cannot arise from effectors. M = 0. This variant therefore evaluates whether a teleonomic dedifferentiation response is required, given the limited number of T cell states that are considered for parallelization in some implementations.

[0215] For the variant model 5, following the standard hierarchy of T cell differentiation, the naive T cell compartment (T N These cells proliferate and become memory T cells in an antigen-dependent manner, for example, via the following scheme: N Differentiate into cells.

number

[0216] With the differentiation of TN cells into the memory cell compartment, the balance equation for TM now becomes, for example:

number

[0217] This introduces five additional free parameters into the model: μ N : Proliferation rate of naive T cells f N : Probability of self-renewal of naive T cells k N : The Hill index relates antigen exposure to naive T cell proliferation d N : Death rate of naive T cells Proportion_T N : Proportion of CAR-T cell dose within the naive T cell compartment

[0218] The fit of the resulting model to the CR / PR / NR population from Fraietta et al. for five structural variants compared to the full model is shown in Figure 29. Note that all variant models are similar in their ability to describe tumor dynamics, but differ in their ability to capture pharmacokinetics.

[0219] Based on the MSE, the original model and the variant 5 (T N A classification (including compartments) best describes the data.

[0220] Variation 2 lacks the fatigue state, but T M and T E (and reversible transfer), but this version does not adequately capture the pharmacokinetics of the NR population. Variant 4 includes all three T cell populations, but does not include the T E From T MVariants 1 and 3 lack a retrograde transition to TM and capture the PK data reasonably well, but fail to capture the peak broadening of the CR and PR populations. Variants 1 and 3 lack a TM compartment and both do not describe any of the population pharmacokinetics, even qualitatively.

[0221] The Akaike Information Criterion was used to rank each model based on the following fitting error (MSE) versus complexity (number of parameters).

number

[0222] where n=number of measurements, k=free parameters, and MSE=mean squared error. The results are summarized in Table 4 below. Note that AIC was originally developed to rank multivariate linear regression models, not nonlinear ODEs, and thus prioritizes restricting free parameters over goodness of fit. Thus, the results are informative rather than discriminatory, and should be balanced against more subjective fitting assessments. [Table 4]

[0223] Based on the MSE, variant 5 (T N The variant 5 (involving both cellular compartments) is the most accurate and outperforms the original model (Variant 0). Examination of the PK curves in Figure 28 reveals that this improvement is due to capturing the last time point of the NR profile (12 months), which is an increase from the previous one (6 months). This may be an artifact of the data (population average) rather than a real phenomenon, suggesting that the model is overfitted. Note that variant 5 includes five additional parameters compared to the original model, and the resulting AIC is more than double, suggesting that this added complexity adds little value.

[0224] Based on the relative AIC, both variants 1 and 2 are superior to the original (full) model. Despite the poorer fit (MSE), the AIC calculations outweigh this with the reduction in free parameters. However, based on the curve evaluation, variants 1 and 2 perform obviously worse than the original model, since they are unable to capture the PK profile of NR (variant 2) or any of the PK profiles (variant 1). Therefore, taking into account the fitting error of the PK curve, the model complexity, and the subjective evaluation, it may be said that the original model is superior to each structural variant model.

[0225] In some cases, MSE considers data points to be of equal value, but subjectively, some data points and readouts may be more or less significant. Given that tumor kinetic profiles appear to be very similar between variant models, tumor fit may not necessarily be an important model selection criterion. In some cases, metrics of CAR-T exposure (Cmax and AUC) can be used to predict clinical efficacy. For example, and as shown below, variant models can be evaluated by the MSE of Cmax (log10-cells) and AUC (log10-cells.days), in addition to the MSE of the data and the AIC corrected for sample size (Table 5, bold indicates models ranked higher by metrics). [Table 5]

[0226] Based on the MSE of this data, variant 5 (T NThe variant 5 (inclusive of the cellular compartment) is the most accurate and outperforms the original model (Variant 0). Examination of the PK curves in Figure 29 reveals that this improvement is due to acquiring the last time point of the NR profile (12 months), which is an increase from the previous one (6 months). This may be an artifact of the data (population average) rather than a real phenomenon, suggesting that the model is overfitted. Note that variant 5 includes five additional parameters compared to the original model, and the resulting AIC more than doubles from 57 to 140, indicating that this added complexity adds little value. To assess the generalizability of the model, two preclinical data sets with dose-response data for PK and tumor kinetics were fitted: NALM xenografts treated with CD19-CAR-T (Figure 30), and MM1.S xenografts treated with BCMA-CAR-T (Figure 31). In both cases, the model described the data with good accuracy.

[0227] FIG. 16A shows a flowchart of a method 1600 for training a machine learning classifier (e.g., machine learning classifier 2304 of FIG. 23) according to one embodiment. In some implementations, the method 1600 may be performed by a processor (e.g., processor 2301). For example, instructions that cause the processor 2301 to execute the method 1600 may be stored in memory 2302 of FIG. 23. In some implementations, by performing steps such as filtering gene signatures and / or grouping filtered gene signatures, a feature vector may be determined more quickly and / or with reduced computational load compared to known methods. A machine learning classifier may then be trained more quickly and / or with reduced computational load using the feature vector.

[0228] At 1601, gene expression data for a plurality of cells is received (e.g., from memory 2302 of computing device 2300, from a computing device communicatively coupled to computing device 2300 via a network, etc.). In some implementations, the cells are immune cells. In some implementations, the cells are immune cells and the immune cells are chimeric antigen receptor expressing T cells (CAR-T cells). In some implementations, the cells are immune cells and the immune cells are chimeric antigen receptor expressing T cells (CAR-T cells) and the immune cells include at least one of autologous or allogeneic cells.

[0229] In 1602, differential gene expression analysis is performed based on the gene expression data to generate differential gene expression data (e.g., differentially expressed genes). In 1603, gene signature enrichment per sample is estimated for a plurality of identified biological pathways based on the differential gene expression data. In 1604, gene signatures having differential enrichment patterns are filtered based on a predetermined threshold between a group of responders and a group of non-responders to the treatment. In some implementations, the filtering in 1604 is based on a statistically significant difference between the group of responders and the group of non-responders. In 1605, the filtered gene signatures are grouped into a plurality of groups based on pairwise correlation between gene signature enrichment scores. In 1606, randomly selecting gene signatures from each group from the plurality of groups to define a set of gene signatures is repeated a predetermined number of times to define a plurality of sets of gene signatures. At 1607, the plurality of sets of gene signatures are provided as input to a feature selection algorithm to determine a feature vector including a reduced set of gene signatures identified from the plurality of sets of gene signatures. In some implementations, the feature selection algorithm is a genetic algorithm. At 1608, a machine learning classifier (e.g., machine learning classifier 2304 of FIG. 23) is trained using the feature vector. In some implementations, the machine learning classifier is at least one of a logistic regression-based classifier, a multinomial logistic regression, a decision tree, a perceptron, a support vector machine, a K-nearest neighbors classifier, a naive Bayes classifier, or a random forest.

[0230] FIG. 16B shows a flowchart of a method 1650 for training a machine learning classifier (e.g., machine learning classifier 2304 of FIG. 23) according to one embodiment. In some implementations, the method 1650 may be performed by a processor (e.g., processor 2301). For example, instructions that cause the processor 2301 to execute the method 1650 may be stored in memory 2302 of FIG. 23. In some implementations, by performing a step of selecting identified biological pathways that are statistically significant, a feature vector may be determined more quickly and / or with reduced computational load compared to known methods. A machine learning classifier may then be trained more quickly and / or with reduced computational load using the feature vector.

[0231] At 1651, gene expression data for a plurality of cells is received (e.g., from memory 2302 of computing device 2300, from a computing device communicatively coupled to computing device 2300 via a network, etc.). In some implementations, the cells are immune cells. In some implementations, the cells are immune cells and the immune cells are chimeric antigen receptor expressing T cells (CAR-T cells). In some implementations, the cells are immune cells and the immune cells are chimeric antigen receptor expressing T cells (CAR-T cells), and the immune cells include at least one of autologous or allogeneic cells.

[0232] At 1652, differential gene expression analysis is performed based on the gene expression data to generate differential gene expression data (e.g., differentially expressed genes). At 1653, per-sample gene signature enrichment is estimated for the plurality of identified biological pathways based on the differential gene expression data. At 1654, a set of identified biological pathways from the plurality of identified biological pathways that are statistically significant (e.g., top 30 most statistically significant) is determined based on the per-sample gene signature enrichment. At 1655, the set of identified biological pathways is provided as an input to a feature selection algorithm to determine a feature vector. At 1608, a machine learning classifier (e.g., machine learning classifier 2304 of FIG. 23) is trained using the feature vector. In some implementations, the machine learning classifier is at least one of a logistic regression-based classifier, a multinomial logistic regression, a decision tree, a perceptron, a support vector machine, a K-nearest neighbors classifier, a naive Bayes classifier, or a random forest.

[0233] FIG. 17 shows a flowchart of a method 1700 for predicting a patient's clinical outcome in response to a cell therapy treatment, according to one embodiment. In some implementations, the method 1700 may include using a machine learning classifier trained using the method 1600. In some implementations, the method 1700 may be performed by a processor (e.g., the processor 2301). For example, instructions causing the processor 2301 to perform the method 1700 may be stored in the memory 2302 of FIG. 23. In some implementations, the predicted clinical outcome may be generated more quickly and / or with reduced computational burden compared to known methods using the method 1700. Furthermore, the predicted clinical outcome may be used to determine corrective actions (e.g., administering a treatment, refraining from administering a treatment, etc.), which may be performed (or not) by a computing device and / or a person.

[0234] At 1701, gene expression data is received for a plurality of cells to be administered to a patient. In some implementations, the cells are immune cells. In some implementations, the cells are immune cells, and the immune cells are T cells expressing a chimeric antigen receptor (CAR-T cells). In some implementations, the cells are immune cells, and the immune cells are T cells expressing a chimeric antigen receptor (CAR-T cells), and the immune cells are selected from the group consisting of autologous cells or allogeneic cells. At 1702, gene signature enrichment per sample is estimated for a plurality of identified biological pathways. At 1703, the differentially enriched gene signatures are filtered between responders and non-responders to the cell therapy. In some implementations, the filtering at 1703 is based on a statistically significant difference between a group of responders and a group of non-responders. At 1704, the filtered gene signatures are provided as input to a machine learning classifier. In some implementations, the machine learning classifier is at least one of a logistic regression-based classifier, a multinomial logistic regression, a decision tree, a perceptron, a support vector machine, a K-nearest neighbor, a naive Bayes classifier, or a random forest. In some implementations, the machine learning classifier is trained according to method 1600. At 1705, a predicted clinical outcome is generated. In some implementations, the predicted clinical outcome includes at least one of (i) a complete response, a non-response, or a partial response, (ii) a reduction in tumor size or tumor burden to a predetermined threshold, or (iii) cytokine release syndrome (CRS) toxicity.

[0235] In some implementations, method 1700 further includes comparing predicted clinical outcomes for different cell populations, cell characteristics, and / or parameters for cell growth and / or manufacturing to identify those associated with a more favorable clinical outcome.

[0236] FIG. 18 shows a flowchart of a method 1800 for generating a predicted clinical outcome, according to one embodiment. In some implementations, the method 1800 may be performed by a processor (e.g., processor 2301). For example, instructions causing the processor 2301 to execute the method 1800 may be stored in memory 2302 of FIG. 23. In some implementations, the predicted clinical outcome may be generated more quickly and / or with reduced computational burden as compared to known methods using the method 1800. Additionally, the predicted clinical outcome may be used to determine corrective actions (e.g., administering a treatment, refraining from administering a treatment, etc.), which may be performed (or not) by a computing device and / or a person.

[0237] At 1801, a cell population is generated. In some implementations, the cells are immune cells. In some implementations, the cells are immune cells, and the immune cells are T cells expressing a chimeric antigen receptor (CAR-T cells). In some implementations, the cells are immune cells, and the immune cells are T cells expressing a chimeric antigen receptor (CAR-T cells), and the immune cells are selected from the group consisting of autologous cells or allogeneic cells. At 1802, gene expression data of the cell population is received. At 1803, the gene expression data is analyzed to identify a set of gene signatures. At 1804, the set of gene signatures is provided as input to a machine learning classifier (e.g., machine learning classifier 2304 of FIG. 23). At 1805, a predicted clinical outcome is generated (e.g., by the machine learning classifier). In some implementations, the machine learning classifier is at least one of a logistic regression-based classifier, a multinomial logistic regression, a decision tree, a perceptron, a support vector machine, a K-nearest neighbors, a naive Bayes classifier, or a random forest. In some implementations, the machine learning classifier is trained according to method 1600. In some implementations, the predicted clinical outcome includes at least one of (i) a complete response, a non-response, or a partial response, (ii) a reduction in tumor size or tumor burden to a predetermined threshold, or (iii) cytokine release syndrome (CRS) toxicity.

[0238] In some implementations, method 1800 further includes comparing predicted clinical outcomes for different cell populations, cell characteristics, and / or parameters for cell growth and / or manufacturing to identify those associated with a more favorable clinical outcome.

[0239] In some implementations, the method 1800 further includes generating a reduced set of gene signatures from the set of identified gene signatures. In some implementations, analyzing the gene expression data includes estimating gene signature enrichment per sample for the plurality of identified biological pathways based on the differential gene expression data. In some implementations, analyzing the gene expression data includes filtering gene signatures having differential enrichment patterns based on a predetermined threshold between groups of responders and non-responders to the treatment. In some implementations, analyzing the gene expression data includes grouping the filtered gene signatures into a plurality of groups based on pairwise correlations between gene signature enrichment scores. In some implementations, analyzing the gene expression data includes repeatedly performing a predetermined number of times to determine a plurality of sets of gene signatures, randomly selecting a gene signature from each group from the plurality of groups to determine the set of gene signatures from the plurality of sets of gene signatures. In some implementations, analyzing the gene expression data includes providing a plurality of sets of gene signatures as input to a feature selection algorithm to define a feature vector comprising a reduced set of gene signatures identified from the plurality of sets of gene signatures.

[0240] FIG. 19 shows a flowchart of a method 1900 for predicting a patient's clinical outcome in response to a cell therapy, according to one embodiment. In some implementations, the method 1900 may be performed by a processor (e.g., processor 2301). For example, instructions causing the processor 2301 to execute the method 1900 may be stored in memory 2302 of FIG. 23. In some implementations, the predicted clinical outcome may be used to determine whether and / or how the cell therapy should be administered. Such insights may be useful, for example, to refrain from administering the cell therapy to patients who are likely not to lead to a desired outcome, or to trigger the administration of the cell therapy to patients who are likely to lead to a desired outcome. Furthermore, the use of mechanism-based kinetic models may allow predictions (e.g., pharmacokinetic responses, pharmacodynamic responses, clinical outcomes, etc.) to be arrived at much more quickly than a human (e.g., given a large amount of data to consider).

[0241] In 1901, one or more characteristics of a patient's tumor burden and a cell therapy are measured. In some implementations, the one or more characteristics of the cell therapy include an initial cell dose and / or a cell subpopulation characteristic. In some implementations, the cell therapy includes an immune cell therapy. In some implementations, the cell therapy includes an immune cell therapy, and the immune cell therapy includes a CAR-T cell therapy. In some implementations, the cell therapy includes an immune cell therapy including an immune cell therapy, an immune cell therapy including a CAR-T cell therapy, and at least one of an autologous immune cell therapy or an allogeneic immune cell therapy. In some implementations, the tumor burden is determined by at least one of a complete blood count, flow cytometry, a radiology (PET / CT) scan, a measurement of circulating tumor-DNA, or an indication-specific circulating tumor biomarker. In 1902, the measured tumor burden and one or more characteristics of the cell therapy are provided as inputs to a mechanism-based kinetic model (e.g., kinetic model 2305 of FIG. 23). In 1903, the mechanism-based kinetic model is used to predict a pharmacokinetic and / or pharmacodynamic response of the cell therapy in the patient. In some implementations, the mechanism-based kinetic model predicts a pharmacokinetic and / or pharmacodynamic response of the cell therapy in the patient. In 1904, a clinical outcome of the patient is predicted in response to the cell therapy.

[0242] FIG. 20 shows a flowchart of a method 2000 for administering cell therapy to a patient, according to one embodiment. In some implementations, the method 2000 may be performed by a processor (e.g., processor 2301). For example, instructions causing the processor 2301 to execute the method 2000 may be stored in memory 2302 of FIG. 23. In some implementations, the predicted clinical outcome may be used to determine whether and / or how the cell therapy should be administered. Such insight may be useful, for example, to refrain from administering cell therapy to patients who are not likely to lead to a desired outcome, to administer cell therapy to patients who are likely to lead to a desired outcome, and / or to determine the appropriate dosage to be administered for a given patient. Additionally, the use of mechanism-based kinetic models may allow predictions (e.g., clinical outcomes, dosages, etc.) to be arrived at much more quickly than a human (e.g., given a large amount of data to consider).

[0243] In 2001, one or more characteristics of the patient's tumor burden and the cell therapy are measured. In some implementations, the one or more characteristics of the cell therapy include an initial cell dose and / or a cell subpopulation characteristic. In some implementations, the cell therapy includes an immune cell therapy. In some implementations, the cell therapy includes an immune cell therapy, the immune cell therapy includes a CAR-T cell therapy. In some implementations, the cell therapy includes an immune cell therapy, including immune cell therapy, including CAR-T cell therapy, and at least one of an autologous immune cell therapy or an allogeneic immune cell therapy. In some implementations, the tumor burden is determined by at least one of a complete blood count, flow cytometry, a radiology (PET / CT) scan, a measurement of circulating tumor-DNA, or an indication-specific circulating tumor biomarker.

[0244] At 2002, the measured tumor mass and one or more characteristics of the cell therapy are provided as inputs to a mechanism-based kinetic model (e.g., kinetic model 2305 of FIG. 23). In some implementations, the one or more characteristics of the cell therapy include an initial cell dose and / or characteristics of a cell subpopulation. In some implementations, the mechanism-based kinetic model predicts a pharmacokinetic and / or pharmacodynamic response of the cell therapy in the patient. At 2003, the mechanism-based kinetic model is used to predict a clinical outcome for the patient. At 2004, the cell therapy is administered to the patient, and the dosage administered to the patient is determined by the mechanism-based kinetic model. In some implementations, the cell therapy is triggered for administration to the patient by the processor sending a signal indicating that the cell therapy is to be administered to the patient (e.g., a display, a speaker, a different computing device, etc.). In some implementations, the cell therapy may then be administered to the patient (e.g., by an intermediate professional, by a robot, etc.). In some implementations, the dosage administered to the patient is determined by a mechanism-based kinetic model based on at least the patient's tumor burden and the Cmax of the cell therapy.

[0245] FIG. 21 shows a flowchart of a method 2100 for generating a cell therapy product, according to one embodiment. In some implementations, method 2100 may be performed by a processor (e.g., processor 2301). For example, instructions causing processor 2301 to perform method 2000 may be stored in memory 2302 of FIG. 23. In some implementations, by performing method 2100, a patient-specific cell dosage may be used to prepare a cell therapy product with a patient-specific cell dosage. The patient may then be administered the cell therapy product with a cell dosage specified for that patient (e.g., not under-obtained or not over-obtained).

[0246] At 2101, a cell population having at least one subpopulation is provided. In some implementations, the one or more characteristics of the cell therapy include an initial cell dose and / or a cell subpopulation characteristic. In some implementations, the cell therapy includes immune cell therapy. In some implementations, the cell therapy includes immune cell therapy, and the immune cell therapy includes CAR-T cell therapy. In some implementations, the cell therapy includes immune cell therapy, including immune cell therapy, including CAR-T cell therapy, and at least one of autologous immune cell therapy or allogeneic immune cell therapy.

[0247] At 2102, the patient's tumor burden is measured. In some implementations, the tumor burden is determined by at least one of a complete blood count, flow cytometry, radiology (PET / CT) scan, measurement of circulating tumor-DNA, or indication-specific circulating tumor biomarkers. At 2103, a predicted proliferation potential (Cmax) of the cell therapy product is determined using a mechanism-based kinetic model (e.g., kinetic model 2305 of FIG. 23). At 2104, a patient-specific cell dosage of the cell therapy product is determined using a mechanism-based kinetic model. In some implementations, the patient-specific dosage is determined based on the tumor burden and the predicted Cmax such that the ratio of Cmax / tumor burden is improved (e.g., optimized). In some implementations, the Cmax and the patient-specific cell dosage depend on the subpopulation of the cell population and the patient's tumor burden. In some implementations, a subpopulation of cells enriched for memory T cells with high proliferation rates, high cytotoxicity potential, and / or lack of exhaustion are associated with more favorable clinical outcomes.

[0248] In some implementations, method 2100 further includes causing preparation of a single dosage form containing the determined patient-specific cell dosage. For example, the processor can send a signal that causes the single dosage form to be prepared (e.g., by a computing device, by a medical professional, etc.). In some implementations, a patient-specific cell dosage form is generated according to method 2100.

[0249] FIG. 22 shows a flowchart of a method 2200 for determining a patient-specific dosage of a cell therapy to be administered, according to one embodiment. In some implementations, the method 2200 may be performed by a processor (e.g., processor 2301). For example, instructions causing the processor 2301 to execute the method 2000 may be stored in memory 2302 of FIG. 23. In some implementations, a patient-specific cell dosage form is generated that includes cells at a dosage determined according to the method 2200. In some implementations, by performing the method 2200, the patient-specific cell dosage can be used to prepare a cell therapy product at a patient-specific cell dosage. The patient can then be administered the cell therapy product at a cell dosage specified for that patient (e.g., not under-obtained or not over-obtained).

[0250] In 2201, a cell population is provided. In some implementations, the one or more characteristics of the cell therapy include an initial cell dose and / or a cell subpopulation characteristic. In some implementations, the cell therapy includes immune cell therapy. In some implementations, the cell therapy includes immune cell therapy, and the immune cell therapy includes CAR-T cell therapy. In some implementations, the cell therapy includes immune cell therapy, including immune cell therapy, including CAR-T cell therapy, and at least one of autologous immune cell therapy or allogeneic immune cell therapy.

[0251] At 2202, gene expression data for a cell population is received. At 2203, the gene expression data is analyzed to identify a set of gene signatures. At 2204, the set of gene signatures is provided as input to a machine learning classifier (e.g., machine learning classifier 2304 of FIG. 23) to generate a predicted proliferation potential (Cmax) for the cell population. In some implementations, the machine learning classifier is trained according to method 1600. At 2205, the Cmax of the cell population and the patient's tumor burden are provided as input to a mechanism-based kinetic model (e.g., kinetic model 2305 of FIG. 23). In some implementations, the tumor burden is determined by at least one of a complete blood count, flow cytometry, radiology (PET / CT) scan, measurement of circulating tumor-DNA, or indication-specific circulating tumor biomarkers. At 2206, a patient-specific cell dosage of the cell therapy product is determined using the mechanism-based kinetic model. In some implementations, determining a patient-specific dosage is determined based on the tumor burden and predicted Cmax, such that the ratio of Cmax / tumor burden is improved (e.g., optimized).

[0252] 23 is a schematic block diagram of a computing device 2300 according to one embodiment. The computing device 2300 may be or include a hardware-based computing device and / or a multimedia device such as a computer, desktop, laptop, smartphone and / or the like. The computing device 2300 includes a memory 2302, a communication interface 2303, and a processor 2301.

[0253] The memory 2302 of the computing device 2300 may be, for example, a memory buffer, a random access memory (RAM), a read only memory (ROM), a hard drive, a flash drive, etc. The memory 2302 may store, for example, code (e.g., a program written in C, C++, Python, etc.) including instructions that cause the processor 2301 to perform one or more processes, methods, and / or functions (e.g., methods 1600, 1700, 1800, 1900, 2000, 2100, and / or 2200).

[0254] The memory 2302 may include a machine learning classifier 2304 and / or a kinetic model 2305. The machine learning classifier 2304 and / or the kinetic model 2305 may be, for example, a machine learning model, an artificial intelligence model, an analytical model, and / or a mathematical model. In some implementations, the machine learning classifier 2304 may be trained using the method 1600 (e.g., in the computing device 2300 and / or a different computing device). In some implementations, the machine learning classifier 2304 may perform one or more steps discussed in the methods 1600, 1700, 1800, and / or 2200. In some implementations, the kinetic model 2305 may perform one or more steps discussed in the methods 1900, 2000, 2100, and / or 2200. In some implementations, the machine learning classifier 2304 and / or the dynamics model 2305 are stored in a different memory (i.e., not memory 2302) contained in a different computing device (i.e., not computing device 2300) that is communicatively coupled to the computing device 2300 via a network (not shown in FIG. 23 ).

[0255] The communication interface 2303 of the computing device 2300 may be a hardware component of the computing device 2300 to facilitate data communication between the computing device 2300 and external devices (e.g., a network, a computing device, and / or a server, not shown). The communication interface 2303 may be operatively coupled to and used by the processor 2301 and / or memory 2302. The communication interface 2303 may be, for example, a network interface card (NIC), a Wi-Fi® module, a Bluetooth® module, an optical communication module, and / or any other suitable wired and / or wireless communication interface.

[0256] The processor 2301 may be, for example, a hardware-based integrated circuit (IC) or any other suitable processing device configured to execute or perform a set of instructions or a set of code. For example, the processor 2301 may include a general-purpose processor, a central processing unit (CPU), an application specific integrated circuit (ASIC), a graphics processing unit (GPU), and / or the like. The processor 2301 is operably coupled to the memory 2302 via a system bus (e.g., an address bus, a data bus, and / or a control bus, not shown). The processor 2301 may be configured to perform the methods 1600, 1700, 1800, 1900, 2000, 2100, and / or 2200.

[0257] To address various problems and advance the art, the entirety of this application (including the cover page, title, headings, background, abstract, brief description of drawings, detailed description of the invention, claims, abstract, figures, appendices, and the like) sets forth, by way of example, various embodiments in which the claimed innovation may be practiced. The advantages and features of this application are those of representative examples of embodiments only and are not intended to be exhaustive and / or exclusive. They are presented for the purposes of understanding and teaching the principles of the invention.

[0258] Examples of computer code include, but are not limited to, microcode or microinstructions, machine instructions such as those produced by a compiler, code used to generate web services, and files containing high-level instructions that are executed by a computer using an interpreter. For example, embodiments may be implemented using Python, Java, JavaScript, C++, and / or other programming languages, packages, and software development tools.

[0259] The drawings are primarily for illustrative purposes and are not intended to limit the scope of the inventive subject matter described herein. The drawings are not necessarily to scale, and in some instances, various aspects of the inventive subject matter disclosed herein may be shown exaggerated or enlarged in the drawings to facilitate understanding of the various features. In the drawings, identical reference characters typically refer to identical features (e.g., functionally similar and / or structurally similar elements).

[0260] Operations performed as part of the disclosed methods may be ordered in any suitable manner. Thus, embodiments may be constructed to perform processes or steps in a different order than illustrated, including performing some steps or processes simultaneously even though illustrated embodiments show sequential operations. In other words, it should be understood that such features are not necessarily limited to a particular order of execution, but rather, may be limited to any number of threads, processes, services, servers, etc., that may execute sequentially, asynchronously, concurrently, in parallel, simultaneously, synchronously, and / or the like, in a manner consistent with the present disclosure. As such, some of these features may be mutually incompatible in that they cannot exist simultaneously in a single embodiment. Similarly, some features may be applicable to one aspect of the innovation and not to other aspects.

[0261] The term "and / or" as used herein and in the embodiments should be understood to mean "either or both" of the elements connected thereby, i.e., in some cases conjunctively and in other cases disjunctively. Multiple elements listed with "and / or" should be interpreted in the same manner, i.e., "one or more" of the elements connected thereby. Other elements other than the elements specifically identified by the "and / or" clause can optionally be present, whether related or unrelated to the specifically identified element. Thus, as a non-limiting example, a reference to "A and / or B" when used with open-ended language such as "comprising" can refer in one embodiment to only A (optionally including elements other than B), in another embodiment to only B (optionally including elements other than A), and in yet another embodiment to both A and B (optionally including other elements), and so forth.

[0262] As used herein in the specification and embodiments, "or" should be understood to have the same meaning as "and / or" as defined above. For example, the disjunctive items "or" or "and / or" within a list should be construed as inclusive, i.e., to include at least one, but also more than one, of the elements or list of elements, and optionally additional items not listed. Only terms clearly indicated to the contrary, such as "only one of" or "exactly one of," or, when used in the embodiments, "consisting of," will refer to the inclusion of the elements or exactly one element of the list. In general, the term "or" as used herein shall be construed as indicating exclusive alternatives (i.e., "one or the other, but not both") only when preceded by an exclusive term, such as "either," "one of," "only one of," or "exactly one of." When used in the embodiments, "consisting essentially of" shall have its ordinary meaning as used in the field of patent law.

[0263] Some embodiments and / or methods described herein can be performed by software (executed on hardware), hardware, or a combination thereof. Hardware modules can include, for example, a processor, a field programmable gate array (FPGA), and / or an application specific integrated circuit (ASIC). Software modules (executed on hardware) can include instructions stored in a memory operably coupled to a processor, and can be expressed in various software languages ​​(e.g., computer code), including C, C++, Java™, Ruby, Visual Basic™, and / or other object-oriented, procedural, or other programming languages ​​and development tools. Examples of computer code include, but are not limited to, microcode or microinstructions, machine instructions such as those generated by a compiler, code used to generate a web service, and files containing high-level instructions executed by a computer using an interpreter. For example, embodiments may be implemented using imperative programming languages ​​(e.g., C, Fortran, etc.), functional programming languages ​​(Haskell, Erlang, etc.), logic programming languages ​​(e.g., Prolog), object-oriented programming languages ​​(e.g., Java, C++, etc.), or other suitable programming languages ​​and / or development tools. Further examples of computer code include, but are not limited to, control signals, encryption code, and compression code.

Claims

1. An apparatus comprising: Memory, a hardware processor operatively coupled to the memory, receiving gene expression data for a plurality of cells; performing a differential gene expression analysis based on the gene expression data to generate differential gene expression data; estimating gene signature enrichment for each sample for a plurality of identified biological pathways based on the differential gene expression data; filtering gene signatures having differential enrichment patterns based on a predefined threshold between groups of responders and non-responders to the treatment; Grouping the filtered gene signatures into groups based on pairwise correlations between gene signature enrichment scores; performing a predetermined number of iterations to determine a plurality of sets of gene signatures, randomly selecting a gene signature from each group from the plurality of groups to determine a set of gene signatures from the plurality of sets of gene signatures; providing the plurality of sets of gene signatures as input to a feature selection algorithm to determine a feature vector comprising a reduced set of gene signatures identified from the plurality of sets of gene signatures; a hardware processor configured to train a machine learning classifier using the feature vector.

2. 2. The apparatus of claim 1, wherein the machine learning classifier is at least one of logistic regression, multinomial logistic regression, a decision tree, a perceptron, a support vector machine, K-nearest neighbors, a naive Bayes classifier, or a random forest.

3. The apparatus of claim 1 , wherein the feature selection algorithm is a genetic algorithm.

4. 2. The apparatus of claim 1, wherein filtering the gene signature is based on statistically significant differences between a group of responders and a group of non-responders.

5. The device of claim 1 , wherein the cell is an immune cell.

6. The device of claim 5, wherein the immune cells are chimeric antigen receptor expressing T cells (CAR-T cells).

7. The device of claim 6 , wherein the immune cells comprise at least one of autologous or allogeneic cells.

8. 1. A method for predicting a patient's clinical outcome in response to a cell therapy treatment, comprising: receiving gene expression data for a plurality of cells to be administered to the patient; Estimating gene signature enrichment for each sample for the identified biological pathways; and filtering gene signatures that are differentially enriched between responders and non-responders to a cell therapy; providing the filtered gene signature as an input to a machine learning classifier; generating a predicted clinical outcome.

9. The method of claim 8 , wherein the machine learning classifier is a logistic regression based classifier.

10. 10. The method of claim 8, wherein filtering the gene signature is based on statistically significant differences between groups of responders and non-responders.

11. The method of claim 8 , wherein the cell is an immune cell.

12. The method of claim 11, wherein the immune cell is a T cell expressing a chimeric antigen receptor (CAR-T cell).

13. 13. The method of claim 12, wherein the immune cells comprise at least one of autologous or allogeneic cells.

14. The method of claim 8 , wherein the machine learning classifier is trained according to one of the devices.

15. 9. The method of claim 8, wherein the method further comprises comparing predicted clinical outcomes for different cell populations, cell features, or parameters for cell growth or manufacturing, and identifying those associated with a more favorable clinical outcome.

16. 9. The method of claim 8, wherein the predicted clinical outcome comprises at least one of: (i) complete response, non-response, or partial response; (ii) reduction in tumor size or tumor burden to a predefined threshold; or (iii) cytokine release syndrome (CRS) or other toxicity.

17. 1. A method comprising: generating a cell population; receiving gene expression data for the cell population; analyzing the gene expression data to identify a set of gene signatures; providing the set of gene signatures as input to a machine learning classifier; generating a predicted clinical outcome.

18. 20. The method of claim 17, wherein the machine learning classifier is a logistic regression based classifier.

19. The method of claim 17 , wherein the machine learning classifier is trained according to one of the devices.

20. 18. The method of claim 17, wherein the cell is an immune cell.

21. The method of claim 20, wherein the immune cell is a T cell expressing a chimeric antigen receptor (CAR-T cell).

22. 22. The method of claim 21, wherein the immune cells comprise at least one of autologous or allogeneic cells.

23. 20. The method of claim 17, wherein the method further comprises comparing predicted clinical outcomes for different cell populations, cell features, or parameters for cell growth or manufacturing to identify those associated with a more favorable clinical outcome.

24. The method further comprises generating a reduced set of gene signatures from the identified set of gene signatures, and analyzing the gene expression data further comprises: estimating gene signature enrichment for each sample for a plurality of identified biological pathways based on the differential gene expression data; filtering gene signatures having differential enrichment patterns based on a predefined threshold between a group of responders and a group of non-responders to the treatment; Grouping the filtered gene signatures into a plurality of groups based on pairwise correlations between gene signature enrichment scores; performing a predetermined number of iterations to determine a plurality of sets of gene signatures, randomly selecting a gene signature from each group from the plurality of groups to determine a set of gene signatures from the plurality of sets of gene signatures; 20. The method of claim 17, comprising: providing the plurality of sets of gene signatures as input to a feature selection algorithm to define a feature vector comprising a reduced set of gene signatures identified from the plurality of sets of gene signatures.

25. 18. The method of claim 17, wherein the predicted clinical outcome comprises at least one of: (i) complete response, non-response, or partial response, (ii) reduction in tumor size or tumor burden to a predefined threshold, or (iii) cytokine release syndrome (CRS) toxicity.

26. 1. A method for predicting a patient's clinical outcome in response to a cell therapy, comprising: Measuring one or more characteristics of the patient's tumor burden and cell therapy; providing the measured tumor mass and the one or more characteristics of the cellular therapy as inputs to a mechanism-based kinetic model; predicting a pharmacokinetic or pharmacodynamic response of the cell therapy in the patient using the mechanism-based kinetic model; and predicting the clinical outcome of the patient in response to the cell therapy.

27. 27. The method of claim 26, wherein the mechanism-based kinetic model predicts the pharmacokinetic or pharmacodynamic response of the cell therapy in the patient.

28. 27. The method of claim 26, wherein the one or more characteristics of the cell therapy include at least one of an initial cell dose or a cell subpopulation characteristic.

29. 27. The method of claim 26, wherein the cell therapy comprises immune cell therapy.

30. 30. The method of claim 29, wherein the immune cell therapy comprises CAR-T cell therapy.

31. 31. The method of claim 30, wherein the immune cell therapy comprises at least one of an autologous immune cell therapy or an allogeneic immune cell therapy.

32. 27. The method of claim 26, wherein the tumor burden is determined by at least one of a complete blood count, flow cytometry, radiology (PET / CT) scan, measurement of circulating tumor-DNA, or indication-specific circulating tumor biomarkers.

33. 1. A method of administering cell therapy to a patient, comprising: determining one or more characteristics of the patient's tumor burden and the cell therapy; providing the measured tumor mass and the one or more characteristics of the cellular therapy as inputs to a mechanism-based kinetic model; predicting a clinical outcome for said patient using said mechanism-based kinetic model; administering the cell therapy to the patient, wherein the dosage administered to the patient is determined by the mechanism-based kinetic model.

34. 34. The method of claim 33, wherein the dosage administered to the patient is determined by the mechanism-based kinetic model based on at least the patient's tumor burden and the Cmax of the cell therapy.

35. 34. The method of claim 33, wherein the mechanism-based kinetic model predicts at least one of a pharmacokinetic or pharmacodynamic response of the cell therapy in the patient.

36. 34. The method of claim 33, wherein the one or more characteristics of the cell therapy include at least one of an initial cell dose or a cell subpopulation characteristic.

37. 34. The method of claim 33, wherein the cell therapy comprises immune cell therapy.

38. 38. The method of claim 37, wherein the immune cell therapy comprises CAR-T cell therapy.

39. 39. The method of claim 38, wherein the immune cell therapy comprises at least one of an autologous immune cell therapy or an allogeneic immune cell therapy.

40. 34. The method of claim 33, wherein the tumor burden is determined by at least one of a complete blood count, flow cytometry, radiology (PET / CT) scan, measurement of circulating tumor-DNA, or indication-specific circulating tumor biomarkers.

41. 1. A method for producing a cell therapy product, comprising: Providing a cell population having at least one subpopulation; Measuring the tumor burden in the patient; determining a predicted proliferative potential (Cmax) of said cell therapy product using a mechanism-based kinetic model; and determining a patient-specific cell dosage of the cell therapy product using the mechanism-based kinetic model; The method, wherein the Cmax and patient specific cell dose are dependent on subpopulations of the cell population and tumor burden of the patient.

42. 42. The method of claim 41, wherein patient-specific dosing is determined based on tumor burden and predicted Cmax such that the ratio of Cmax / tumor burden is improved.

43. 42. The method of claim 41, wherein a subpopulation of cells enriched for memory T cells having a high proliferation rate, high cytotoxic potential, or lack of exhaustion is associated with a more favorable clinical outcome.

44. 42. The method of claim 41, wherein the cell therapy comprises immune cell therapy.

45. 45. The method of claim 44, wherein the immune cell therapy comprises CAR-T cell therapy.

46. 46. ​​The method of claim 45, wherein the immune cell therapy comprises at least one of an autologous immune cell therapy or an allogeneic immune cell therapy.

47. 42. The method of claim 41, wherein the tumor burden is determined by at least one of a complete blood count, flow cytometry, radiology (PET / CT) scan, measurement of circulating tumor-DNA, or indication-specific circulating tumor biomarkers.

48. 42. The method of claim 41, further comprising the step of preparing a single dosage form containing the determined patient-specific cell dosage.

49. 42. A patient-specific cell dosage form made according to the method of claim 41.

50. 1. A method for determining a patient-specific dosage of a cell therapy to be administered, comprising: Providing a population of cells; receiving gene expression data for the cell population; analyzing the gene expression data to identify a set of gene signatures; providing the set of gene signatures as input to a machine learning classifier to generate a predicted proliferative potential (Cmax) of the cell population; providing the Cmax of the cell population and the patient's tumor burden as inputs into a mechanism-based kinetic model; and determining a patient-specific cell dosage of a cell therapy product using the mechanism-based kinetic model.

51. 51. The method of claim 50, wherein the step of determining a patient-specific dose is determined based on tumor burden and predicted Cmax such that the ratio of Cmax / tumor burden is improved.

52. 51. The method of claim 50, wherein the machine learning classifier is trained according to one of the devices.

53. 51. The method of claim 50, wherein the cell therapy comprises immune cell therapy.

54. 54. The method of claim 53, wherein the immune cell therapy comprises CAR-T cell therapy.

55. 55. The method of claim 54, wherein the immune cell therapy is selected from the group consisting of autologous immune cell therapy or allogeneic immune cell therapy.

56. 51. The method of claim 50, wherein the tumor burden is determined by at least one of a complete blood count, flow cytometry, radiology (PET / CT) scan, measurement of circulating tumor-DNA, or indication-specific circulating tumor biomarkers.

57. 51. A patient-specific cell dosage form comprising cells at a dosage determined according to the method of claim 50.

58. 1. A method comprising: receiving gene expression data for a plurality of cells; performing a differential gene expression analysis based on the gene expression data to generate differential gene expression data; estimating gene signature enrichment for each sample for a plurality of identified biological pathways based on the differential gene expression data; determining a set of identified biological pathways from the plurality of identified biological pathways that are statistically significant based on the gene signature enrichment for each sample; providing the set of identified biological pathways as input to a feature selection algorithm to define a feature vector; and training a machine learning classifier using the feature vector.

Images