Tumor microenvironment by liquid biopsy

Abstract

The present invention provides a method for non-invasively profiling tumors or tumor microenvironments using bodily fluid samples, such as blood samples. The disclosed method uses liquid biopsy analysis to profile the abundance and diversity of immune cells in the tumor microenvironment and predict patient response to treatment, particularly immunotherapy. Specifically, the disclosed method is useful for predicting immunotherapy toxicity (and, more generally, treatment toxicity) from cell-free DNA methylation cellular status analysis.

Description

[Technical Field]

[0001] Technical Field The present disclosure relates to non-invasive profiling of the tumor microenvironment from markers in cell-free DNA and prediction of immunotherapy response and toxicity. [Background technology]

[0002] background Tumors are complex ecosystems that contain cancer cells along with a variety of other cells, some of which support tumor growth, some of which attempt to kill the cancer cells, and some of which act as bystanders. Tumors may also contain blood vessels and extracellular matrix. The environment in which cancer cells reside is known as the tumor microenvironment (TME). See Anderson, 2020, The tumor microenvironment, Current Biol 30(16):R921-R925 (incorporated by reference).

[0003] Leukocytes (immune cells) are an important part of the TME and are also found circulating in the blood. Leukocytes that infiltrate tumors are called tumor-infiltrating leukocytes (TILs), while those that circulate in the blood are peripheral blood leukocytes (PBLs). See Chen, 2013, Oncology meets immunology: the cancer-immunity cycle, Immunity 39(1):1-10 (incorporated by reference). Examples of some important leukocyte types are cytotoxic T cells, Tregs, B cells, NK cells, and neutrophils. Tumor evolution is highly dependent on the TME, especially TILs. For example, cytotoxic T cells can help kill cancer cells, while Tregs can help tumor growth. The role of TILs within the TME is variable, and these cells can potentially be tumor cell killers or tumor cell growth promoters, or even change roles depending on the conditions. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Anderson, Current Biol (2020)30(16):R921~R925 [Non-patent document 2] Chen,2013,Oncology meets immunology:the cancer-immunity cycle,Immunity 39(1):1-10 Summary of the Invention [Means for solving the problem]

[0005] overview The present invention provides a method for non-invasively profiling tissue microenvironments using bodily fluid samples. The method of the present invention is useful for analyzing and detecting abnormal tissue microenvironments (ATMs), such as those caused by tumors (tumor microenvironment), inflammatory conditions (inflammatory microenvironment), tissue transplants (transplant microenvironment), and various pathogens (infectious microenvironment). The present invention provides a non-invasive, computational, end-to-end framework for profiling ATMs using liquid biopsy samples (LiquidTME) and sequencing associated nucleic acids, particularly cell-free DNA (cfDNA). In a preferred embodiment, the LiquidTME method of the present invention analyzes and profiles ATMs (including tumor ATMs) using analysis of cell-free DNA methylation and cellular states in nucleic acids obtained from liquid biopsy samples. The method of the present invention provides a cellular profile in a specific microenvironment.

[0006] The LiquidTME method disclosed herein provides a robust and sensitive profile of the tissue microenvironment distributed in a subject using only liquid biopsy samples. As described herein, the various microenvironment analyses and profiles generated using the methods of the present invention have been validated using data generated from different assays. The methods disclosed herein have been validated and are more sensitive than many conventional first-line disease analysis techniques. For example, the methods of the present invention provide early detection of TMEs that have transitioned or are transitioning from a benign to a malignant state. The methods of the present disclosure detect malignant TMEs even in patients with low tumor burden using predictions obtained from liquid biopsy samples. Thus, the methods of the present invention provide an advance in the standard of care for assessing cancer.

[0007] Similarly, the disclosed method uses liquid biopsy analysis to profile the immune cell repertoire (e.g., immune cell abundance and diversity) in the microenvironment and predict patient response to treatment. In certain embodiments, the disclosed method is used to predict immunotherapy toxicity (and, more generally, treatment toxicity) and / or response durability based on cell-free DNA methylation patterns and cell state analysis. In many diseases, immunotherapy is the only effective treatment for some individuals, while for others it results in no improvement or even adverse effects. Therefore, immunotherapy is often a second-line treatment option only after the first-line treatment has proven unsuccessful.

[0008] For example, it is understood that some patients, particularly cancer patients, experience a category of toxic effects in response to immunotherapy known as immune-related adverse events (irAEs). It is also understood that not all patients (irAEs notwithstanding) will respond positively to immunotherapy. The method of the present invention provides a profile of the cellular composition of ATM that predicts immunotherapy toxicity and patient response. Before administering pre-immunotherapy, such as immune checkpoint inhibitors, to a patient, analysis can be performed to generate a predictive profile.

[0009] The provided profile may include, for example, a combined model that provides a measure of both activated CD4 T effector memory (TEM) cell abundance and T cell receptor (TCR) diversity. Data shows that CD4 TEM abundance or TCR diversity can predict immunotherapy response or toxicity, and the present invention provides measures of CD4 TEM abundance and TCR diversity by analyzing cell-free DNA (cfDNA) from non-invasively obtained blood or plasma samples.

[0010] Specifically, embodiments of the present invention involve identifying epigenetic modifications in cfDNA and measuring TEM abundance or TCR diversity from the patterns of epigenetic modifications. For example, using techniques such as bisulfite sequencing or enzymatic methyl-seq (EM-seq), the methods of the present invention provide data describing the patterns or locations of methylated DNA on nucleic acids derived from ATMs. Deconvolution and assembly software identifies which genes in ATMs have methylation (e.g., promoter methylation) levels. These results can be added to or compared with a liquid biopsy atlas that maps cfDNA sequence patterns to cellular states, disease states, or physiological conditions. After deconvolution and assembly (identifying patterns of epigenetic modifications at the gene-specific level), patterns of epigenetic gene modifications or sets of modified genes can be searched in the atlas to quantify or identify cell or tissue types or conditions, and to predict or monitor the risk or status of a wide range of physiological conditions, disorders, infections, and diseases.

[0011] The lead assembly and atlas search identifies cells or specific cellular states (e.g., transcriptional states) present or present in a patient's ATM. Identifying an ATM as having cells in a particular set of cellular states strongly predicts disease progression and / or irAEs (e.g., toxicity) in immunotherapies such as immune checkpoint inhibitors; and also predicts patient response (i.e., whether the patient will respond well to immunotherapy). The lead assembly and atlas may also provide measures of T cell abundance and diversity (e.g., T cell abundance and TCR diversity), thereby providing a profile of the TIL and immune repertoire in the ATM.

[0012] The disclosed methods are useful for granularly profiling peripheral blood cell states and / or quantifying activated immune cells from cell-free DNA. Such methods provide pre-treatment and early treatment prediction of immunotherapy toxicity (i.e., immune-related adverse events or irAEs). Furthermore, the disclosed methods are useful for simultaneously predicting both treatment response and toxicity using the same assay. Furthermore, the disclosed methods are useful for simultaneously quantifying a wide range of cellular states that comprehensively represent human health and disease, essentially providing an atlas for predicting and monitoring the risk of a wide range of physiological conditions, disorders, infections, and diseases (i.e., by measuring multiple cell, tissue, and microbial types / states, including those sequenced or present in public / published methylation datasets). Thus, the present invention provides methods for analyzing cfDNA methylation of cell / tissue states to predict treatment response and toxicity.

[0013] The disclosed methods are useful for non-invasive early / pre-treatment prediction of severe immune-related adverse events by cell-free DNA analysis; simultaneous prediction of both severe immune-related adverse events and immunotherapy response by early / pre-treatment cell-free DNA analysis; providing a liquid biopsy atlas of human health and disease with the ability to quantify multiple cell / tissue / microbial types / states; and predicting and monitoring risk / status for a wide range of physiological conditions, disorders, infections, and diseases.

[0014] In certain embodiments, the present invention provides a method for sequence read deconvolution and read counting applicable to cell-free DNA methylation data, comprising: (i) identifying CpG sites at a fragment level in cell-free DNA; (ii) comparing the methylation level for each CpG site per fragment with a ground-truth reference table of known cell / tissue states; and (iii) counting or matching across the CpG sites in this manner, fragment by fragment, to assign the cell-free DNA fragments to cell states in the reference table. Preferably, the method comprises: (iv) continuing to count cell-free DNA fragments until substantially all fragments have been analyzed for cell-state assignment, and / or (v) determining the cell-state composition of the cell-free DNA mixture (i.e., derived from a liquid biopsy sample) from the cell-state assignments. The CpG sites per fragment may be provided as input to a machine learning algorithm that assigns fragments to cell states. The cell-state composition may be provided as a report of ATM-infiltrating immune cells, such as leukocytes, in the tumor microenvironment. Thus, the present disclosure provides a liquid biopsy measure of ATM. The method of the present disclosure provides high-resolution methylation cell state analysis of plasma cell-free DNA. The method herein is useful for quantifying any number of (e.g., >20 or >50 or more) different cell states in blood plasma. The method herein provides the ability to simultaneously predict immunotherapy response and toxicity from the very same plasma sample and sequencing results, and do so before treatment. The method of the present disclosure has important clinical implications for pre-treatment / early immunotherapy response and toxicity prediction.

[0015] In certain embodiments, the present invention provides a method for predicting therapeutic efficacy. The method includes identifying epigenetic modifications in nucleic acids from a sample, measuring the abundance and / or diversity of T cells from the pattern of the epigenetic modifications, and predicting response to the treatment based on the abundance and / or diversity of the T cells. The measured T cell diversity can be the diversity of T cell receptors (TCRs) in the patient's immune repertoire. The method can include predicting the risk of adverse events in response to immunotherapy or predicting responsiveness to immunotherapy when the TCR diversity is below a predetermined threshold.

[0016] In preferred embodiments, the sample is a blood or plasma sample, a urine sample, or other biological fluid sample obtained by liquid biopsy. Preferably, the nucleic acid is cell-free DNA. The identified epigenetic modifications may include promoter methylation. The method may include querying a cell state atlas for patterns of epigenetic modifications. That is, a set of methylated gene promoters in the sample can be searched in the atlas to identify a cell state associated with that set of methylated gene promoters. The method may include providing a profile describing the cell state of multiple cells in the tumor microenvironment of the patient. Such a profile may include, for example, a list of cells present and the transcriptional state of the cells, such as the transcription levels of certain genes in those cells. In some embodiments, the method includes providing a profile of tumor-infiltrating leukocytes in the tumor microenvironment based on the pattern of epigenetic modifications. Preferably, the method includes inferring a more detailed state of tumor-infiltrating leukocytes and other tumor microenvironmental elements within this cell-free DNA TME / ATM compartment (i.e., inferring the tumor microenvironment composition from cell-free DNA analysis).

[0017] In certain embodiments, the present invention provides methods for predicting disease outcomes and / or risk stratifying patients. Exemplary methods include identifying epigenetic modifications in sequence data, including methylation states of nucleic acid bases, from cell-free nucleic acids derived from a liquid biopsy sample obtained from a subject; identifying an abnormal tissue microenvironment and assigning the cell-free nucleic acids to a tissue of origin based on the identified pattern of epigenetic modifications; determining the state of cells in the abnormal tissue microenvironment using the epigenetic modification pattern; and predicting disease outcomes in the subject based on the state of cells in the abnormal tissue microenvironment.

[0018] In certain embodiments, the abnormal tissue microenvironment is selected from the group consisting of tumor microenvironment (TME), inflammatory microenvironment (IME), tissue transplant microenvironment (TTME), and pathogen microenvironment (PME). In some embodiments, the circulating cell-free nucleic acids are derived from microenvironment-infiltrating lymphocytes. In certain embodiments, the disease is cancer, and the abnormal tumor microenvironment is TME.

[0019] In certain methods, predicting the outcome of cancer in a subject comprises one or more of the following: diagnosing the subject with cancer, predicting remission, predicting recurrence, assessing minimal residual disease, predicting the response to immunotherapy, and predicting the toxicity of immunotherapy.Generally, the method of the present invention allows risk stratification in patients with locally advanced cancer or metastatic cancer; and is useful for distinguishing between patients with molecularly low-risk disease and patients with molecularly high-risk disease.This type of stratification is applicable across cancers, and helps to enable and enhance personalized treatment.

[0020] In certain embodiments, the methods of the present invention can detect tumor microenvironment and thus cancer in subjects with low measured tumor burden or in cancer screening settings (patients without a known cancer diagnosis). Predicting cancer outcome in a subject using the methods of the present invention can include determining that the cellular state in the TME exceeds a premalignant threshold. This can be based on the detailed composition of TME cellular states derived from patterns of epigenetic modifications and / or the abundance and / or diversity of T cells. The measured T cell diversity can be the diversity of T cell receptors (TCRs) in the patient's immune repertoire. Certain methods of the present invention can include inferring the abundance of T cell effector memory cells from the pattern.

[0021] In certain embodiments, the methods of the present invention comprise predicting the risk of an adverse event in response to immunotherapy, wherein the measured TCR diversity and / or predicted abundance of T cell effector memory cells exceeds a predetermined threshold. The methods may further comprise predicting a positive outcome in response to immunotherapy, wherein the measured cell-free DNA-derived TME profile, the abundance of a particular cellular state within the TME, or the abundance of total tumor-infiltrating leukocyte-derived cell-free DNA exceeds a predetermined threshold.

[0022] In a preferred method, the step of determining the cell state includes sequencing nucleic acids from a purified version of the cell state to generate sequence data comprising methylated bases; mapping the sequence data to a reference to identify promoters of a plurality of genes; and identifying the cell state based on the set of genes with hypomethylated promoters.

[0023] The methods of the invention may include providing a profile of tumor-infiltrating leukocytes in the tumor microenvironment based on patterns of epigenetic modifications.

[0024] In certain methods, the cell-free nucleic acids include nucleic acids derived from non-tumor cells, and determining the cellular state in the abnormal tissue microenvironment includes creating a TME profile. The non-tumor cells may include one or more of stromal cells, immune cells, and / or cells derived from the tumor margin. The TME profile may include one or more epigenetic patterns that correlate with tumor progression and / or tumor regression.

[0025] The methods of the present invention can predict the toxicity of immunotherapy based on the epigenetic modification levels of promoters in genes involved in the transcriptional status of a given T cell and / or the status of other immune cells; and can predict the response to immunotherapy based on the promoter methylation levels in genes corresponding to a given tumor microenvironment cell community. The transcriptional status of a given T cell can be specific to CD4 T cells. The methods of the present invention can also predict the toxicity of immunotherapy based on the epigenetic modification levels of promoters in genes involved in the expression of one or more cytokines; and can predict the response to immunotherapy based on the promoter methylation levels in genes in a given tumor ecotype.

[0026] In certain embodiments, the epigenetic modification is predictive of immunotherapy toxicity and is selected from one or more of interleukin-10 promoter hypomethylation; interleukin-6 promoter hypermethylation; and interleukin-7 promoter hypermethylation.

[0027] In a preferred embodiment, the present invention involves selecting fragments of a preferred size for TME analysis. For example, selecting cfDNA fragments with sizes consistent with those produced by tumors allows subsequent aneuploidy or other analyses to focus on tumor-derived DNA. Thus, liquid TME analysis can be based on a fragmentomic screen that focuses on fragments known to be tumor-associated and containing specific terminal motifs in cell-free DNA fragments. In one embodiment, the copy number of cfDNA fragments is used as a surrogate for aneuploidy. In another embodiment, aneuploidy analysis is combined with other diagnostic criteria, such as tumor mutation burden, to stratify patients for immunotherapy response. In this way, blood samples can be deconvoluted into matrices containing fragment size, copy number, mutation load, and / or epigenetic signatures, providing detailed resolution regarding future immunotherapies and potential disease progression and outcomes.

[0028] In addition, the method of the present invention utilizes tumor-derived LTME signal to determine aneuploidy and predict disease outcome. Generally, aneuploidy predicts response to immunotherapy in most patients. According to the present invention, determining copy number in LTME samples is used as a surrogate for aneuploidy. Highly altered copy number predicts high aneuploidy, and low copy number predicts low aneuploidy. Therefore, LTME signal is useful as a surrogate for aneuploidy to predict response to immunotherapy.

[0029] In another embodiment, LTME samples are used to select nucleic acid fragment size as a proxy for aneuploidy.For example, the cfDNA derived from tumor in liquid biopsy samples is typically shorter than the cfDNA derived from non-tumor cells.Therefore, cfDNA is size-selected and used as an aneuploidy filter in LTME, thereby stratifying patients for the risk of severe disease and / or response to immunotherapy.Generally, LTME analysis based on cfDNA fragmentomics is a good predictor of the success of immunotherapy (for example, the likelihood of efficacy or adverse events).

[0030] An alternative embodiment is a hierarchical strategy for addressing multicollinearity. Broadly speaking, in a first step, cell types are grouped into broader classes (e.g., all T cells, all B cells) based on measures of similarity between cell-type methylation profiles (e.g., genome-wide promoter-level methylation, cell-type-enriched CpGs detected by feature selection, etc.). In this way, cell types within a given class will exhibit multicollinearity with each other (e.g., CD4 effector vs. central memory T cells), but cell types between classes will not (e.g., CD4 effector memory T cells vs. naive B cells).

[0031] In the second step, read counting is used to distribute the reads / fragments into each class with high precision. This is possible because each class is easily separated by a highly specific CpG profile. The reads assigned to each class are used to generate class-specific methylation profiles.

[0032] In the third step, cell-type CpGs within a given class are used to deconvolve the bulk CpG mixture of each class (i.e., derived from class-specific reads). This can be achieved by a number of statistical learning methods (e.g., gradient boosting decision trees [XGBoost], non-negative least squares regression with L2 norm regularization, deep learning) that regularize the results and deal with multicollinearity.

[0033] In the final step, the cell type-specific fractions within each class are combined by weighting them by the global fractional abundance of each class (e.g., weighting each T cell subset by its fraction of total T cells in the mixture; repeat for all cell types).

[0034] Further examples and advantages of the present invention are provided below in the detailed description thereof. [Brief explanation of the drawings]

[0035] [Figure 1] FIG. 1 shows the workflow of the method of the present invention. [Figure 2] Figure 2 shows the results demonstrating the performance of the read counting. [Figure 3] Figure 3 shows that the more CpGs per fragment, the lower the false positive rate (FPR). [Figure 4] Figure 4 shows the detection limit analysis. [Figure 5] FIG. 5 graphically illustrates that the number of available fragments decreases as the CpG cutoff increases. [Figure 6] FIG. 6 shows the FPR and cell fraction for different numbers of CpGs per fragment. [Figure 7] Figure 7 shows the workflow for creating a signature matrix. [Figure 8] Figure 8 shows the results from tumor microenvironment-based deep deconvolution. [Figure 9] Figure 9 shows the signature for CD8 TILs. [Figure 10] FIG. 10 shows the correlation of tumor microenvironment-estimated TIL content. [Figure 11] FIG. 11 shows liquid biopsies for predicting immunotherapy response and toxicity. [Figure 12] FIG. 12 is a boxplot and ROC plot showing prediction of patient response. [Figure 13] FIG. 13 shows box plots and ROC plots for toxicity prediction. [Figure 14] FIG. 14 shows the tumor microenvironment profile. [Figure 15] FIG. 15 shows the relative abundance of the 20 cell states. [Figure 16] Figure 16 shows cell state abundance (scRNA-seq) versus future irAE status. [Figure 17] Figure 17 shows TCR diversity by scV(D)J-seq. [Figure 18] FIG. 18 shows CD4 memory T cell levels associated with irAEs. [Figure 19] Figure 19 shows a combined model predicting irAE grade. [Figure 20] Figure 20 shows time to severe irAE in concomitant ICI patients. [Figure 21] FIG. 21 shows TCR clone dynamics. [Figure 22] FIG. 22 shows CD4 TEM levels measured by CyTOF. [Figure 23] FIG. 23 shows the CD4 TEM expression profile. [Figure 24] FIG. 24 shows Tph levels measured by CyTOF. [Figure 25] FIG. 25 shows pre-treatment cell-free DNA analysis. [Figure 26] FIG. 26 shows activated CD4 TEM and CE9 scores. [Figure 27] FIG. 27 shows the results of the method herein. [Figure 28] Figure 28 shows pre-treatment CD4 Tph cells and irAE risk. [Figure 29] FIG. 29 shows the method steps. [Figure 30] Figure 30 is a schematic diagram of the proposed project. [Figure 31] Figure 31 compares TCR clonotypes with a database of CDR3 sequences. DETAILED DESCRIPTION OF THE INVENTION

[0036] Detailed Description The present invention provides a method for noninvasively profiling tissue microenvironments using bodily fluid samples. The method of the present invention can be used to analyze and detect abnormal tissue microenvironments, such as those caused by tumors (tumor microenvironment), inflammatory conditions (inflammatory microenvironment), tissue transplants (transplant microenvironment), and various pathogens (infectious microenvironment). The present invention provides a noninvasive, computational, end-to-end framework for profiling abnormal tissue microenvironments (ATMs) using liquid biopsy samples (LiquidTMEs) and sequencing associated nucleic acids, particularly cell-free DNA (cfDNA). The method of the present invention is useful for inferring closely related cell types and states (including cellular states associated with abnormal tissue microenvironment (e.g., tumor microenvironment) profiles) from blood plasma-derived cell-free DNA, which can be applied to different clinical situations. The present disclosure provides a novel platform for noninvasive profiling of infiltrating immune cells from abnormal microenvironments in patients, specifically, profiling and decoding infiltrating immune cell-derived methylation signatures identified from plasma-derived cell-free DNA molecules.

[0037] In a preferred embodiment, the LiquidTME method of the present invention uses liquid biopsy samples to analyze and profile tissue microenvironments. Using only liquid biopsy-derived samples, the method of the present invention can profile the cellular composition of abnormal tissue microenvironments, such as the tumor microenvironment (TME).

[0038] The methods of the present invention are useful for predicting the response and side effects of certain therapies, particularly immunotherapy treatments, in patients, such as cancer patients. In certain embodiments, the present invention uses methylation-based cell-free DNA analysis to detect circulating DNA from specific abnormal microenvironments and predict its tissue of origin. The methods of the present invention may also include detecting and / or quantifying ATM-infiltrating immune cells (e.g., tumor-infiltrating leukocytes in the TME) in a sample. The present invention utilizes the insight that the methylation profile of certain cells, particularly immune cells, derived from specific ATMs differs from the methylation profile of normal cells unrelated to the microenvironment. The methods of the present invention also include methods for deconvoluting methylated bulk mixtures into purified cell states. The methods of the present invention can use databases or atlases of cell state-specific methylation profiles, and the results can contribute to the databases or atlases.

[0039] To successfully ameliorate certain conditions, immune cells such as cytotoxic T cells (CTLs) must migrate into abnormal tissue microenvironments to destroy any infected or diseased cells. For example, in cancer, CTLs (as tumor-infiltrating lymphocytes [TILs]) must be able to migrate into the tumor microenvironment and destroy cancer cells. Studies have shown that TILs are often in a weakened state, limiting their ability to kill cancer cells. Similar states are known to be associated with other conditions, such as inflammatory states or pathogen-induced diseases. This weakened state is known as exhaustion. Exhausted T cells express inhibitory receptors to which cancer cells bind their corresponding ligands to avoid attack. As a result, despite the presence of TILs within the TME (or other relevant microenvironment), diseased cells may evade immune-mediated killing, thereby promoting their spread, growth, and ultimately the patient's death. See Wherry, 2015, Molecular and cellular insights into cell exhaustion, Nature Rev Immunol 15(8):486-499 (incorporated by reference).

[0040] The present invention provides a non-invasive liquid biopsy approach for profiling the cellular composition of such abnormal tissue microenvironments. Specifically, the method of the present invention is useful for inferring closely related cell types and states (including cellular states associated with ATM profiles) from cell-free DNA derived from blood plasma, which has clear utility and applicability in various clinical settings. The present invention provides methods for detecting, analyzing, and predicting the outcome of treatments targeting such microenvironments. In particular, the method of the present invention can be used to predict durable responses and / or toxic side effects to potential treatments such as immunotherapy. For example, the method of the present invention is useful for predicting the durability of treatment response and / or side effects of immune checkpoint inhibitor therapy in patients with cancers such as melanoma or colorectal cancer.

[0041] The first aspect of the present disclosure provides an ultrasensitive framework for profiling closely related cell types and states using DNA methylation. Based on methylation data, analytical platforms such as deep deconvolution algorithms can classify closely related cell states, including cellular states within specific ATMs such as tumor microenvironments.

[0042] A second aspect of the present disclosure provides tools for the development and use of atlases of cell state-specific methylation profiles. Using purified cells from in-house and public databases, the methods of the present invention can be used to identify distinct methylation profiles for a wide range of cell types (e.g., immune cells, somatic cells, tumor cells, cells originating from various organs) and states (e.g., healthy, diseased, undergoing tumor progression / regression, under stress, pre-treatment, or post-treatment). Such atlases can be used as a reference for deconvolving cell states from any given mixture. In doing so, the atlases can provide a reference for use in assessing and / or detecting one or more specific tissue microenvironments in a subject using liquid biopsy samples.

[0043] A third objective of the present disclosure is to apply the disclosed assays to multiple conditions leading to ATM. For example, this may include using ATM methylation profile data to diagnose diseases such as cancer or specific cancer types (e.g., colorectal and melanoma). Similarly, ATM methylation profiles can be used to provide insight into the location of ATMs, such as organs harboring tumors or infections. Similarly, the methods of the present disclosure may use ATM methylation profiles to assess whether the TME has crossed the threshold from benign to malignant. Additionally, the methods of the present disclosure may be used to assess TME malignancy even in patients with a low measured tumor mutation burden (TMB). Early detection of crossing the threshold for malignant disease is crucial, as premalignant tumors are better candidates for surgical resection and treatment.

[0044] In certain embodiments, the method of the present invention can be applied to advanced stage patients who have received certain treatments before treatment to profile ATM methylation profile response signatures. These signatures can be used in conjunction with patient-derived ATM methylation profile data to predict durable treatment response and / or toxic side effects. In particular, cfDNA-derived methylation patterns can provide predictions regarding the response / side effects of immunotherapy or immune checkpoint inhibitor treatment in cancer patients. These predictions can lead to dramatic changes in standard treatment. For example, although immune checkpoint inhibitors are often considered second-line treatment options, they provide dramatic positive responses in certain cancer patients. Identifying such patients means providing patients with the best possible treatment without exposing them to potentially harmful treatment modalities such as radiation and chemotherapy.

[0045] In certain embodiments, the method of the present invention can be applied to the advanced stage patients who have received certain treatment before treatment, and profile their response signature.In this way, the method of the present invention can be used to predict the patient's response and toxicity to potential treatment (for example, immune checkpoint inhibitor [ICI] or immunotherapy).For example, the method assay can be applied to the advanced stage melanoma patients who are treated with ICI before treatment, and can be used to identify response signatures and validate them in a held-out test set.In this way, the method of the present invention is useful for predicting the patient's response and toxicity to immune checkpoint inhibitor (ICI).

[0046] Solid tumors can be viewed as having two components: malignant cancer cells and other cells of the body intermixed with the malignant cancer cells. The tumor microenvironment (TME) is complex and plays an important role in driving inflammation, promoting tumor growth, and / or promoting cell death. The TME can include vascular cells, the cancer cells themselves, non-malignant immune cells, somatic cells surrounding the tumor margin, and the cell-derived extracellular environment. See Anderson, 2020, The tumor microenvironment, Current Biol 30(16):R921-R925 and Joyce, 2009, Microenvironmental regulation of metastasis, Nat Rev Can 9(4):239-252 (both incorporated by reference). Malignant cells can alter the TME in ways that prevent immune cells in the TME from effectively killing cancer cells. See Gajewski, 2013, Innate and adaptive immune cells in the tumor microenvironment, Nature Immunol 14(10):1014-1022 and Tumeh, 2014, Pd-1 blockade induces responses by inhibiting adaptive immune resistance, Nature 515(7528):568-571 (both incorporated by reference).

[0047] Other tissue microenvironments contain similar components to TME or are themselves part of TME. Just as TME is formed by the activity of cells, especially immune cells, within TME, other abnormal tissue microenvironments (such as inflammatory microenvironment (IME), tissue transplantation microenvironment (TTME), pathogenic microenvironment (PME)) are also formed. Therefore, the method of the present invention can be used to evaluate, profile and predict the toxicity / success of treatments targeting IME, TTME and / or PME.

[0048] Thus, in certain embodiments, the methods of the present invention can be used to predict whether a patient will respond positively or negatively to a treatment, such as immunotherapy.

[0049] For example, immune checkpoint inhibitors (ICIs) are a promising method for treating certain advanced cancer patients. ICIs block inhibitory receptors on TILs, a phenomenon that has revolutionized the field of cancer treatment. However, predicting ICI response in patients remains challenging, with success rates ranging from 1% to 60%. Furthermore, standard imaging techniques cannot reliably assess treatment response at early time points. Studies have shown that earlier assessment of ICI response can be achieved using serial biopsy analysis of the TME, a promising result but clinically impractical. Therefore, it is crucial to develop methods for noninvasively monitoring the TME and other ATMs using "liquid biopsies."

[0050] Malignant cells can alter the TME in ways that prevent immune cells in the TME from effectively killing cancer cells. See Gajewski, 2013, Innate and adaptive immune cells in the tumor microenvironment, Nature Immunol 14(10):1014-1022 and Tumeh, 2014, Pd-1 blockade induces responses by inhibiting adaptive immune resistance, Nature 515(7528):568-571 (both incorporated by reference). Immune checkpoint inhibitors (ICIs) can "release the brakes" on these immune cells, turning them into more potent cancer killers. The ability of ICIs to kill otherwise unresponsive tumors represents a potential game changer for the treatment of advanced cancers.

[0051] However, not all patients benefit from this treatment, and it is difficult to know which patients will benefit and which will not.Severe side effects associated with ICI treatment can occur, highlighting the importance of improving patient selection so that only patients who will benefit are treated.The response to ICI treatment can be predicted early by tumor biopsy analysis.However, there has not previously been a non-invasive method for such prediction.

[0052] The present invention provides a computational, non-invasive liquid biopsy approach for profiling the cellular composition of abnormal tissue microenvironments. Specifically, the method of the present invention is useful for estimating closely related cell types and states (including cell states related to tumor profiles) from cell-free DNA derived from blood plasma, and applying it to different clinical situations. The method of the present invention is useful for predicting the response and side effects of ICI treatment for patients with cancer, such as melanoma or colorectal cancer.

[0053] Because ICIs convert immune cell compartments in the TME into cancer-killing cells, treatment response is highly dependent on the cellular composition of the tumor. See, e.g., Thommen 2018, T cell dysfunction in cancer, Cancer Cell 33(4):547-562 (incorporated by reference). For example, a tumor's TME may lack immune cells with cancer-killing potential. Therefore, monitoring the TME before or during treatment would be beneficial. However, TME monitoring has previously required invasive biopsies. Serial patient biopsies are impractical and can suffer from sampling bias due to tumor heterogeneity. The present invention provides a noninvasive, computational, end-to-end framework for profiling the tumor microenvironment and sequencing cell-free DNA via liquid biopsy (herein, LiquidTME).

[0054] Biopsy is an invasive method of extracting cells directly from tissue (such as from a tumor) for detailed examination and is one of the first steps used by clinicians to diagnose many conditions, including cancer. Based on analysis of the extracted cells, physicians determine a course of treatment. This invasive method is standard practice for solid tumor malignancies, but it can be expensive, risky, and sometimes impractical. In particular, accurate monitoring of treatment response would require serial biopsies, which may not be feasible. Liquid biopsies are an alternative in which researchers attempt to test tumors from bodily fluids, such as blood.

[0055] Scientists discovered the existence of cell-free DNA (cfDNA) in blood plasma over 100 years ago. See Mandel, 1948, Nucleic acids in blood plasma in 1 man, CR Seances Soc Biol Fil 142:241-243 (incorporated by reference). As the name cfDNA suggests, these DNA fragments are not contained within cells but rather circulate within blood plasma. When cells die, some of their DNA fragments are released into the blood circulation, where they can be captured and measured within cfDNA. It is understood that a fraction of these circulating DNA fragments originate from malignant cells in cancer patients or similarly stressed cells in other abnormal tissue microenvironments (e.g., IME, TTME, and PME). These microenvironment-specific cell-free DNA fragments, when originating from the TME, are known as circulating tumor DNA (ctDNA). For clarity, the equivalent circulating DNA originating from diseased or stressed cells in the IME, TTME, and PME is referred to herein as ctDNA.

[0056] It is understood that, like cancer cells, TME itself also releases DNA into the blood circulation. In cancer, cell-free DNA fragments from the microenvironment can be referred to as circulating tumor-infiltrating lymphocyte TIL DNA (ctilDNA). IME, TTME, and PME produce equivalents, which are referred to herein as ctilDNA for clarity.

[0057] While several methods for detecting ctDNA are available, the present disclosure provides a method for detecting or quantifying ctilDNA, including providing its tissue of origin. The purpose of this disclosure is to provide a robust computational framework for detecting ctilDNA, called "LiquidTME," for liquid biopsies of specific ATMs. One important feature of TIL-derived cfDNA is the methylation of CpG dinucleotides.

[0058] In DNA, a cytosine (C) followed by a guanine (G) is known as CpG (the "p" represents the phosphate bond between them). Through epigenetic mechanisms, a methyl (CH3) group is added to the C at the CpG site. This phenomenon is known as CpG methylation. It has been shown that different cell types and states have specific methylation patterns that control gene expression. To quantify methylation patterns in DNA, next-generation sequencing after bisulfite treatment is commonly used. Briefly, in this bisulfite-based sequencing method, if the C in a CpG site is unmethylated, it is converted to uracil (U) and then recognized as thymine (T). On the other hand, if the C in a CpG site is methylated, it remains untouched. Finally, the sequenced reads are aligned with a reference genome, and the ratio of C to T is calculated for every CpG position. Recent studies have demonstrated the potential utility of cfDNA methylation for providing early detection of certain diseases and conditions, such as cancer, including determining the tissue of origin. However, TILs have not previously been profiled from cfDNA using CpG methylation. This disclosure provides a useful method for profiling TIL DNA using CpG methylation of cell-free DNA by developing a novel liquid biopsy approach. These profiles can be used to assess the composition of ATM and provide predictive assessment of certain diseases or conditions.

[0059] The literature on ctDNA detection suggests certain considerations that find the same applicability for similar cfDNA arising from non-tumor ATM.

[0060] Circulating tumor DNA or ctDNA is DNA fragments originating from cancer cells. Cancer is a disease that begins with genomic mutations. Based on tracking these mutation signatures, ctDNA can be quantified from cfDNA sequencing. Several ctDNA detection technologies exist, but they generally work by querying likely mutated genomic locations in cancer cells and sequencing these locations in plasma cfDNA in detail (also known as targeted sequencing). After targeted sequencing, predefined genomic locations are investigated for mutations, thus detecting and quantifying ctDNA molecules. See Newman, 2014, An ultrasensitive method for quantitating circulating tumor DNA with broad patient coverage, Nature Med 20(5):548-554; Forshew, 2012, Noninvasive identification and monitoring of cancer mutations by targeted deep sequencing of plasma DNA, Science Trans Med 4(136):136ra68-136ra68; and Murtaza, 2013, Non-invasive analysis of acquired resistance to cancer therapy by sequencing of plasma DNA, Nature 497(7447):108-112 (all incorporated by reference).

[0061] It is also understood that ctDNA detection is affected by background noise. Noise can be introduced during sample preparation and sequencing. This can confound results when quantifying rare ctDNA fragments in patients with low disease burden (e.g., early-stage cancer or minimal residual cancer after curative-intent treatment). See Chaudhuri, 2017, "Early detection of molecular residual disease in localized lung cancer by circulating tumor DNA profiling," Cancer discovery, 7(12):1394-1403; Chin, 2019, "Detection of solid tumor molecular residual disease (MRD) using circulating tumor DNA (ctDNA)," Molecular Diag Ther, 23(3):311-331; and Newman, 2016, "Integrated digital error suppression for improved detection of circulating tumor DNA," Nature Biotech, 34(5):547-555 (all incorporated by reference). Double-stranded variant support, in which both the plus and minus strands are sequenced and mutational variants are supported on both parental strands of DNA, significantly reduces this noise. However, requiring double-stranded variant support is an inefficient approach, as 80–90% of recovered cell-free DNA sequencing reads are typically single-stranded without double-stranded support. Another approach to noise reduction is to profile background error patterns by sequencing cell-free DNA from healthy donors and consider them when querying mutations in patient cfDNA. Other studies have shown that it is possible to reduce background noise by requiring simultaneous detection of adjacent mutations within the same cfDNA fragment.

[0062] The present invention uses methylation-based cell-free DNA analysis to detect circulating DNA from specific ATM and predict its tissue of origin.In particular, the method of the present invention can use methylation-based cell-free DNA analysis to detect cfDNA from ATM and predict its tissue of origin.For example, the method of the present invention can use ctDNA to predict tumor tissue of origin. For background, see: tumor detection and classification using plasma cell-free DNA methylomes,Nature 563(7732):579-583, Guo,2017,Identification of methylation haplotype blocks aids in deconvolution of heterogeneous tissue samples and tumor tissue-of-origin mapping from plasma DNA,Nature Genetics 49(4):635-642, and Li,2018,CancerDetector:ultrasensitive and non-invasive cancer detection at the resolution of individual reads See using cell-free DNA methylation sequencing data, Nucleic Acids Res 46(15):e89-e89 (all incorporated by reference).

[0063] Methylation-based ctDNA detection primarily involves two steps: (1) first identifying a cancer-specific methylation signature, similar to a mutation signature; and (2) using deconvolution techniques to detect ctDNA and infer the tumor's tissue of origin. By extension, methylation-based detection of cfDNA originating from non-cancerous ATM is also possible, since the presence of base mutations is not required.

[0064] Differentially methylated CpG sites can be used to identify the tissue origin of ctDNA. Instead of single CpG sites, adjacent CpG sites (termed methylation haplotype blocks, or MHBs) associated with each other can be used to more accurately deconvolute methylation data. Differentially methylated MHBs can be identified, and the tissue of origin of ctDNA and cfDNA can be identified using tools such as random forest classifiers. Another approach involves individually classifying aligned reads using the methylation pattern of each read. CancerDetector [Li 2018] can be useful in methods that use a beta-binomial model to assign every cfDNA sequencing read as either cancerous or non-cancer.

[0065] As discussed above, tumor-infiltrating leukocytes (TILs) (and equivalents in other microenvironments) are leukocytes (white blood cells) that infiltrate tumors and contribute to the composition of the tumor microenvironment. Based on the immune cells in the TME, tumors can be broadly classified into three main classes: immune-infiltrated, immune-excluded, and immune-silent. In the case of immune-infiltrated tumors, TILs infiltrate tumors and become resident within the tumor tissue. If TILs are found only at the tumor border, the tumor is classified as immune-excluded. Finally, some tumors completely lack immune cells and are classified as immune-silent.

[0066] Considering that one purpose of this disclosure is to detect TIL content from cfDNA, the molecular profile of TIL must be different from that of PBL. ATAC-seq can be used to obtain different epigenetic programs in microenvironment-specific immune cells, such as tumor-specific CD8 T cells. See Philip, 2017, Chromatin states define tumor-specific T cell dysfunction and reprogramming, Nature, 545(7655):452-456 (incorporated by reference).

[0067] The present invention utilizes the insight that the methylation profiles of certain cells, particularly immune cells, derived from particular abnormal microenvironments differ from the methylation profiles of normal cells unrelated to the microenvironment. Thus, the methods of the present invention can be used to detect and / or assess the presence of abnormal tissue microenvironments (e.g., TME) based on detecting the presence of nucleic acids with methylation profiles that correlate with cells, such as immune cells, arising from such microenvironments.

[0068] For example, it has been shown that the methylation profile of CD8 T cells originating from tumor tissue differs from that of normal CD8 T cells. See Yang, 2020, Distinct epigenetic features of tumor-reactive CD8+ T cells in colorectal cancer patients revealed by genome-wide DNA methylation analysis, Genome Biol 21(1):1-13 (incorporated by reference). Gene promoters from CD8 T cells isolated from the TME are hypomethylated for the tumor-reactive marker genes CD39 and CD103. Using such epigenetic patterns, these insights make the methods of the present invention useful for distinguishing TILs from PBLs using methylation. By extension, the methods of the present invention can determine whether cfDNA originates from ATM or normal tissue.

[0069] Heterogeneous tissues are composed of different cell types and states. In certain embodiments, deconvolution methods are used to computationally estimate the cell proportions of these different cell types from bulk sequencing data. Tissue deconvolution has been primarily developed for gene expression data, where the gene expression of tissue is modeled as a weighted sum of the gene expression of underlying cell types. CIBERSORT is such a common method, which first identifies signatures from 22 cell types, and then uses support vector regression to estimate the 22 cell types from bulk expression data. See Newman, 2015, Robust enumeration of cell subsets from tissue expression profiles, Nature Meth 12(5):453-457 (incorporated by reference).

[0070] CIBERSORTx is a recent extension of CIBERSORT that builds signature matrices from single-cell RNA sequencing data and provides the ability to profile different cell states (e.g., exhausted vs. non-exhausted CD8 T cells) within each deconvolved cell type. See Newman, 2019, Determining cell type abundance and expression from bulk tissues with digital cytometry, Nature biotechnology, 37(7):773-782 (incorporated by reference).

[0071] The concept of deconvolution can be extended from gene expression data to ATM methylation profile data to ascertain the underlying cell type status, thereby providing an assessment of ATM. In certain embodiments, such deconvolution involves considering the methylation status of CpG sites as a weighted sum of the methylation status of the underlying cell type. Based on these insights, MethylCIBERSORT [Chakravarthy, 2018, Pan-cancer deconvolution of tumor composition using DNA methylation, Nature Comm 9(1):1-13 (incorporated by reference)] uses CIBERSORT applied to methylation sequencing data, while MethylResolver [Arneson, 2020, MethylResolver—a method for deconvoluting bulk DNA methylation profiles into known and unknown cell contents, Communications biology, 3(1):1-13 (incorporated by reference)] uses least trimmed squares regression for methylation deconvolution. The present invention employs such methods suitable for deconvolution of methyl-seq data. For example, in addition to modeling the deconvolution as a system of linear equations, the present invention may involve the use of machine learning classifiers that are trained on several bulk samples and then applied to the provided data.

[0072] The present disclosure provides a novel platform for non-invasive profiling of patient-derived TILs. In a preferred embodiment, non-invasive profiling of TILs involves decoding TIL-derived methylation signatures identified from plasma-derived cell-free DNA molecules. Such signatures can, for example, indicate disease status (e.g., remission, relapse, minimal residual disease). Signatures can also predict the durability and / or toxicity of a particular patient's expected response to a potential treatment. Similarly, signatures can indicate favorable ATM for disease progression. In the context of cancer, this can include evaluating methylation signatures from TILs and / or non-TME cells identified as originating from the tumor or tumor margin. Use of the detected profiles of margin-derived cells and / or non-TME cells can provide loci indicative of tumor progression or regression (e.g., via somatic regression). Such information provides crucial insights into treatment planning.

[0073] In certain aspects, the system of the present invention is used to deconvolute methylation data and build, contribute to, or use a methylation cell atlas, which may contain reference data for several different human cell types and conditions, using both in-house and public data resources (such as BLUEPRINT and ENCODE). The platform is useful for assessing immunotherapy response and toxicity. Embodiments use archived and de-identified biospecimens from individuals with known conditions (e.g., biospecimens available from Yale SPORE in Skin Cancer (YSPORE-SC)). These specimens, including melanoma biopsies, plasma samples, and peripheral blood leukocyte samples, are collected with participants' signed informed consent in accordance with Health Insurance Portability and Accountability Act (HIPAA) regulations.

[0074] As the methods of the present invention contribute to such atlases, the atlases may grow to encompass a wide range of cell types and / or cell states with methylation profiles that exhibit various ATMs that correlate with a particular disease or condition. For example, the atlases may include methylation patterns associated with inflammatory conditions (e.g., macular degeneration and rheumatoid arthritis), autoimmune diseases, transplanted tissues (including rejection and graft-versus-host conditions), pathogenic diseases, immune responses, and other conditions. Consequently, with a sufficient atlas, the methods of the present invention may be used as a general-purpose diagnostic for a few or many conditions.

[0075] Algorithms for deconvolving methylation data with high sensitivity and specificity can use any read assembly, alignment, or mapping tool. Many alignment tools for methylation data create BedGraph files as their final output, and most traditional deconvolution tools work with the BedGraph file format. See Hoang, 2014, CWig: compressed representation of Wiggle / BedGraph format, Bioinformatics 30(18):2543-2550 and Kent, 2010, BigWig and BigBed: enabling browsing of large distributed datasets, Bioinformatics 26:2204-2207 (both incorporated by reference). However, some embodiments of the present invention operate using sequence alignment map (SAM) or binary alignment map (BAM) file formats [Li, 2009, The sequence alignment / map (SAM) format and SAMtools, Bioinformatics 25:2078-2079 (incorporated by reference)], which may provide deeper read-level sequencing data. BAM files store aligned reads of a mixture. And because the read pairs (i.e., DNA fragments) of a mixture are generated from constituent single cells, fragment-based classification will apply more powerful detection power to mixture deconvolution. Therefore, aligned DNA fragments of a given mixture will preferably be classified using the fragment-derived SAM or BAM files.

[0076] In BedGraph files, for every CpG site, an average methylation value is calculated after aligning the reads to the corresponding reference genome. This is a reasonable approach, but it may miss aspects of the methylation patterns within the reads. On the other hand, BAM files contain aligned reads with read-level information, which allows us to capture read-level patterns and use these patterns to deconvolute mixtures at the per-fragment level.

[0077] Based on the above observations, the present invention provides a method for deconvoluting a methylated bulk mixture into a purified cellular state.

[0078] Figure 1 shows the steps of an exemplary method for deconvolving bulk methylation data into data representing refined cell states. First, differentially methylated regions (DMRs) are identified for a cell type and / or cell state of interest. These DMRs for each cell state serve as specific signatures. To generate these cell state-specific signatures, the cell state is compared to all other cell states in an atlas or database, traditionally known as a one-versus-rest comparison. However, for deconvolving multiple (≧20) cell states, some of which are closely related to each other, we provide a method for generating better signatures for each cell state.

[0079] To identify differentially methylated regions and create a signature matrix, first obtain methylation profiles of purified cell types and context-dependent cell states. These regions are investigated in bulk methylation mixtures to deconvolute closely related context-dependent cell types and states. Given a signature matrix of a cell state of interest, every read pair or fragment in the mixture is examined one by one. Considering that each sequencing read is obtained from a fragment derived from a single cell, it is preferable to use a deconvolution method that quantifies at the per-fragment level.

[0080] As fragments are tested, the fragment methylation patterns are tested for a match with a predefined signature matrix. If they match the signature of a specific cell state, the fragment is classified into that cell state. Finally, to obtain the cell fraction, all fragments classified into that cell state are counted and divided by the number of available fragments tested for that cell state. This method is referred to herein as read counting. As a proof-of-principle study, six refined cell types were obtained from the BLUEPRINT public database and 25 in silico mixtures were prepared. Read counting shows good correlation with the ground truth of these mixtures. See Fernandez, 2016, The BLUEPRINT Data Analysis Portal, Cell Syst 3(5):491-495 (incorporated by reference).

[0081] Figure 2 shows results demonstrating the performance of read counting. As a proof-of-principle using six cell types, 25 in silico mixtures were prepared where read counts show a strong correlation with known ratios. One crucial parameter of the read counting approach is how many CpG sites are desired to consider for each fragment. This is important because increasing the number of CpG sites required per fragment reduces the false positive rate.

[0082] Figure 3 shows that the false positive rate decreases as the number of CpG sites increases. However, if too many CpG sites are required per fragment, the number of available fragments may decrease, resulting in a very low sensitivity of the resulting deep deconvolution approach.

[0083] Figure 4 compares read counts with different numbers of CpG sites. Multiple CpG sites per fragment help achieve a lower detection limit (lower line in Figure 4), but the number of available fragments with multiple CpG sites decreases.

[0084] FIG. 5 shows that the number of available fragments with multiple CpG sites per fragment is low.

[0085] Figure 5, on the other hand, shows that when all available fragments (CpG≧1) are considered, there is a false discovery rate of over 80% (Figure 6) at cell fractions of less than 1%. Taking all these findings into account, the method can be optimized by taking the following measures: instead of setting any fixed minimum CpG cutoff, consider all fragments but weight fragments with more CpG sites. This approach will consider all fragments with the advantage of multiple CpG sites. Several different cell types and cell states can be deconvoluted. These cell types and states may be very closely related, and their signatures may be difficult to distinguish. In that case, it may be useful to consider the discrimination power of CpGs for a given fragment. The discrimination power of CpG sites lies in the fact that each such site creates and defines the distance between the cell type / state of interest and others.

[0086] Detection limit analysis indicates that there is significant noise below 1%. From a predefined signature matrix, it is possible to generate true positive rates (TPR) and false positive rates (FPR). Since theoretical predictions indicate that binomial modeling can be used to predict read counting outcomes, it is anticipated that noise can be reduced by incorporating the FPR of the signature matrix into the read counting algorithm using a Monte Carlo-based approach.

[0087] Figures 3 to 6 show the technical evaluation of read counts.

[0088] FIG. 3 is a box plot showing that the false positive rate (FPR) decreases with increasing number of CpG sites per fragment.

[0089] Figure 4 shows the results of a detection limit analysis for a cell type (neutrophil) within an in silico mixture. As the number of CpG sites increases, the FPR decreases, thus achieving a better detection limit. The theoretical prediction is shown as a dashed line. Figure 5 graphs the decrease in the number of available fragments as the CpG cutoff is increased. Figure 6 shows the change in FDR with cell fraction for different numbers of CpG sites per fragment.

[0090] The methods of the invention can use, and their results can contribute to, databases or atlases of cell state-specific methylation profiles that correlate with particular ATM types and ATM states. The methods can use (or copies of) existing such resources (e.g., including those named scMethBank, m5C-Atlas, DNAm-atlas, and MethAtlas).For example, Zong,2022,scMethBank:a database for single-cell whole genome DNA methylation maps,Nucleic Acids Research 50(D1):D380-D386,Ma,2022,m5C-Atlas: comprehensivea database for decoding and annotating the 5-methylcytosine(m5C)epitranscriptome,Nucleic Acids Res 50(D1):D196-D203, Zhu,2022,A pan-tissue DNA methylation atlas enables in silico decomposition of human tissue methylomes at cell-type resolution,Nat Meth 19:296-306, Loyfer,2022,A human DNA methylation atlas reveals principles of cell type-specific methylation and identifies thousands of cell type-specific regulatory elements,bioRxiv:477547(28 pages), Moss,2018,Comprehensive See human cell-type methylation atlas reveals origins of circulating cell-free DNA in health and disease. Nat Commun 9:5068, and Katsman, 2022, Detecting cell-of-origin and cancer-specific methylation features of cell-free DNA from Nanopore sequencing, Genome Biol 23:a158 (all incorporated by reference).

[0091] In the LiquidTME framework shown in Figure 1, the differentially methylated region or signature matrix is ​​another key component that the deconvolution algorithm uses as a reference for fitting. Traditionally, signatures for a cell type are generated in a one-versus-all manner, where all other cell types are considered together when preparing a reference for a cell type.

[0092] More specifically, cell types are divided into two groups: one group contains the cell type of interest, and the other contains the rest. While this technique is efficient and reasonable for a smaller number of distinct cell types, the one-versus-all approach can be problematic when the signature matrix contains a large number of closely related cell types and states (e.g., derived from public and / or in-house data). For the atlas, it may be useful to develop a methylation signature matrix (comprising several different cell types and states, many of which will be closely related). In this scenario, instead of using a one-versus-all approach, the method uses a hierarchical approach, in which existing biological information would be used. Specifically, certain steps are used to generate the methylation signature matrix.

[0093] Figure 7 shows the workflow for creating a signature matrix. In step 701, the workflow begins by considering the methylation profiles of different cell types (e.g., immune cells and somatic cells) and states (e.g., immune cell exhaustion or expansion) in the human body.

[0094] Cell types are first grouped based on their known biology. In step 701, all cell types and states will be grouped into smaller groups based on biological similarity, such that closely related cell types / states are grouped together. For example, CD4 T cell, CD8 T cell, and Treg cell states are all T cells and therefore grouped together in a single group. This group information will be user-defined (step 701 in Figure 7), so that the granularity can be adjusted based on the situation. In this group-based framework, every cell state belongs to a unique group. When a cell state profile is generated, there are two types of groups: the group containing the cell state is named the own group, and the other group is named the rest group. To generate a cell state profile, the methylation distances between the cell state and all the rest groups will be calculated separately.

[0095] In step 707, for any cell type / state (such as the last one listed in group 2 in step 701), the distance from each group is measured. For group 2, the distance for all cell states is measured separately because the last cell is biologically similar to the rest of the cells in group 2. In step 713, different distance and CpG number thresholds are used to obtain different candidate signatures for that cell. Step 721 provides the optimal signature for the cell state from the previous step, where the columns are cell types / states and the rows are CpG positions. Cells in the first column are mostly bright (from a high-low scale) because hypomethylation is used as the cell state-specific signature. Because all cell states in the group are closely related, they can be compared one by one (step 707 in Figure 7).

[0096] Third, after comparing all group-wise distances, the candidate signatures can be combined using different distance thresholds. These combinations are candidate signatures for cell states. To find the optimal signature, generate an in-silico mixture for training. Because testing all candidate signatures can be time-consuming, there is an option to adopt a subset of them based on the number of CpG sites available in the candidate signatures (step 713). Finally, identify the signature that maximizes performance within the in-silico training mixture (step 721).

[0097] Figure 8 shows results from LiquidTME-based deep deconvolution. LiquidTME is compared to wet lab ground truth for 10 immune cell subsets from the methylated cell atlas in peripheral blood from seven healthy human subjects. In this non-limiting example, flow cytometry or time-of-flight cytometry (CyTOF) is used for wet lab ground truth.

[0098] The above approach can be extended for use with all cell types and states (e.g., >50) in the atlas to generate a large signature matrix that can be referred to as a methylated cell atlas. The disclosed read counting method combined with the use of this methylated cell atlas is named LiquidTME, which demonstrates the ability to profile ATMs and immune cells in ATMs from liquid biopsy samples.

[0099] Using a database of cell state and its epigenetic markers, such as a methylated cell atlas, the method of the present invention can provide several types of predictive analysis using the methylation data from liquid biopsy samples.The method of the present invention preferably comprises identifying the epigenetic modification in the nucleic acid from the sample, such as by bisulfite sequencing or enzymatic methyl sequencing of the cfDNA from blood or plasma samples.The pattern of epigenetic modification is compared with the database or atlas, or is read by machine learning algorithms.

[0100] A database of cell states and their epigenetic markers, such as a methylated cell atlas, can be used to implement a method for predicting the efficacy of a treatment. The method of the present invention is useful for predicting the toxicity of immunotherapy (and, more generally, the toxicity of treatment) from cell-free DNA methylation cell state analysis. For example, from a liquid biopsy sample, the method of the present invention can use cfDNA to deeply profile the immune cell repertoire, which may include immune cell diversity, activation, and / or abundance. The measured abundance and / or diversity form the basis for predicting the toxicity or response of immunotherapy.

[0101] In certain preferred embodiments, the method involves measuring the abundance of activated CD4 T effector memory (TEM) cells to assess immunotherapy-related adverse events (irAEs) in patients with cancer. Toxicity prediction is based on the fact that abundant activated CD4 TEM cell levels are associated with the occurrence of severe irAEs. Certain embodiments may also measure TCR diversity from patterns of epigenetic modifications. Higher TCR clonotype diversity in bulk peripheral blood is understood to predict the occurrence of severe irAEs. Prediction of irAEs is made when one or both of TEM abundance and TCR diversity exceed a threshold in pre-ICI patients. A preferred embodiment uses a combined model that integrates features of both activated CD4 TM cell abundance and bulk TCR diversity. In some embodiments, the combined model produced an AUC of 1.0 (P=0.04) in bulk cohort 2 and an AUC of 0.86 by leave-one-out cross-validation (LOCCV) to predict severe irAEs.

[0102] For ground-truthing and cell state atlas construction, methods may include droplet-based scRNA-seq and time-of-flight mass cytometry of peripheral blood from patients treated with a particular therapy. In an exemplary embodiment, the method includes droplet-based scRNA-seq and time-of-flight mass cytometry of peripheral blood from cancer patients treated with combination immunotherapy (anti-PD1 / anti-CTLA4) to determine activated CD4 memory T cell subsets that best predict the occurrence of severe irAEs.

[0103] Time-of-flight cytometry and bulk RNA sequencing (deconvolved with CIBERSORTx and TCR assembly with MiXCR) can be performed to validate the combined model (abundance of activated CD4 memory T cells integrated with TCR diversity). CIBERSORTx is a machine learning tool that infers cell type-specific gene expression profiles. See Newman, 2019, Determining cell type abundance and expression from bulk tissues with digital cytometry, Nat Biotech 37:773-782 (incorporated by reference). MiXCR is a software package for immune profiling. See Bolotin, 2015, MiXCR: software for comprehensive adaptive immunity profiling, Nat Methods 12(5):380-1 (incorporated by reference). RNA sequencing and / or mass spectrometry with gene expression and immune profiling can be used to provide ground truth or gold standard data.

[0104] Using this gold-standard data, gene expression signatures of irAE toxicity and immunotherapy response can be inferred from promoter-level methylation in cell-free DNA, thereby demonstrating the feasibility of predicting both irAE severity and immunotherapy response from pretreatment plasma. This provides a liquid biopsy prediction of both immunotherapy response and toxicity from a single cell-free DNA assay. To validate such predictions, cell-free DNA is extracted from patients whose pretreatment plasma (cycle 1, day 1) is tested for ground truth validation (e.g., using scRNA-Seq, TOF mass cytometry, and bulk RNA-Seq), subjected to methylation sequencing, and queried against a gene set previously reported to be associated with immunotherapy toxicity (CD4 T5+35) and response (CE9 ecotype 32) (Lozano, 2022, Nat Med 28:353, incorporated by reference). Pretreatment promoter methylation signatures correlate with irAE severity and immunotherapy response. [Example]

[0105] Example Validation of LiquidTME The following example demonstrates the performance of the disclosed method used to accurately predict TIL content within ATM (in this case, tumor mass) noninvasively via plasma cfDNA analysis. To test the performance of LiquidTME, a less detailed version of LiquidTME was implemented, which combines states and performs group-wise comparisons instead of comparing all cell states separately. Methylation data for 22 refined cell states were collected from the BLUEPRINT public database [see Fernandez, 2016, The BLUEPRINT Data Analysis Portal, Cell Syst 3(5):491-495 (incorporated by reference)] and used to generate a signature matrix for these 22 cell states. Using the generated signature matrix or methylated cell atlas, seven actual bulk PBMC samples were used for wet-lab ground truthing of several cell states using flow cytometry or CyTOF (Figure 8). In addition to using wet-lab ground truth for multiple cell states, in-silico mixtures are useful for evaluating performance.

[0106] Figure 9 shows the signature for CD8 TILs. The columns are different cell types / states and the rows are CpG positions. CD8 TILs show a distinct pattern compared to other cell types and states.

[0107] Furthermore, this hierarchical signature matrix generation approach was used to generate signatures for TIL, specifically CD8 TIL.Purified CD8 TIL, tumor cells (MelTumor) from melanoma patients, and CD8 PBL (normal CD8 PBL from melanoma patients) were isolated and methylation sequenced.This example shows the performance of the method herein to accurately predict the TIL content in tumor mass non-invasively through plasma cfDNA analysis.

[0108] In addition to BLUEPRINT and in-house data, more cell types and states from other public resources are included and used in the preparation of a comprehensive methylation cell atlas. As discussed above, when signatures for closely related cell states are unclear and deconvolution approaches face challenges, problematic cell states are compared pairwise. In addition to generating profiles from methylation data, methylation profiles can be predicted from scRNA-seq data, if necessary. Studies have shown that there is a negative correlation between promoter methylation and gene expression levels. See Anastasiadi, 2018, "Consistent inverse correlation between DNA methylation of the first intron and gene expression across tissues and species," Epigenetics & chromatin, 11(1):1-17 (incorporated by reference). Because methylation data is collected from purified cells, predicting methylation profiles from scRNA-seq can be useful.

[0109] Non-invasive TIL profiling for multiple cancer types The method of the present disclosure can accurately profile TIL, and thus profile the ATM (tumor) of origin.This can include, for example, determining the type of cancer that produces TME that can be detected through the epigenetic modification found in cfDNA.In addition, TIL methylation profile can be used to provide prediction of the durability and / or toxicity of response to treatment (in this case, immunotherapy).

[0110] To assess whether noninvasive profiling of tumor-infiltrating leukocytes (TILs) has in vivo utility, it is important to compare the estimated TIL composition in plasma from colorectal cancer (CRC) tumors and melanoma patients with orthogonal measures of TIL content (e.g., by flow cytometry) in paired tumors. Viably preserved tumor, plasma, and PBL samples from 30 patients with advanced melanoma will be analyzed. Patients with matched epidemiological and clinical characteristics will be analyzed without intentionally excluding specific gender / sex or minority groups. A subset of patients underwent tumor biopsy and blood collection before treatment. A subset of patients with recurrent specimens will also be evaluated to allow for assessment of changes in TIL content from baseline. In parallel, archived whole blood samples (plasma and PBLs) from 10 age-matched healthy non-pregnant donors (in whom TILs should be absent) will be obtained from a local blood bank and processed, regardless of demographic characteristics or specific gender / sex. In addition to non-invasive TIL profiling, the methods of the present invention are applied to predict toxicity in melanoma patients.

[0111] Figure 10 shows how LiquidTME applied to CRC cfDNA, specifically blood plasma-derived cfDNA, compares to ground truth TIL content (measured by analysis of tumor biopsies).

[0112] Figure 11 shows the correlation of LiquidTME-estimated cfDNA-derived TIL content with paired tumor biopsy results. TIL levels in tumors are obtained by standard FACS and SLD imaging techniques.

[0113] As a proof-of-principle study, TIL signatures were obtained using the disclosed method and tested against cfDNA from six CRC patients for whom matched bulk tumors were available. It was hypothesized that TIL content in bulk tumor tissues would correlate with cfDNA levels quantified by the LiquidTME approach. Indeed, this method demonstrates a significant positive correlation with wet-lab ground truth (Figure 11).

[0114] Next, the method of the present disclosure is applied to the plasma cfDNA of 23 melanoma patients to non-invasively detect TIL content and predict the response to immunotherapy.All these patients have been diagnosed with metastatic melanoma and have been receiving immunotherapy.The results show that ctilDNA is higher in patients who respond to treatment (Figure 13).

[0115] We model this response prediction task as a machine learning problem to improve performance in larger cohorts. This allows us to use additional cell-free DNA features (e.g., fragment length) to improve response prediction. Studies have shown that the distribution of ctDNA fragment lengths differs from that of healthy cfDNA. Studies have also shown that ctDNA fragments have different terminal motifs than healthy cfDNA. See Underhill, 2016, Fragment length of circulating tumor DNA, PLoS genetics, 12(7):e1006162; Mouliere, 2018, Enhanced detection of circulating tumor DNA by fragment size analysis, Science Trans Med 10(466):eaat4921; Cristiano, 2019, Genome-wide cell-free DNA fragmentation in patients with cancer, Nature, 570(7761):385-389; and Mathios, 2021, Detection and characterization of lung cancer using cell-free DNA fragmentomes, Nature Comm 12(1):1-14 (all incorporated by reference). The distribution of ctDNA fragment lengths is found to differ as it also originates from the TME and therefore differs from healthy cfDNA. Fragment length can be integrated as another feature in predictive models.

[0116] Toxicity prediction In this example, a method for predicting immune-related adverse events was tested. The method of the present invention can determine the TCR repertoire for ATM, and this information can be used to predict the durability and toxicity of treatment response. In this non-limiting example, immune-related adverse events were tested using samples from a cohort of 15 patients. This cohort of 15 patients had previously undergone bulk TCR beta repertoire profiling in peripheral blood monocyte (PBMC) samples by immunoSEQ and RNA-Seq, as described in Lozano, 2022, Nat Meth 28(2):353-362 (incorporated by reference). These results provide a ground truth for testing the results of LiquidTME methodology.

[0117] Immunotherapy can be toxic, known as immune-related adverse events (irAEs). A recent study in Nature Medicine showed that higher levels of activated CD4 memory T cells in the peripheral blood of melanoma patients predicted higher rates of severe irAEs. See Lozano, 2022, T cell characteristics associated with toxicity to immune checkpoint blockade in patients with melanoma, Nature Med 28(2):353-362 (incorporated by reference).

[0118] The present invention includes an ultrasensitive deconvolution method and a comprehensive methylated cell atlas for classifying closely related cellular states. The disclosed method can detect activated CD4 memory T cell levels from cfDNA. To test this, toxicity information was collected from 15 patients in the Lozano 2022 melanoma cohort and the LiquidTME method was applied. The method described herein provides a framework for inferring constituent cell proportions from cell-free DNA methylation data. By developing a novel deconvolution algorithm in conjunction with a comprehensive methylated cell atlas, the method of the present invention can, for the first time, noninvasively profile the TME. These methods are feasible given these preliminary experiments.

[0119] Figures 12-14 show application to melanoma cfDNA.

[0120] Figure 12 shows the application of LiquidTME to patient cfDNA to predict response and toxicity before initiation of ICI in melanoma patients.

[0121] FIG. 13 is a boxplot and ROC plot showing prediction of patient response.

[0122] Figure 14 shows box plots and ROC plots for toxicity prediction using CD4 memory cells estimated by LiquidTME. Box plot p-values ​​calculated by two-sided MWU test.

[0123] Figure 14 shows that LiquidTME predicts higher CD4 memory cells in patients with severe toxicity. Based on this finding, LiquidTME is consistent with other sequencing approaches, with results predictive of severe irAEs.

[0124] Thus, the methods of the present invention use a computational liquid biopsy framework to facilitate more accurate and personalized cancer treatment, including the prediction of immunotherapy response and toxicity.

[0125] In certain embodiments, the present invention utilizes the insight that clonally diverse activated CD4+ memory T cells, more specifically CXCR5-PD1hi peripheral helper T (Tph) cells, specifically support ICI-mediated toxicity in melanoma patients. To address this hypothesis, we performed CyTOF, scRNA-seq, scV(D)J-seq, and immunoSEQ to comprehensively assess T and B cell status in blood before and during treatment, thereby (1) determining whether pretreatment blood Tph levels predict the occurrence of severe irAEs in melanoma patients treated with combination immunotherapy and whether Tph clonotypes preferentially expand during treatment with combination immunotherapy in patients who develop severe toxicity.

[0126] While predicting severe irAEs from peripheral blood is clinically important, it has also been shown that patients who experience some degree of toxicity have a better and more durable immunotherapy response rate. Therefore, making clinical decisions regarding immunotherapy without also considering the likelihood of a durable response can be difficult. The present invention alleviates these difficulties by using cell-free DNA methylation sequencing to simultaneously predict 1) immunotherapy toxicity and 2) durable immunotherapy response from pretreatment plasma using both cell state signatures and an agnostic machine learning approach (validated in the provided cohort (Objective 3)). By doing so, the present invention provides a tool that can be used to lay the foundation for future clinical trials in which immunotherapy decision-making is guided by the risks versus benefits of combination immunotherapy using the liquid biopsy biomarkers defined herein.

[0127] Severe, life-threatening, or fatal immune-related adverse events (irAEs) occur in approximately 60% of melanoma patients treated with combination immune checkpoint inhibitors (ICIs). While recent data suggest that autoreactive lymphocytes play a key role in promoting irAEs, the pathophysiology underlying the development of severe irAEs remains unclear. This lack of knowledge has led to an absence of a method for predicting who will experience these severe toxicities in clinical practice, thereby resulting in unacceptably high treatment-related morbidity and mortality. Furthermore, there is no minimally invasive clinical assay that can simultaneously predict the risk of irAE development and the benefit of immunotherapy response before treatment, making treatment individualization difficult. According to one embodiment of the present invention, CXCR5-PD1hi peripheral helper CD4 T cells (Tph) are critical determinants of future toxicity in melanoma patients receiving ICIs; they are functionally involved in irAE development and are applicable to both toxicity and early response assessment by cell-free DNA methylation profiling. Clonally diverse activated CD4+ memory T cells are significantly elevated in pretreatment blood from melanoma patients who develop severe or life-threatening irAEs after anti-PD1 and anti-CTLA4 combination therapy (Nature Medicine, 2022). Data from 13 melanoma patients found that these same T cells have a gene expression profile enriched for Tph cells (a circulating T cell state elevated in autoimmune disorders such as rheumatoid arthritis, lupus, and type 1 diabetes), but this has yet to be rigorously evaluated in human irAE development.

[0128] Furthermore, plasma cell-free DNA (cfDNA) methylation profile data from 21 melanoma patients revealed distinct signatures of immunotherapy response and toxicity, the latter reflecting activated CD4+ memory T cells in the blood.

[0129] Here, we treated peripheral blood from 100 melanoma patients with anti-PD1 / anti-CTLA4 dual therapy to determine whether Tph cells underlie the occurrence of severe irAEs at baseline, whether they are associated with clonal expansion during treatment, and whether they can be utilized in conjunction with tissue-based signatures to determine ICI response and toxicity from baseline cfDNA. Two innovative approaches underpin this work. First, we developed a key, feasible platform for benchmarking Tphs from pretreatment blood to predict irAE development in a graded, organ-system-independent manner (Nature Medicine, 20226). Previous studies predicted irAEs from a single organ system, either poorly generalized to multiple organs (AUC < 0.68) or failed to identify life-threatening grade 4 irAEs. Second, we used machine learning methods to profile cellular ecosystems at scale to identify seven cellular ecosystems that better predict ICI response than over 100 competing measures, including specialized biomarkers, and are detectable from cfDNA. Previous biomarkers of ICI response have been limited to small cohort sizes, single genes, bulk signatures, or direct tissue biopsies.

[0130] Tph cells underlie particularly severe irAEs (but not responses) in melanoma patients treated with combination ICIs. To address this, we performed mass cytometry on pretreatment PBMCs prospectively obtained from 100 melanoma patients treated with combination ICIs, and performed scRNA-seq / scV(D)J-seq on 50 of these samples. The frequency of cellular states was compared with the incidence of irAEs and responses.

[0131] Activated CD4+ memory T cell clonotypes, specifically Tph cells, preferentially expand clonally before the onset of severe irAEs. This may provide mechanistic insight, as any T cell state that preferentially expands before irAE onset is likely involved in irAE development. Therefore, we performed paired scRNA-seq / scV(D)J-seq early in treatment (day 1 of cycle 2) from the same 50 patients profiled pretreatment by scRNA-seq in Aim 1. To validate these findings, we performed immunoSEQ on the same T cell states from pretreatment and early treatment blood from the remaining 50 melanoma patients in Aim 1.

[0132] Based on data from 21 patients, we hypothesized that ICI toxicity and response signatures could be inferred from pretreatment cfDNA methylation profiles. Therefore, we applied NEBNext enzymatic methyl-seq to pretreatment cfDNA from the same 100 patients in Aim 1. After splitting into training / validation cohorts, we evaluated published toxicity and response gene signatures, refined toxicity signatures, and an independent machine learning strategy to identify predictive signatures.

[0133] This study sheds light on irAE biology and novel techniques for improving melanoma outcomes. Additionally, the study validated the utility of ATM-derived cfDNA methylation profiles to accurately predict treatment response and / or toxicity.

[0134] Severe and debilitating toxicities known as immune-related adverse events (irAEs) are common side effects of numerous types of treatments for myriad conditions. This is especially true in the setting of dual immune checkpoint blockade, a paradigm-shifting treatment that has revolutionized outcomes for patients with melanoma, with 100,000 new cases diagnosed annually in the United States. The standard of care for patients with metastatic melanoma is dual anti-PD1 / anti-CTLA4 blockade (hereafter referred to as combination ICI), which results in better responses and improved survival rates compared with single-agent ICI. Unfortunately, combination ICI can cause severe and potentially life-threatening irAEs (grade 3+) in up to 60% of ICI-treated melanoma patients. This can lead to premature discontinuation of anticancer treatment, hospitalization, admission to intensive care units, and even death. ICI-induced irAEs affect various organ systems, including the lungs, heart, joints, thyroid, pituitary gland, liver, colon, nervous system, and skin. There are no clinical assays to determine which patients will experience severe irAEs before treatment, and the mechanisms underlying their development are not well understood.

[0135] Life-threatening (grade 4) and fatal (grade 5) irAEs occurred at rates of 11% and 1.2% in patients treated with dual ICIs, respectively, and 5% and 0.4% in patients receiving anti-PD1 monotherapy, respectively. Fatal irAEs typically occur early after the initiation of immunotherapy, with a median onset of only 14.5 and 40 days after the initiation of combination ICIs and anti-PD1 monotherapy, respectively. Unfortunately, no clinical assays exist that can determine who will experience life-threatening or fatal (grade 4-5) irAEs. Indeed, elucidating the biological basis of grade 4+ irAEs may help clinicians select alternative treatment options (i.e., withholding adjuvant immunotherapy in stage III melanoma after surgical resection, or prioritizing anti-PD1 treatment, which has a lower severe irAE rate than combination ICIs) in patients at risk of ICI-induced mortality.

[0136] Several groups have investigated potential biomarkers of ICI-induced toxicity based on blood or tumor analyses. However, these studies generally focused on early treatment prediction or a single organ system, with only limited evidence of organ-system-independent pretreatment performance and no demonstrated ability to discriminate between multiple irAE grades. More recently, serum antibody approaches have been used to predict severe ICI-induced irAEs before treatment in melanoma. However, these signatures differed significantly based on the immunotherapy regimen used, and there was no evidence that they could discriminate between multiple irAE grades. Thus, previous studies have not identified a common immune state preceding irAE occurrence that has the ability to predict irAE severity independently of organ systems and ICI regimens.

[0137] Herein, we show that clonally diverse activated CD4 T effector memory cells are strongly associated with the onset and severity of irAEs, providing implications for identifying the cellular basis of irAE development.

[0138] Melanoma patients who developed severe immunotherapy-related toxicity had significantly higher levels of clonally diverse, activated CD4 T effector memory cells in their peripheral blood before treatment. Furthermore, these cells preferentially expanded during treatment in an exploratory cohort of patients who developed severe toxicity after combination ICI treatment. The abundance of clonally diverse, activated CD4 T effector memory cells remained a significant pretreatment predictor of toxicity onset and severity, independent of organ system involvement.

[0139] A major barrier to improving outcomes in melanoma is the inadequate ability to simultaneously determine the likelihood of both ICI response and irAE development prior to treatment. While predicting immunotherapy-induced toxicity is crucial and can help prevent substantial morbidity and even death, it has been shown that patients who experience toxicity are more likely to respond to treatment. Therefore, making important clinical decisions using assays that only determine the probability of toxicity without considering response can be challenging, complicating risk-benefit calculations. Cell-free DNA (cfDNA) liquid biopsy is a proven approach for minimally invasive tumor profiling and can now be applied to investigate tumor and blood cellular states that predict clinical outcomes.

[0140] Cell-free DNA, continuously released into blood plasma, is a useful analyte for measuring a variety of physiological and pathological conditions. While tumor-derived cfDNA has been shown to decrease from pre- to mid-treatment in durable ICI responders, most cancer-focused cfDNA assays measure somatic mutations and are limited to cancer cells. In contrast, epigenomic changes, such as methylation levels, can be used to detect cell type of origin from cfDNA and are generally applicable to cell state profiling. While pre-treatment cfDNA assays have not been shown to predict both ICI response and irAE risk, our novel pilot data demonstrate proof-of-concept capabilities for addressing this challenge.

[0141] As a proof-of-concept, the disclosed method was used to determine whether CXCR5-PD1hi peripheral helper CD4 T cells (Tph)—a critical subset of CD4 memory T cells involved in diverse manifestations of autoimmune diseases—underlie the development of severe irAEs, whether they are associated with irAE-associated clonal expansion, and whether they can be used in conjunction with tissue-based signatures to predict ICI response and toxicity from pretreatment cfDNA. Tph cells were identified as a likely phenotype underlying the relationship between activated CD4 memory T cells and irAE development. Additionally, a proof-of-principle study was conducted using whole-genome cfDNA methylation profiling from the pretreatment plasma of 21 melanoma patients, identifying novel cfDNA signatures of ICI response and irAE development.

[0142] If successfully completed, this will advance our understanding of the cellular conditions underlying irAEs, which in the future may facilitate therapeutic targeting of this population to prevent or reverse ICI-induced irAEs without sacrificing efficacy, facilitating clinical application via cfDNA assays that allow prediction of both ICI toxicity and response.

[0143] The scientific premises underlying this study include the following: First, clonally diverse activated CD4+ memory T cells in pretreatment peripheral blood are associated with irAE development in patients with melanoma receiving ICI therapy, independent of organ system. Second, CD4+ memory T cells consist of diverse phenotypic subsets, including Tph cells, which are involved in several autoimmune disorders such as rheumatoid arthritis, lupus, and type 1 diabetes, but these have not yet been rigorously evaluated in human irAE development. Third, elucidating the T cell phenotypic state(s) that preferentially expand during treatment with respect to irAE development can identify the phenotypic state(s) underlying irAE development. Fourth, minimally invasive cfDNA methylation profiling can be used to investigate the composition of tumor- and blood-derived cellular states from peripheral blood plasma.

[0144]

[0003] Method embodiments of the present invention teach the use of circulating T cell biomarkers to predict the risk of severe irAEs from pretreatment blood. Previously reported features associated with irAE occurrence have shown modest performance in predicting irAEs from pretreatment samples (AUC<0.68); have only been described for a single organ system; lacked clear mechanistic insight (e.g., autoantibodies); or have not demonstrated predictive value for irAE grade. A biomarker combining the abundance of activated CD4+ memory T cells and T cell receptor diversity can predict severe irAEs from pretreatment blood of melanoma patients with high accuracy (AUC 0.90 in patients treated with concomitant ICIs), is organ system independent, biologically interpretable with mechanistic significance underlying this proposal, and can also differentiate irAE grade (including distinguishing between life-threatening grade 4 irAEs and non-life-threatening but severe grade 3 irAEs).

[0145] In cancer, a complex ecosystem of interacting cell types forms powerful signaling networks that shape tumor development. While single-cell genomics, spatial transcriptomics, and multiplexed imaging can provide a high-resolution picture of tumor cellular ecosystems, practical considerations limit the scale, scope, and depth of these assays. To overcome these limitations, we developed a machine learning framework for large-scale identification of cellular states and ecosystems from single-cell, bulk, and spatially resolved expression data. The machine learning framework identified a novel cellular ecosystem that was localized to the tumor center in carcinomas and melanomas; comprised of seven lymphoid and myeloid lineage states; correlated more strongly with ICI response than 121 competing measures, including dedicated biomarkers; and was strongly associated with response to combination ICIs when measured in a pilot study of 21 pretreatment cfDNA methylation profiles from melanoma patients.

[0146] Single-cell profiling of pretreatment blood reveals two T cell signatures—activated CD4+ memory T cells and T cell receptor (TCR) diversity—associated with the development of severe irAEs in patients with melanoma. In a recently published study, high-dimensional single-cell profiling was applied to a discovery cohort of pretreatment blood samples from 18 patients with advanced melanoma, eight of whom experienced severe irAEs after initiating ICIs.

[0147] Figures 15 and 16 show analysis of pre-treatment peripheral blood for cellular determinants of severe irAEs.

[0148] Figure 15 shows the relative abundance of 20 cell states by CyTOF in 18 patients and their association with irAE occurrence. Figure 15 shows that by CyTOF, we found that elevated levels of CD4 effector memory T cells (TEM) cells were significantly associated with severe irAE occurrence.

[0149] Figure 16 shows cell state abundance (scRNA-seq) versus future irAE status and CD4 TEM cell frequency (CyTOF) in the same patients. Figure 16 confirms these findings with scRNA-seq, showing that CD4 TEM cells expressing activation markers are more strongly associated with severe irAEs. Given these results, we hypothesized whether pretreatment TCR diversity might also correlate with severe ICI-induced irAEs.

[0150] Figure 17 shows that single-cell TCR clonotype diversity (Shannon entropy) of activated CD4 TEM cells was elevated in patients who experienced severe irAEs (AUC = 0.90, P = 0.05). This association was driven by TCR abundance, i.e., the number of unique clonotypes relative to total PBMCs. While this relationship was weaker or absent in other T cell subsets, when combined with all evaluable T cells, a significant trend was observed between bulk TCR diversity and severe irAE occurrence, driven by CD4 TEM and TEM-like T cells (AUC = 0.80). These findings suggest that a more diverse TCR repertoire at baseline in CD4 TEM cells, broadly reflected in bulk peripheral blood, is associated with the occurrence of severe irAEs. Figure 17. TCR diversity by scV(D)J-seq. Pretreatment TCR clonotype diversity in activated CD4 TEM, bulk CD4 T cells, and bulk T cells stratified by irAE status.

[0151] Bulk RNA-seq profiling of pre-treatment blood confirms the relationship between activated CD4 memory T cell levels, TCR diversity, and severe irAE occurrence.

[0152] Using bulk RNA sequencing to analyze pre-treatment blood from 53 additional melanoma patients treated with ICIs and divided into two bulk cohorts (1 and 2), we queried immune composition using CIBERSORTx66 and TCR clonotype diversity using MiXCR67.

[0153] Figure 18 shows that of the 13 cellular states queried by CIBERSORTx, only activated CD4 memory T cell levels were associated with severe irAE occurrence. Furthermore, higher TCR clonotype diversity in bulk blood predicted severe irAE occurrence across organ systems.

[0154] Next, we used integrated modeling to predict the risk and grade of pretreatment blood-borne irAEs. We hypothesized whether a model integrating the abundance and TCR diversity of activated CD4+ memory T cells from bulk RNA-seq data might be superior to either feature alone.

[0155] Figure 18 shows pre-treatment activated CD4 memory T cell levels and TCR clonotype diversity versus irAE occurrence in 53 patients.

[0156] Figure 19 shows the relationship between the combined model and the highest irAE grade.

[0157] Figure 19 shows that using a logistic regression framework to train a bivariate model, the resulting pretreatment model highly predicts severe irAE occurrence across organ systems (AUC = 0.90, P = 0.0004), including patients treated with anti-CTLA-4 / anti-PD1 combination therapy (AUC = 1.0, P = 0.04) and distinguishes between different irAE grades (Figure 19). The model still predicts severe irAE occurrence across clinical and epidemiological subgroups and is not significantly associated with durable clinical benefit, highlighting its specificity for irAE biology. To test whether the pretreatment model can predict time to severe irAE, patients were assigned to high versus low groups by defining optimal cutpoints in bulk cohort 1. In the provided bulk cohort 2, patients in the high group experienced a severe irAE within a median of 1.74 months after starting treatment, whereas the majority of patients in the low group never experienced a severe irAE (P<0.0001, hazard ratio [HR]=11.6).

[0158] Figure 20 shows time to severe irAE in concomitant ICI patients stratified by composite model score.

[0159] Figure 20 shows that similar results were observed for each ICI type separately, whether evaluated in bulk cohort 2 (P<0.025, HR=8.3 and 14.8 for combination and PD1, respectively) or across cohorts by leave-one-out cross-validation (LOOCV) (P=0.0028 and HR=12.2 for combination therapy, Figure 20; P=0.03 and HR=9.0 for PD1 treatment). This candidate biomarker also predicted time to severe irAE in a multivariate model, independently of ICI type, age, sex, and other clinical parameters (P=0.005, Z=2.81).

[0160] TCR clonal expansion after ICI initiation correlated with the development of severe irAEs, suggesting a role for CD4 TEMs in irAE pathogenesis. Here, we used immunoSEQ to profile the bulk TCR-β repertoire in paired pre- and early-treatment PBMC samples from 15 melanoma patients treated with combination ICIs. Figure 21 shows TCR clonal dynamics associated with severe irAE development in patients receiving combination ICIs. Figure 21 demonstrates that increased TCR clonal expansion was preferentially associated with severe irAEs. We also performed scRNA-seq and scTCR-seq in three of these severe irAE patients and found preferential expansion of the activated CD4 TEM compartment among clones detected in both blood draws. As outlined in Aim 2, we will extend this analysis to a larger cohort to further elucidate whether specific T cell states preferentially expand upstream of clinical irAE development.

[0161] As a first step to verify the findings described herein, 24 new melanoma patients were recruited from Yale University and Washington University, and blood was drawn on day 1 of cycle 1. Sixty-three percent of patients were treated with combination ICIs, and the remaining patients were treated with anti-PD1 monotherapy. Of these patients, eight developed severe irAEs across nine organ systems (including four with life-threatening irAEs), while 16 did not. Patients treated with combination ICIs who did not develop severe irAEs had lower circulating CD4 T cells measured by CyTOF compared with healthy controls. Figure 22 shows preliminary prospective validation and the relationship between pretreatment Tph levels and severe irAE status. Figure 22 shows CD4 T cells measured by CyTOF stratified by ICI regimen and irAE status. Figure 23 shows CD4 TEM expression profiles (mean log2 fold change vs. other CD4 subsets) from our published scRNA-seq data compared to the expected Tph profile.

[0162] Figure 24 shows Tph levels measured by CyTOF and stratified as in A. Group comparisons were performed using a two-sided Wilcoxon test. Figure 22 shows that patients who developed severe irAEs had significantly higher CD4 TEM levels. Indeed, dividing the cohort into two groups based on median CD4 TEM levels confirmed that patients with low CD4 TEM had significantly better freedom from severe irAEs than patients with high CD4 TEM levels (P=0.013; HR=6.7).

[0163] CXCR5-PD1hi Tph cells are a pathogenic CD4 T cell state implicated in several autoimmune disorders, including rheumatoid arthritis, lupus, type 1 diabetes, IgA nephropathy, and IG4-related disease. Using pretreatment blood scRNA-seq data from a recent study, we examined activated CD4 TEM cells as a correlate of severe irAE occurrence. Figure 23 shows activated CD4 TEM cells correlated with severe irAE occurrence, with expression profiles remarkably similar to those of Tph cells. As such, we hypothesized whether profiling Tph cells might improve irAE prediction beyond quantification of CD4 TEM levels alone. While Tph cells could not be assessed in published CyTOF cohorts due to the lack of necessary markers, a novel CyTOF panel was designed to delineate these cells in detail.

[0164] Elevated Tph cell levels in baseline blood strongly predict severe irAEs in pilot data. Within CD4 memory T cells, Tph markers (CXCR5-PD1hi) were highly enriched by CyTOF. Figure 24 further shows that pretreatment CD4 memory T cell Tph levels correlate more strongly with severe irAE occurrence than CD4 TEM and may distinguish between grade 4 and 3 irAEs.

[0165] Pretreatment cfDNA methylation profiles are promising for predicting ICI toxicity and response. To determine whether future ICI toxicity and ICI can be detected from baseline cfDNA, we extracted pre-ICI cfDNA from the plasma samples of 21 patients from our retrospective melanoma cohort treated with combination ICI. Figure 25 shows pretreatment cell-free DNA analysis for simultaneously predicting response and toxicity. Figure 25 shows a schematic diagram showing activated CD4 TEM (for irAE prediction) and cell-free DNA derived from the tumor microenvironment (CE9; for response prediction). Figure 26 shows that activated CD4 TEM score and CE9 score correlated with toxicity and response status, respectively.

[0166] Figure 27 shows an inverted outcomes analysis (compare with Figure 26) to query biomarker specificity. * *P<0.05, two-tailed Wilcoxon test. Figure 25 shows whole-genome enzymatic methylation (EM)-seq performed to a median coverage of 30' and analyzed promoter methylation levels of genes associated with ICI toxicity and response. For the former, we evaluated the top 20 signature genes of CD4 TEM cells profiled by scRNA-seq from a recent study.

[0167] Figure 26 shows that patients treated with concomitant ICIs who experienced severe irAEs had significantly elevated CD4 TEM scores (AUC = 0.84) from this promoter-level methylation analysis compared with patients who did not experience severe irAEs. Furthermore, patients who experienced durable clinical benefit from concomitant immune checkpoint blockade had significantly higher CE9 scores derived from plasma cell-free DNA compared with non-responders (AUC = 0.84). Figure 27 demonstrates specificity by showing that CE9 did not predict toxicity and that CD4 TEM scores were not associated with immunotherapy response. These data suggest that pretreatment cell-free DNA methylation analysis can be used to simultaneously predict both immunotherapy response and toxicity by specifically interrogating the corresponding cellular states evident in plasma cell-free DNA derived from tumor tissue and blood, respectively.

[0168] These data support the notion that pretreatment blood samples from melanoma patients who develop severe irAEs are commonly enriched for activated CD4+ memory T cells, including Tph cells, and that the latter may underlie irAE development. These results were validated and refined in a larger cohort of melanoma patients, and cfDNA-interrogated blood- and tissue-based signatures robustly predict ICI toxicity and response in de novo patients.

[0169] The above data demonstrate that: (i) clonally diverse activated CD4 TEM cells are enriched pretreatment in the peripheral blood of melanoma patients at risk for severe toxicity from immune checkpoint blockade; (ii) activated CD4 TEM cells appear to undergo preferential clonal expansion during treatment in melanoma patients treated with concomitant ICIs who develop severe irAEs; (iii) CXCR5-PD1hi Tph cells outperform CD4 TEM cells in predicting the occurrence of severe irAEs in a newly enrolled melanoma cohort and have the potential to distinguish between severe and life-threatening irAEs; and (iv) cell-free DNA methylation profiling is a promising approach for simultaneously predicting immunotherapy toxicity and response from pretreatment plasma.

[0170] It will be determined whether circulating pre-treatment CD4 Tph cells are associated with irAE risk (Figure 28), whether they preferentially expand during treatment in irAE patients (Figure 29), and whether liquid biopsy assays have utility for simultaneous assessment of ICI toxicity and response (Figure 30).

[0171] FIG. 28 is a schematic diagram in melanoma (n=100) to determine whether CD4 Tph best predicts severe irAE development.

[0172] Figure 29 is a schematic illustrating the use of the present invention to determine which T cell subpopulations are preferentially expanded in patients with severe irAE. Figure 30 is a schematic illustrating the use of plasma cell-free DNA methylation to predict ICI toxicity and response.

[0173] A multicenter cohort of 100 patients with advanced, unresectable melanoma treated with combination ICI (anti-PD1 / anti-CTLA4) (standard of care, see Section A.1) was recruited from Yale Cancer Center and Washington University Siteman Cancer Center. Peripheral blood samples were collected pre- and early-treatment with ICI. In the first study, droplet-based scRNA-seq and time-of-flight mass cytometry were performed on all major pre-treatment PBMC populations to determine whether baseline levels of Tph cells best predicted the occurrence of severe irAEs. Next, paired scRNA-seq / scV(D)J-seq of pre- and early-treatment blood samples was performed along with immunoSEQ TCR sequencing to determine whether Tph (or another T cell population) preferentially expanded during treatment in patients experiencing severe irAEs. Finally, we determined whether pre-treatment cell state-derived signatures or independent machine learning could effectively predict ICI toxicity and response using a novel cfDNA methylation assay. All results were correlated with the incidence, grade, and timing of irAEs.

[0174] Peripheral blood samples from melanoma patients were collected from Yale Cancer Center and the University of Washington Siteman Cancer Center. Therefore, peripheral blood samples were readily available for all three purposes of this study within the first 1–2 years, with a follow-up period sufficient for the proposed analyses. PBMC and plasma samples were collected from all patients before treatment (day 1 of cycle 1) and during treatment (day 1 of cycle 2). Clinical data collection included age, sex, race, histological subtype, disease stage, irAE severity (CTCAE v5), timing of irAE, organ system(s) affected by irAE, durable clinical response, number of ICI cycles, overall survival, progression-free survival, and center. We determined whether pretreatment blood Tph cell levels predicted severe irAE occurrence in melanoma patients treated with anti-PD1 / anti-CTLA4 therapy.

[0175] The diversity and predictive power of irAE-associated factors across melanoma patients has important implications for the diagnostic and therapeutic breadth likely required to maximize ICI efficacy. Based on preliminary data, we hypothesized that elevated baseline levels of Tph cells underlie severe irAEs (but not responses), particularly in melanoma patients treated with combination ICIs. An alternative hypothesis is that Tph cells are not specifically associated with irAEs, and other cell subsets, including parental CD4 memory T cell populations, may be more predictive. To address this, we performed mass cytometry on pretreatment PBMCs prospectively obtained from 100 melanoma patients treated with combination ICIs, and performed scRNA-seq / scV(D)J-seq on 50 of these samples. We compared the frequency of cellular states with irAE incidence and response to determine which hypothesis is correct.

[0176] Several risk groups are important in melanoma pathogenesis. For example, melanoma can be clinically subdivided into cutaneous, mucosal, and ocular subtypes, with ocular melanoma generally having a poorer prognosis and a lower response rate to immunotherapy. Separately, melanoma is approximately 20 times more common in whites than in blacks, and metastatic melanoma is more common in men than women. Studies have also shown that patients experiencing irAEs, especially those related to certain organ systems (i.e., skin), have a higher immunotherapy response rate. Extensive clinical data are collected for each subject, and the inventors consider known covariates (e.g., age, race, histological subtype, durable response status) and the site of organ tissue toxicity within this scope.

[0177] Peripheral blood was collected from 100 melanoma patients treated with combination ICI (anti-PD1 / anti-CTLA4). Peripheral blood was collected into approximately two K2EDTA Vacutainer tubes (approximately 20 mL) (Becton Dickinson) before the start of combination ICI in melanoma patients (day 1 of cycle 1) and processed within 1 hour of phlebotomy. The expected yield is approximately 20 million PBMCs from approximately 20 mL of blood. PBMCs were isolated using Lymphoprep (Stem Cell Technologies) according to the manufacturer's instructions, resuspended in freezing medium (90% FBS / 10% DMSO), and cryopreserved in 10% dimethyl sulfoxide / 90% fetal bovine serum in Mr. Frosty containers (Nalgene) at -80°C for 24 hours, then stored in liquid nitrogen until further cell processing.

[0178] Thaw the cryopreserved pre-treatment PBMC sample by holding the cryovial without submerging the cap in a 37°C water bath for 1-2 minutes. Approximately 3 × 10 PBMCs in single-cell suspension were then incubated in Human TruStain FcX (BioLegend) for 10 minutes at room temperature to block nonspecific antibody binding, followed by incubation with Cell-ID Cisplatin (Fluidigm) and 50 metal-conjugated antibodies (Fluidigm) against cell surface and intracellular molecules (including 38 recently described + additional markers for central and effector memory T cells (i.e., CD62L, CCR7, CD27), T cell exhaustion (i.e., PD1, TIGIT), and CD4 T cell subsets (i.e., Th1, Th2, Th17, T peripheral helper [Tph] (including CXCR5-PD1hi Tph cells)) according to the manufacturer's instructions, similar to the method we utilized to generate our recent pilot data (Figure 24). This may help determine whether activated CD4 TEMs predictive of severe toxicity are preferentially CD4 Th1 / 2 / 17 / ph.

[0179] Cells are then washed, stained with Cell-ID Intercalator-IR (Fluidigm), diluted in PBS containing 1.6% paraformaldehyde (Electron Microscopy Sciences), and stored at 4°C until acquisition. After the washing step, sample acquisition is performed using a Helios System (Fluidigm) with an event rate of <400.

[0180] To reduce technical variability between samples, Ce beads were used in each sample and output files were normalized collectively using Bead Normalizer v0.3. To further minimize technical variability, sample processing and acquisition batches were limited, identical reagent lots were used across all samples, and no major adjustments were made to Helios calibration settings between sample runs.

[0181] CyTOF data are analyzed using Cytobank v9.4 (Beckman Coulter) using the FlowSOM algorithm for hierarchical cluster optimization and the viSNE algorithm for visualization of high-dimensional data.75,76 Identification of cell subpopulations and visualization of data are performed by manual gating with standard markers using Cytobank v9.4.

[0182] Single-cell RNA sequencing and single-cell V(D)J sequencing were performed on pretreatment PBMCs of the first 50 patients enrolled in the prospective cohort. Single-cell suspensions from PBMC samples were obtained as described above and prepared to a concentration of approximately 1,000 viable cells per 1uL using a hemocytometer (Thermo Fisher Scientific) for cell counting according to the manufacturer's instructions. The single-cell suspensions were then subjected to library preparation for paired scV(D)J-seq and scRNA-seq using a 5' transcriptome kit (10xGenomics) according to the manufacturer's instructions. Complementary DNA libraries, targeted at 5,000 cells per sample, were sequenced using a NovaSeq instrument (Illumina) with 2 x 100 base pair paired-end reads targeting 20,000 read pairs per cell.

[0183] Raw scRNA-seq reads were barcode-deduplicated and aligned to the hg38 reference genome using Cell Ranger v3.1.0 to obtain a sparse digital count matrix, which was then analyzed to identify cell type and state using Seurat v4+. Outlier cells were identified and removed based on the following criteria: (1) >25% mitochondrial content or (2) cells with <100 or >1,500–3,000 expressed genes, depending on the sample-level distribution. After normalization (NormalizeData) and variable feature identification (FindVariableFeatures), we applied FindIntegrationAnchors to identify anchors and IntegrateData to perform batch correction. The most variable genes and top principal components were used to apply principal component analysis (PCA) and uniform manifold approximation and projection (UMAP). We applied FindClusters to identify cell types and states, assigning these to major cell lineages based on the expression of canonical marker genes, and separately used a reference-guided annotation framework within Seurat v4 (Azimuth78) to project the scRNA-seq dataset onto a PBMC atlas of 161,764 cells across six major lineages and 27 more detailed subsets.

[0184] Raw scV(D)J-seq reads were mapped to the reference refdata-cellranger-vdj-GRCh38-altensembl-4.0.0 using Cell Ranger v7.0, and the resulting clonotype assemblies were downloaded from Loupe V(D)J browser v3.0.0 (10xGenomics). To calculate T cell receptor (TCR) diversity, we used Shannon entropy (R package vegan v.2.5-379) for total PBMCs for each T cell subset. B cell receptor (BCR) clonotypes were similarly analyzed across B cell subsets for the IGK, IGL, and IGH chains.

[0185] All identified immune cell states were correlated with the incidence of symptomatic irAEs (grade 2+), severe irAEs (grade 3+), and life-threatening irAEs (grade 4+) across CyTOF, scRNAseq, and scV(D)J-seq data (including subset-specific TCR / BCR Shannon entropy) using receiver operating characteristic area under the curve (ROC AUC) analysis, Kaplan-Meier analysis, Cox proportional hazards regression, Wilcoxon rank-sum test when comparing two groups, and Kruskal-Wallis test when comparing more than two groups simultaneously. Identified T / immune cell states were correlated by grade with irAE occurrence using the Jonkheel-Tapstra test for ordinal data. Additionally, in an exploratory fashion, the abundance of circulating immune phenotypic states was correlated with organ-specific toxicities (i.e., colitis, myocarditis, pneumonitis, hepatitis, thyroiditis, hypophysitis, etc.). Multivariate models were used to verify independence from other clinical indicators. Carry out multiple hypothesis correction using the Benjamini-Hochberg method if necessary.

[0186] For patients receiving concomitant ICIs and profiled by CyTOF (n = 30), pooling the retrospective cohort and the new prospective analysis (Section C.5) revealed that the relationship between pretreatment levels of CD4 T cells and the occurrence of severe irAEs had an effect size of 1.07. Based on this, at least 22 patients per group were required to achieve α = 0.01 and 1 - β = 0.8. For Tph cells, which were only evaluable in the new prospective cohort, the effect size to identify concomitant ICI-related severe irAEs was 2.05, requiring 7 patients per group (alpha = 0.01, power 80%). A cohort of 100 patients would meet these requirements, as approximately 60 of the 100 patients would be expected to have severe irAEs. The ability to identify cell populations from scRNA-seq data depends on three key variables: the number of cells profiled, the fold change between population-specific genes, and the number of population-specific genes. To identify individual differentially expressed genes, we predict that at least 1,000 cells will be required with an effect size >0.89 to detect rare populations down to 1% (10 cells) (alpha = 0.05, power 80%). Because we plan to sequence approximately 250,000 cells in total (approximately 5,000 per sample for 50 samples), we expect sufficient power to detect modest effect sizes.

[0187] The data (i) validate previously published results6 that CD4 effector memory T cells are enriched in the pre-treatment peripheral blood of melanoma patients who develop severe irAEs; (ii) CXCR5-PD1hi CD4 Tph cells are the CD4 memory T cell state that best predicts the occurrence of severe irAEs, and their baseline levels correlate gradedly with the severity of toxicity; (iii) Tph cells enriched pre-treatment in patients experiencing toxicity also have elevated TCR diversity; and (iv) demonstrate whether additional immunophenotypic cell states are associated with organ-specific toxicity.

[0188] The heterogeneity of potential mechanisms underlying ICI-induced irAEs complicates the development of therapeutic strategies to mitigate or avoid them. However, any T cell state that preferentially expands during treatment before irAE onset is likely directly or indirectly involved in irAE pathogenesis. In a pilot analysis of 15 melanoma patients treated with concomitant ICIs, we identified a clonal expansion of bulk T cells during treatment associated with severe irAEs. In three patients for whom we performed combined scRNA-seq and scV(D)J-seq, we identified a shift toward clonally expanded, activated CD4 T cells, rather than CD8 T cells, in the blood during treatment preceding the onset of severe irAEs. Based on our preliminary data, we hypothesize that activated CD4 effector memory T cell clonotypes, specifically Tph cells, preferentially expand during treatment before the onset of severe irAEs. Our alternative hypothesis is that alternative T / B cell subsets (or non-T / B cell subsets) are preferentially expanded during treatment prior to the development of severe irAEs.

[0189] For the first 50 patients enrolled in the prospective cohort, we will perform single-cell RNA sequencing and single-cell V(D)J sequencing on early-treatment PBMCs (in addition to the pre-treatment PBMCs profiled in Aim 1) using a protocol identical to that described in Section D.4.d.4.

[0190] Paired pre- and post-treatment scRNA-seq and scV(D)J-seq data were analyzed as described in sections D.4.d.5 and D.4.d.6 and our previous study. To maximize rigor and avoid classification artifacts, when annotating CD4 and CD8 T cell-derived clonotypes, we first considered only TCR clonotypes that uniformly expressed positive lineage markers (CD4>0 and CD8A / B=0 for CD4 T cells; CD8A or CD8B>0 and CD4=0 for CD8 T cells). In our previous study, 69% of all clonotypes could be unambiguously labeled using this approach.

[0191] For each T cell state identified in both pre- and on-treatment scRNA-seq data pairs, we assess the richness and evenness of the TCR-β repertoire using Pielou's evenness, which is robust to the number of clones per sample and correlates with increased clonality. Clonal expansion of each T cell state is then inferred by analyzing clonality differences between on-treatment and on-treatment time points. We also analyze individual clonotypes unique to pre- and on-treatment samples, as well as persistent clones, defined as productive TCR-β CDR3 nucleotide sequences shared between pairs of pre- and on-treatment blood samples. For the latter, we analyze differences in productive frequency between paired samples for each clonotype, allowing for assessment of clonotype dynamics both individually and collectively to determine whether persistent clonotypes in any T cell state preferentially expand after ICI initiation but before the onset of severe irAEs. These analyses are repeated for B cell states.

[0192] After confirming (or identifying) the clonally expanded immunophenotypes most significantly associated with severe irAEs in 50 patients (D.5.c.4), we sorted these immunophenotypes from PBMCs collected pre- and early-treatment from the remaining 50 patients in the cohort. Specifically, approximately 5 million PBMCs were treated with TruStain FcX Fc receptor blocking solution (BioLegend) for 10 minutes at room temperature, then stained with fluorophore-tagged surface antibodies specific for the desired T / B cell state for 30 minutes at room temperature. After exclusion of DAPI-positive cells and putative doublets based on forward and side scatter analysis, the PBMCs were sorted with operator assistance using a Sony SY3200 Synergy instrument at the Siteman Flow Cytometry Core at the University of Washington. Sorting performance was confirmed by analysis of flow cytometry output using FlowJo v10.

[0193] Here, we performed immunoSEQ TCR-b chain (or BCR) profiling on pairs of PBMCs sorted pre- and early-treatment. First, genomic DNA was extracted from each sorted population using the DNeasy Blood & Tissue Kit (Qiagen) and submitted to survey-resolution immunoSEQ (Adaptive Biotechnologies). Data from productive rearrangements were exported using the immunoSEQ Analyzer online tool and assessed for repertoire richness and diversity using Pielou's evenness. Pre-treatment normalized clonality for sorted cell states was then compared across irAE grades to confirm the predicted phenotype.

[0194] Additionally, time to severe irAEs was assessed based on the degree of immunophenotypic clonal expansion, which may further suggest functional relevance. Kaplan-Meier analysis was used after division into tertiles, and multivariate Cox regression was also performed, taking into account the time between blood draw and other covariates.

[0195] TCR clonotypes expanding in peripheral blood may be more likely to recognize self- and disease-associated antigens in patients destined to develop severe irAEs. Figure 31 compares the similarity of expanding TCR clonotypes with an external database (McPAS-TCR) of CDR3 sequences for known antigens in autoimmunity, cancer, and pathogenic infections. The similarity index per clonotype and per patient was determined as the sum of the % TCR matches at edit distances of 1, 2, and 3 and compared between patients based on severe irAE status. Figure 31 shows TCR clonotypes from Lozano et al. compared with the McPAS-TCR database of CDR3 sequences with known antigen specificities, stratified by antigen type and irAE status. Wilcoxon tests were used for group comparisons.

[0196] In patients treated with concomitant ICIs, the relationship between TCR clonal expansion (1-Pielou equivalence) and the occurrence of severe irAEs has an effect size of 1. Based on this, 15 patients per group (with and without severe irAEs) are needed to achieve α=0.01 and 1-β=0.8. If we can successfully isolate the T cell state that drives this signal, we expect an even higher effect size for this association. Therefore, training / validation cohorts of 50 patients each (including approximately 30 with severe irAEs and approximately 20 without severe irAEs) are expected to be adequately powered for the proposed study.

[0197] We hypothesize that ICI toxicity and response signatures can be detected in pretreatment cfDNA methylation profiles and can reliably predict corresponding clinical outcomes in melanoma. Consistent with this, data from 21 patients demonstrate that promoter methylation profiling of pretreatment cfDNA can predict ICI toxicity and benefit using published signatures (Figures 25-27). First, we apply NEBNext enzymatic methyl-seq to pretreatment cell-free DNA from the same 100 patients described above. After dividing into training / validation cohorts, we evaluate our published toxicity and response gene signatures, refined toxicity signatures, and an independent machine learning strategy to identify predictive signatures.

[0198] Briefly, approximately 20 mL of blood was collected into K2EDTA Vacutainer tubes (Becton Dickinson) at 1,200 g for 10 minutes to separate the plasma from the PBMCs, and the plasma was then spun again at 1,800 g to remove any remaining PBMCs, after which the double-spun plasma was stored at -80°C in approximately 2 mL aliquots.

[0199] cfDNA was extracted from plasma using the QiaAmp Circulating Nucleic Acid Kit (Qiagen) according to the manufacturer's instructions. cfDNA concentration was measured using a Qubit 4.0 fluorometer with a dsDNA High Sensitivity Assay Kit (Thermo Fisher Scientific), and fragment size was assessed using a High Sensitivity DNA Kit (Agilent Technologies) on an Agilent 2100 Bioanalyzer. A median of 50ng was used for EM-Seq library preparation according to the manufacturer's instructions. Libraries were sequenced using a NovaSeq 6000 (Illumina) targeting 30x genome-wide coverage.

[0200] After alignment and determination of methylation sites using Bismark with default parameters, promoter methylation levels were analyzed (1 kb upstream of the gene body) in gene signatures associated with ICI toxicity. The top 10, 20, 50, and 100 signature genes were examined (log2 fold change) from activated CD4+ memory T cells profiled by scRNA-seq to identify the cellular states most associated with severe irAE development. These analyses were performed across the entire gene and promoter region, as well as focusing on the 1 kb promoter region.

[0201] For each gene set, the signature score—defined as 1 minus the average promoter methylation level—was corrected for each sample's background by randomly sampling the same number of genes across the transcriptome and calculating 1 minus the average promoter methylation level. This was performed 10 times, then averaged and subtracted from the original signature score. Corrected signature scores were derived from permutations of different gene sets / regions, and these scores were compared with the incidence of severe irAEs in the first 50 patients enrolled in the cohort (discovery). A discriminatory cutpoint was trained by applying Youden's J statistic to receiver operating characteristic (ROC) analysis. Using CpGs from the top 100 signature genes (selected as above), we also trained a series of machine learning models (SVM, logistic regression, random forest) for comparison. All predictors were evaluated by LOOCV to identify the best approach.

[0202] As validation, we compared the best-performing predictors with cfDNA EM-seq data from the remaining 50 similarly treated patients and with data from our pilot study. Their performance in predicting severe irAEs was assessed by receiver operating characteristic (ROC) analysis, two-sided Wilcoxon test, Kaplan-Meier analysis of freedom from severe toxicity analyzed by the log-rank test, and multivariate Cox proportional hazards model for freedom from severe toxicity. We also assessed the graded relationship of these EM-seq-derived gene signatures with irAE grade by plotting their values ​​against irAE grade, and statistical testing was performed using the nonparametric Jonkheel-Tapstra test for ordinal data.

[0203] Next, we investigated cell state-specific genes comprising the cellular ecosystems we previously identified in tumor tissues, including CE9, a pro-inflammatory ecosystem that strongly predicts immunotherapy response. Using the data analysis pipeline described above, promoter methylation levels in plasma cell-free DNA were analyzed for durable clinical benefit (DCB) to immunotherapy (defined as no progression by RECIST 1.1 criteria on standard of care imaging at least 6 months after ICI initiation). Gene sets were defined as the top 10, 20, 50, and 100 signature genes per cell state, encompassing each ecosystem.

[0204] Using the same sequencing alignment and methylation analysis protocols as described above, we used an independent machine learning system to identify patients at risk for severe irAEs and those likely to achieve durable clinical benefit from concomitant ICIs. The algorithm also addresses the predictive ability to distinguish irAE grades. Considering computational efficiency, as an initial feature selection step, we used a two-tailed Wilcoxon test to identify all CpGs with significant differences between 1) patients who developed severe irAEs and those who did not, and 2) patients who achieved durable clinical benefit (DCB) and those who did not (NDB). Next, we used CpGs with nominal significance (P<0.05) to train an extreme gradient-boosted decision tree model (XGBoost) to optimize decision tree parameters by LOOCV, first in our initial 50-patient discovery cohort and then in the provided validation cohorts 1 and 2 (D.6.c.3), to identify defined outcomes. Performance will be compared to the results in the above section using Kaplan-Meier analysis and multivariate Cox proportional hazards regression, using ROC AUC analysis for dichotomous variables and PFS / OS.

[0205] In the cfDNA data (Figures 25-27), the association between the activated CD4 TEM signature and severe irAE incidence in patients treated with concomitant ICIs had an effect size of 1.43. The relationship between the biologic signature and DCB had an effect size of 0.9. Based on this, 20 patients per clinical group are required to achieve α = 0.05 and 1 - β = 0.8. Because a severe irAE rate of approximately 60% and a DCB rate of approximately 60% are expected, training and validation cohorts of 50 patients each provide sufficient power.

[0206] The data show that (i) the signature of activated CD4 memory T cells, more specifically CD4 Tph cells, can predict severe irAE occurrence in the discovery and validation cohorts in the genome-wide methylation profiles of pretreatment cfDNA, and (ii) the signature obtained using a machine learning algorithm can be applied to pretreatment cfDNA methylation data to predict durable response and survival in the discovery and validation cohorts. We also (iii) determine whether machine learning can surpass (i) and (ii) in predicting ICI toxicity and benefit from the same data.

[0207] The method of the present invention was applied to the combined genomic and epigenomic analysis of cfDNA in high-risk castration-resistant prostate cancer to identify prognostic liquid biopsy signatures. LiquidTME technology can be applied to risk stratify patients with locally advanced or metastatic cancer, distinguishing between patients with molecularly low-risk disease and those with molecularly high-risk disease. In this way, seamless risk stratification across cancer types can help clinicians adjust / intensify / individualize treatment regimens.

Claims

1. 1. A method of predicting a disease outcome in a subject, the method comprising: identifying epigenetic modifications in the sequence data, said epigenetic modifications comprising methylation status, fragment size, and / or terminal motifs in cell-free nucleic acids obtained from a liquid biopsy sample; identifying an abnormal tissue microenvironment based on the epigenetic modifications and assigning a tissue of origin to the cell-free nucleic acids; Using the pattern of epigenetic modifications to determine the state of cells in the abnormal tissue microenvironment; and predicting the outcome of a disease in the subject based on the cellular state in the abnormal tissue microenvironment. A method comprising:

2. 2. The method of claim 1, wherein the abnormal tissue microenvironment is selected from a tumor microenvironment (TME), an inflammatory microenvironment (IME), a tissue transplant microenvironment (TTME), and a pathogen microenvironment (PME).

3. The method of claim 2, wherein the circulating cell-free nucleic acid is derived from microenvironment-infiltrating lymphocytes.

4. 4. The method of claim 3, wherein the disease is cancer and the abnormal tumor microenvironment is TME.

5. 5. The method of claim 4, wherein predicting the outcome of cancer in the subject comprises one or more of diagnosing the subject with cancer, predicting remission, predicting recurrence, assessing minimal residual disease, predicting response to immunotherapy, and predicting toxicity of immunotherapy.

6. 6. The method of claim 5, wherein the subject has a measured low tumor mutational burden.

7. 5. The method of claim 4, wherein predicting the outcome of cancer in the subject further comprises determining that the cell state in the TME is above a pre-malignant threshold.

8. 5. The method of claim 4, wherein determining the cell state comprises measuring T cell abundance and / or diversity from the pattern of epigenetic modifications.

9. 9. The method of claim 8, wherein the T cell diversity measured is the T cell receptor (TCR) diversity of the patient's immune repertoire.

10. 9. The method of claim 8, further comprising the step of inferring the abundance of T cell effector memory cells from said pattern.

11. 11. The method of claim 10, further comprising predicting the risk of an adverse event in response to immunotherapy or non-response to immunotherapy, wherein the measured TCR diversity and / or the predicted abundance of T cell effector memory cells is below a predetermined threshold.

12. 11. The method of claim 10, further comprising predicting a favorable outcome in response to immunotherapy, wherein the measured TCR diversity and / or the predicted abundance of T cell effector memory cells exceeds a predetermined threshold.

13. 2. The method of claim 1, wherein the step of determining the cell state comprises: sequencing the nucleic acid to generate sequence data comprising methylated bases; mapping the sequence data to a reference to identify promoters of a plurality of genes; and identifying the cell state based on the set of genes having methylated promoters.

14. The method of claim 1 , wherein the identified epigenetic modification comprises promoter methylation.

15. 10. The method of claim 1, further comprising querying a cell state atlas for the pattern in the epigenetic modifications.

16. 16. The method of claim 15, further comprising providing a profile describing a cellular state of a plurality of cells in an abnormal tissue microenvironment in the patient.

17. 5. The method of claim 4, further comprising providing a profile of tumor-infiltrating leukocytes in a tumor microenvironment based on the pattern in the epigenetic modifications.

18. 5. The method of claim 4, wherein the cell-free nucleic acids comprise nucleic acids derived from non-tumor cells, and determining the cellular state in the abnormal tissue microenvironment comprises creating a profile of the TME.

19. 19. The method of claim 18, wherein the non-tumor cells comprise one or more of somatic cells, immune cells, and / or cells from the tumor margin.

20. 20. The method of claim 19, wherein the TME profile comprises one or more epigenetic patterns that correlate with tumor progression and / or somatic regression.

21. 21. The method of claim 20, further comprising using the TME profile to predict immunotherapy toxicity in the subject.

22. The step of predicting the toxicity of the immunotherapy is based on the level of epigenetic modifications of promoters in genes involved in the transcriptional state of a given T cell; and 22. The method of claim 21, wherein predicting immunotherapy response is based on the level of promoter methylation in said gene in a given tumor ecotype.

23. 23. The method of claim 22, wherein the predetermined T cell transcriptional state is specific for CD4 T cells.

24. 23. The method of claim 22, wherein the predetermined tumor ecotype is carcinoma ecotype 9 or carcinoma ecotype 1.

25. The step of predicting the toxicity of the immunotherapy is based on the level of epigenetic modification of promoters in genes involved in the expression of one or more cytokines; and 22. The method of claim 21, wherein predicting immunotherapy response is based on the level of promoter methylation in said gene in a given tumor ecotype.

26. the epigenetic modifications predict the toxicity of immunotherapy; interleukin-10 promoter hypomethylation; interleukin-6 promoter hypermethylation; and Interleukin-7 promoter hypermethylation 26. The method of claim 25, wherein the ion exchange is selected from one or more of:

Images