Method for using the t cell repertoire as a predictive and prognostic biomarker in cancer

Abstract

The present invention relates to a method of establishing a prediction regarding the development of a type of solid cancer in an individual, comprising: (i) collecting a biological sample that includes lymphocytes of said individual, (ii) sequencing, using NGS techniques, the mRNA associated with the V(D)J regions of the TCRs of the lymphocytes collected in the previous step, (iii) removing from the TOR repertoire obtained in the previous step the TOR sequences that are specific to other diseases than said solid cancer, (iv) using artificial intelligence algorithms to compute a predictive score for the probability that said TOR repertoire is specific for said type of cancer and (v) based on said predictive score determining the probability of said individual of developing said type of solid cancer.

Description

Method for using the T cell repertoire as a predictive and prognostic biomarker in cancer

[0001] The present invention relates to a method for using the T cell repertoire for the prediction of the occurrence of a specific type of cancer in humans.Background

[0007] The T lymphocytes of an individual respond to a disease such as cancer by generating and expressing on their surface specific receptors (T cell receptors, known as TCRs) that recognize and bind to disease-specific antigens, such as tumour-specific antigens (TSAs) and tumour-associated antigens (TAAs). TSAs and TAAs are typically portions of intracellular molecules that are expressed on the tumour cell surface as part of the major histocompatibility complex (MHC). The TCRs are themselves protein complexes usually comprising two variable alpha (a) and beta ( ) chains, each comprising three hypervariable domains also known as complementarily determining regions (CDRs). Of these hypervariable regions, the region CDR3 of chain [3 (CDR3[3) is considered the primary CDR that recognizes the antigen.

[0008] The sum of all TCRs sequenced from a sample containing T cells of one individual is termed the TCR repertoire or TCR profile or total TCRs. The highly diverse multitude of TCRs needed for recognizing and binding to any potential antigen is generated during the maturation process of the T lymphocites by the process known as somatic recombination or V(D)J recombination. This process leads to the production of a multitude of different possible V(D)J regions, which are responsible for the hypervariability of the CDR3 domains of the TCRs. V(D)J recombination of the different TCR genes could theoretically generate between 1015and 1 O20TCR chains. However, the actual TCR diversity in a human's body is estimated at around 1013different TCR variants (clonotypes), which implies that the TCR development is not entirely random and is subject to certain limitations. Moreover, while there are TCRs that are common in the general population, recent high-resolution studies have demonstrated that the majority of TCRs are rare, meaning that extremely precise methods would be needed in order to elucidate the TCR repertoire of an individual.

[0009] In recent years, highly specific methods based on next generation sequencing (NGS) have been developed that allow for the simultaneous analysis of multiple TCR sequences.Their results are organized in interactive panels of data that allow for the representation of a TCR repertoire based on the length, diversity and position of the CDR3, respectively the V(D)J regions, in the TCRs. Other data may also be available such as the number / frequency of clones identified for each variable combination gene-length of CDR3 nucleotide, the Shannon diversity, etc., allowing for identifying and searching for individual (rare) clones of interest. Nevertheless, T-cell receptor (TCR) repertoires identified through next-generation sequencing (NGS) methods have a significant magnitude, which is potentially larger than 107clones in a single individual. These repertoires also display substantial inter-individual diversity, attributed to the varied distribution of MHC class I and II alleles across global populations. This variability can impede the recognition of cancerspecific T-cell clones. Thus, we are still far from being able to compile a comprehensive mapping of all the TCRs in the human population and their association with all relevant antigens, and also far from being able to reliably use such maps for diagnostic purposes.State of the art

[0002] Various studies have been performed in order to develop blood tests that can quickly and accurately detect cancer prior to the development of specific symptoms. For example, it has been documented that, for all cancers combined, the following blood test abnormalities increase the probability of cancer: low albumin, increased platelets, increased calcium, and increased inflammatory markers.

[0003] Although these test abnormalities give an indication that a cancer may be present, they leave uncertainty over which cancer(s) should be investigated, and have to be interpreted in conjunction with other tests or previous results of the same test.

[0004] The article "T-cell receptor repertoire analysis for the diagnosis and treatment of solid tumors: A methodology and clinical applications" by Na Li et al (2020) shows that NGS technologies can be employed to identify TCR repertoires of patients with cancer and suggests that the degree of diversity of a TCR repertoire can be associated with the effectiveness of immunotherapies in already diagnosed cancers.

[0005] A multi-cancer early detection (MCED) blood testing, presented at the ESMO Congress 2022, is said to detect a common cancer signal from over 50 different types ofcancer and predict where the signal has come from in the body. The signal is derived from small clusters of circulating tumour DNA in the blood that have different patterns of methylation from the rest of the DNA. However, the efficacy of this test is contingent on the presence of cancer cells in the blood at the moment of sampling and on the release of freely circulating tumour DNA, which may happen only at specific times during the development of the disease.

[0006] As a result, there is still a need for an effective rapid test that can accurately indicate the presence of specific types of cancers prior to the development of symptoms.Summary of the invention

[0010] The present invention refers to a method of establishing a prediction regarding the development of a type of solid cancer in an individual, said method comprising the steps:(i) collecting a biological sample that includes lymphocytes of said individual, preferably a sample of peripheral blood,(ii) sequencing, using NGS techniques, the mRNA associated with the V(D)J regions of the TCRs of the lymphocytes collected in the previous step, to obtain a list of TCR sequences that make up a TCR repertoire of said individual,(iii) removing from the TCR repertoire obtained in the previous step the TCR sequences that are specific to diseases or conditions other than said solid cancer, to obtain a working TCR repertoire,(iv) using artificial intelligence algorithms based on neural networks with attention layers trained on datasets of: TCR repertoires that are cancer-specific for said type of cancer and TCR repertoires that are cancer-free, to compute a predictive score for the probability that said working TCR repertoire is a cancer-specific TCR repertoire for said type of cancer or a cancer-free TCR repertoire for said type of cancer;(v) based on the predictive score obtained in the previous step, determining the probability of said individual of developing said type of solid cancer.

[0011] By using the prediction method of the present invention, it is possible to accurately, easily and quickly predict whether an individual who shows no cancer symptoms or inconclusive symptoms, and / or whose other cancer tests are negative or inconclusive, is atrisk of developing a certain type of cancer in the near future. Thus, it will become possible to closely monitor the individuals who are at risk and to establish a very early and highly individualized diagnostic and treatment.

[0012] The present method of establishing a prediction is carried out on said biological sample ex vivo or in vitro. The method relies on the capacity of T lymphocytes, especially memory T lymphocytes, to remain in circulation for a long period of time; this increases the probability of detecting at a specific moment in time the presence in a biological sample (such as in the peripheral blood of the patient) of tumour-specific TCRs, as compared to the chances of detecting tumour cells or tumour DNA.

[0013] The present method offers an intervention window wherein to specifically investigate the individual for the type of cancer for which the high risk has been detected, in order to start the treatment or prevention therapy at the earliest possible moment. The inventors of the present disclosure have been able to determine the effectiveness of the method for colorectal and lung cancers, but it will be obvious to the specialist in the art that the method will yield similar results for any other cancer type.

[0014] In step (i) of the prediction method of the present invention, by biological sample containing lymphocytes is meant any bodily fluid containing lymphocytes such as blood, lymph, ganglions, etc. However, for ease and rapidity of sampling, peripheral blood is preferred, such as peripheral venous blood.

[0015] In step (ii) of the method, the mRNA associated with the V(D)J regions of the TCRs of the sampled lymphocytes is sequenced.

[0016] The present inventors have found that, for the purposes of the present method, sequencing of the mRNA associated with the V(D)J regions of the TCR has proven to have substantial advantages over sequencing the tumour-associated DNA or antigens. T lymphocytes start circulating in the bodily fluids of an individual from the moment the first tumour cells appear, i.e. from the earliest stages of the cancer, long before the tumour nucleic acids / antigens can be reliably detected themselves. Moreover, T lymphocytes, especially memory T lymphocytes, remain in circulation for a long period of time, which increases the probability of being detected at a specific moment in time in a biologicalsample (such as in the peripheral blood of the patient), as compared to the chances of detecting tumour cells or tumour DNA.

[0017] The present inventors have also found that, for the purposes of the present method, sequencing of the mRNA associated with the V(D)J regions of the TCRs has proven to have substantial advantages over sequencing the DNA of the TCRs, having greater sensibility and allowing for a more comprehensive identification of the individual (rare) TCR variants. This is because, on one hand, the DNA is less common and does not have the same breadth of coverage. On the other hand, the sequencing of the mRNA can be used to identify the expressed TCRs, meaning the TCR sequences that were already subject to the splicing process and had undergone post-criptional processing, and are thus susceptible to producing functional TCRs. Per comparison, DNA sequencing does not identify the TCR sequences in their transcripted forms, and thus inevitably numerous unproductive sequences, which are functionally irrelevant, can be produced, which reduces the accuracy of the testing. Additionally, because the relevant nucleotide sequences are not situated next to one another in the DNA, the DNA templates are significantly greater than the corresponding mRNA templates; as a result, complex amplifications are necessary that are susceptible to mistakes. Conversely, mRNAs are shorter, simpler in design and may be employed to amplify the hypervariable regions of the TCR with a single pair of primers, which leads to reducing the number of amplified defects and allows for the analysis of both the a and [3 chains of the TCR in the same procedure.

[0018] The mRNA sequencing of step (ii) is accomplished using NGS techniques, by which the nucleotide sequences of the relevant hypervariable regions of the TCRs are obtained, as well as the associated amino acid sequences of the corresponding TCR peptides. NGS methods also facilitate the rapid identification of a large number of V(D)J regions. The processes of NGS sequencing are known to the specialist, who has the knowledge to perform the sequencing step as described above.

[0019] By sequencing the mRNA coding for the V(D)J regions of the TCRs, at the end of step (ii) a list of TCR sequences is obtained that constitutes a TCR repertoire of the individual. Said TCR repertoire can be visualized in the form of a list of aligned TCR sequences (nucleotides and / or amino acids) in letter code, for example, in the form of a .tsv file.

[0020] The TCR repertoire is then compared in step (iii) of the method with established TCR databases such as ImmunoMap, CDRdist, GLIPH. These and other known TCR databases are equipped with search engines (such as, for example, the TCRmatch library tool), able to compare and match the TCR repertoire with various TCR sequences that are known to be specific to certain diseases and conditions. Thus, the TCR sequences that are specific to diseases or conditions other than the target of the method (other than the specific cancer type) can be identified and removed from the TCR repertoire file. As a result, a TCR repertoire of smaller dimensions, also called a working TCR repertoire, is obtained at the end of step (iii), which is more easily accessible for analysis. This comparison and removal can be accomplished, for example, through the use of the search tools associated with each established database.

[0021] In a preferred embodiment, said working TCR repertoire is further reduced by selecting the top most prevalent TCRs that have the greatest degree of clonality, such as preferably selecting the top 10,000 such TCRs, and removing the rest of the TCRs from the working TCR repertoire. Using the resulting reduced working TCR repertoire in the further steps of the present method has proven to lead to an advantageously higher accuracy of the reading.

[0022] The working TCR repertoire or the reduced working TCR repertoire thus obtained is then analyzed in step (iv) of the present method by using a machine learning (ML) tool. This is an artificial intelligence tool that employs specialized neuronal networks with attention layers that rely on predictions and statistics, which enables the collection, interpretation, and analysis of large amounts of data. However, it has been found that such a machine learning tool cannot be effectively trained using only the known cancer-related TCR sequences from the existing public databases. At present, these databases contain a limited number of samples from cancer patients, which samples are not sufficient to train a reliable tool that can predict the occurrence and development of the respective cancer. Additionally, the TCR repertoires are diverse across individuals; this is attributed to the diverse distribution of MHC class I and II alleles across global populations. This variability complicates the recognition of T-cell clones associated with cancer for a specific individual.

[0023] However, it has been surprisingly found by the inventors that, despite said substantial inter-individual diversity and insufficient number of TCR sequences associated with specificcancers, a machine learning tool can be developed that is capable of producing a prediction score regarding the probability of developing a specific type of cancer. Such a ML tool can be developed by using training and validation data comprising a sufficient number TCR repertoires sampled from individuals who are known to have developed the target disease (the specific type of cancer to be analyzed) and from individuals who are known to be free of the target disease (healthy individuals).

[0024] Thus, it represents another aspect of the present invention a method of training artificial intelligence to be used in step iv) of the above disclosed method, in order to predict the probability that a person will develop a specific type of cancer, said method of training comprising the following steps:A) receiving, at least on one server, a training dataset that includes a sufficient number of TCR repertoire files that are encoded and labelled as follows:- the TCR repertoire files and the TCR sequences contained therein, which were sampled from patients that are confirmed to have a specific type of cancer, are labelled as cancer-specific,- the TCR repertoire files and the TCR sequences contained therein, which were sampled from patients that are known to be healthy, i.e. free of said specific type of cancer, are labelled as cancer-free;B) feeding the encoded labeled data to a neural network,C) executing by said neural network the following steps:1 ) randomly initializing a set of weights and biases as function parameters for obtaining a prediction for the probability of a TCR repertoire to be cancerspecific or cancer-free,2) on each TCR sequence in each TCR repertoire:(a) using the function of step 1 ) to obtain a prediction score,(b) using a loss function to compare the prediction score obtained in step (a) with the label of said TCR sequence and calculate a measure of dissimilarity between them,(c) using said measure of dissimilarity to optimize said weights and biases of the neural network using backpropagation;3) calculating the accuracy of said predictions based on the results of the comparisons, and4) repeating steps 2) to 3) until a pre-designated condition is met.

[0025] In step 4) of the above method, said pre-designated condition may be:- the neural network reaching a configured threshold of success, or- a pre-determined number of epochs, meaning a predetermined number of repetitions of steps 2) to 3), or- until the accuracy does no longer increase for a predetermined number of consecutive epochs.Preferably, the pre-determined condition in step 4) is reaching a configured threshold of success.

[0026] Thus, this Al tool can use the above method to generate a prediction score expressing the probability of a particular TCR repertoire file to contain TCRs identified as cancer-specific. This prediction score being derived from the sequences identified as cancer-specific in the TCR repertoire of the individual, this information indicates the degree to which the T-cells of said individual have encountered cancer cells of said type and have developed an immunity response thereto. As a result, this score is indicative of the presence and number of cancer cells of said specific type of cancer in that individual, and describes the probability of said individual of developing said cancer type in the near future.

[0027] In a preferred embodiment, for the purposes of setting and training the machine learning tool to be used in step (iv) of the prediction method of the present invention, the inventors have incorporated in a new database the working TCR repertoires obtained according to the steps i) to iii) of the prediction method as detailed above, taken from: 500 patients suffering from colorectal cancers, 500 patients suffering from lung cancers and 500 healthy (cancer-free) patients. Using these cancer-specific-labelled and cancer-free- labelled TCRs as training datasets, a machine learning (ML) tool was developed and trained by the Romanian company Altfactor. This ML tool used 70% of the labelled working TCR repertoires in each category as learning data and the other 30% labelled working TCR repertoires in each category as validation data. This machine learning tool has been trained as per the method disclosed above to calculate a predictive score of probability that a TCR repertoire to be analyzed belongs in a group of cancer-specific repertoires (such as, in this embodiment, colorectal or lung cancer labelled TRC repertoires) or in the group cancer-free TCR repertoire. During the validation phase, this machine-learning tool had a predictedaccuracy of 70-72%; after only one training iteration, the accuracy increased to 71 -73%, and is expected to further increase as new data and feedback are incorporated.

[0028] Based on the prediction score obtained in step (iv) of the method of prediction of the present invention by the ML tool trained as described above, it can be determined in step (v) whether or not the tested individual is at risk of developing the specific type cancer tested. The inventors have documented during the follow-up studies that a score of 70% or higher of similarity with the cancer-specific TCRs is associated with a high probability of cancer development, a score of 50-70% similarity is associated with a moderate probability of cancer development, a score of 30-50% similarity is associated with a low probability of cancer development, and a score of under 30% similarity is associated with lack of evidence of cancer development.

[0029] It represents another aspect of the current invention, the use of the described method of predicting cancer risk for assisting physicians in determining the degree of risk associated with a specific type of cancer for a specific individual, and in determining if and how often additional tests and / or exams, such as for example an Rx, or colonoscopy, should be conducted.

[0030] Additionally, the physician will be able to use the results of the present method of prediction to determine if the individual should benefit from any prophylactic therapies such as immunostimulation, vaccination, antibiotic administration, and / or dietary changes, or behavioral changes. To accomplish this, the predicted values should be interpreted in the clinical context; they should be associated with the personal / family medical history of the individual, as well as be correlated with additional parameters such as: age and gender, the history of familial cancer, the personal history of infection (such as with HIV, CMV, or hepatitis B), the subject's lifestyle (such as smoking, drugs or alcohol, etc), medication (such as hormones or immunotherapies), imaging techniques results (such as ultrasound, x-rays, or PET-CT), blood test results (such as CBC), the number of lymphocytes, or the genetic makeup of said individual.Example

[0031] For implementing the method of prediction of the present invention, in a first stage, biological samples were collected in the form of blood lymphocytes from the peripheral veins of a subject.

[0032] The mRNA associated with the V(D)J regions of the TCRs in the lymphocytes was sequenced using NGS techniques, which resulted in a repertoire of TCRs that was specific to the individual in question. Specifically, a study of the human T cell receptor beta chain amplified form from fresh mRNA using the Framework 3 and Joining Gene method was performed. Short amplicons (average length 80 bp) were used to cover the CDR3 region, which enabled the tracking of clones and the measurement of diversity in cells derived from degraded materials, including RNA that is treated with FFPE. The IMGT® database of variants, diversity and gene-splicing combinations was utilized as the basis for this, it was released with: The Ion ReporterTM Software.Classification of readings

[0033] After the first step of data processing, the raw sequencing reads were organized and graphed in a stacked bar format. The actual numbers of reads associated with each classification were documented in a table wherein each line represents a specific clone, listed from the most common to the least common. For each of the clones, the variables and the junction regions are listed along with the corresponding amino acid and nucleotide sequences.

[0034] The TCR repertoire file containing the sequences was filtered using the method as presented in the ISI-WOS paper ‘TCRMatch: Predicting T-Cell Receptor Specificity Based on Sequence Similarity to Previously Characterized Receptors’ (doi: 10.3389 / fimmu.2021 .640725). Thus, the working TCR repertoire was obtained by removing the TCR sequences associated with known diseases. During this filtering, around 25% of the amino acid sequences were removed, with an average number of 100k sequences remaining in the file.

[0035] The first 10% of these sequences, in the order of ratio value (how many of them were present in the blood sample), representing the reduced working TCR repertoire, were fed to the neural network developed and trained by Romanian firm Altfactor as per the method described above.

[0036] The results (the output data) were returned to the graphical / user interface in a JSON (JavaScript Object Notation) format file containing the fields: a) filename - which is the denomination of the analyzed TCR repertoire in .tsv file, b) class_name - which is the name of the category of TCR repertoires: neutral (for healthy), lung cancer, and colorectal cancer, c) prediction - which is the prediction score representing the probability of the analyzed TCR repertoire of belonging to the respective category (class) specified by class_name.

[0037] For instance, in the following exemplary output result, the neural network has calculated a prediction score of 47.27% probability of lung cancer:

[0038] The obtained results are then presented in a friendly user interface, wherein, for example, the calculated exemplary prediction score of 47.27% is shown to be in the interval of 30-50%, and associated with a low probability of cancer development. The subject and their physician may be given further information and suggestions associated with the calculated probability.

Claims

Claims1. A method of establishing a prediction regarding the development of a type of solid cancer in an individual, comprising the steps:(i) collecting a biological sample that includes lymphocytes of said individual, preferably a sample of peripheral blood,(ii) sequencing, using NGS techniques, the mRNA associated with the V(D)J regions of the TCRs of the lymphocytes collected in the previous step, to obtain a list of TCR sequences that make up a TCR repertoire of said individual,(iii) removing from the TCR repertoire obtained in the previous step the TCR sequences that are specific to diseases or conditions other than said solid cancer, to obtain a working TCR repertoire,(iv) using artificial intelligence algorithms based on neural networks with attention layers trained on datasets of TCR repertoires that are labelled cancer-specific for said type of cancer and TCR repertoires that are labelled cancer-free, to compute a predictive score for the probability that said working TCR repertoire is a TCR repertoire cancer-specific for said type of cancer or a TCR repertoire cancer-free for said type of cancer;(v) based on the predictive score obtained in the previous step, determining the probability of said individual of developing said type of solid cancer.

2. Method according to claim 1 , wherein the solid cancer is a colorectal cancer or a lung cancer.

3. Method according to any preceding claim, wherein, after step (iii), said working TCR repertoire is further reduced by selecting the top most prevalent TCRs that have the greatest degree of clonality, such as preferably by selecting the top 10,000 such TCRs and removing the rest of the TCRs from the working TCR repertoire.

4. Method according to any preceding claim, wherein, in step (v), a predictive score of 70% or higher is associated with a high probability of cancer development, a predictive score of 50-70% is associated with a moderate probability of cancer development, a predictive score of 30-50% similarity is associated with a low probability of cancer development, and a predictive score of under 30% is associated with lack of evidence of cancer development.

5. Use of a method according to any preceding claim for assisting a physician in determining whether an individual should benefit from further exams, such as a colonoscopy or X-ray, and / or from any prophylactic therapy such as immunostimulation, vaccination, antibiotic therapy, or from dietary supplements, and / or from changes in diet, behaviours, or in the life and work environment.

6. A method of training artificial intelligence to be used in step iv) of the method according to claims 1 to 4, to calculate a score of probability that a TCR repertoire is cancer-specific or cancer-free for a specific type of cancer, said method of training comprising the following steps:A)receiving, at least on one server, a training dataset that includes a sufficient number of different TCR repertoire files that are encoded and labelled as follows:- the TCR repertoire files and the TCR sequences contained therein, which were sampled from individuals that are confirmed to have a specific type of cancer, are labelled as cancer-specific,- the TCR repertoire files and the TCR sequences contained therein, which were sampled from individuals that are known to be healthy, i.e. free of said specific type of cancer, are labelled as cancer-free;B) feeding said encoded labeled data to a neural network,C) executing by said neural network the following steps:1 ) randomly initializing a set of weights and biases as function parameters for obtaining a prediction for the probability of a TCR repertoire to be cancerspecific or cancer-free,2) on each TCR sequence in each TCR repertoire:(a) using the function of step 1 ) to obtain a prediction score,(b) using a loss function to compare the prediction score obtained in step (a) with the label of said TCR sequence and calculate a measure of dissimilarity between them,(c) using said measure of dissimilarity to optimize said weights and biases of the neural network using backpropagation;3) calculating the accuracy of said predictions based on the results of the comparisons, and4) repeating steps 2) to 3) until a pre-designated condition is met.

7. A method of training artificial intelligence according to claim 6, wherein in step 4), said pre-designated condition may be:- the neural network reaching a configured threshold of success, or- a pre-determined number of epochs, meaning a predetermined number of repetitions of steps 2) to 3), or- until the accuracy does no longer increase for a predetermined number of consecutive epochs, preferably, the pre-determined condition being reaching a configured threshold of success.

Images