DESCRIPTION
“PROGNOSTIC METHOD FOR DETERMINING A PROBABILITY OF ALLOGRAFT REJECTION”
TECHNICAL FIELD
This invention is related to the field of organ transplantation and more particularly provides a method for the prognostic of rejection of transplanted organs. In particular, the present invention relates to a prognostic method for determining a probability of allograft rejection using an infrared spectroscopybased method coupled with a machine learning model.
BACKGROUND OF THE INVENTION
Transplantation is the removal of living, functioning cells, tissues, or organs from the body and then their transfer back into the same body or into a different body. When a tissue or organ is transplanted from one person to another, it is called an allograft transplant. Allograft transplants are typically used when tissue from one’s own body, called an autograft transplant, cannot be used.
Transplant rejection can be classified as hyperacute, acute, or chronic. Hyperacute rejection is usually caused by specific antibodies against the graft and occurs within minutes or hours after grafting. Acute rejection (AR) occurs days or weeks after transplantation and can be caused by specific lymphocytes in the recipient that recognize human leukocyte antigen (HLA) in the tissue or organ grafted. AR may be mediated by T-cell (TCMR) or by antibodies (ABMR). Finally, chronic rejection (CR) usually occurs months or years after organ or tissue transplantation. Various mechanisms involving chronic inflammation, humoral, and cellular immune reactions play an essential role in the immunopathogenesis of chronic rejection [https://pubmed.ncbi.nlm.nih.gov/30571056/].
Transplantation is the whenever possible is the only mode of therapy for most end-stage organ failure affecting kidneys, liver, heart, lungs, and pancreas. However, the graft organ can be lost by diverse causes from pre-transplant organ stress, mismatch between recipient metabolic demands and graft renal mass, post-transplant infections, the graft immune rejection or even due to side effects of the immunosuppressive drugs therapy [M. O. Hamed et al., “Early graft loss after kidney transplantation: Risk factors and consequences,” Am. J. Transplant. , vol. 15, no. 6, pp. 1632-1643, 2015], [Z. M. El- Zoghby et al., “Identifying specific causes of kidney allograft loss,” Am. J. Transplant. , vol. 9, no. 3, pp. 527-535, 2009. 3], Transplant patients are frequently monitored in order to reduce graft immune rejection risk, minimize the immunosuppressant side effects and prevent complications (such as renal and cardiac dysfunction, infections and malignancies).
Biomarkers may play an essential predictive role in transplant rejection. A biomarker is a characteristic that is objectively measured and evaluated as an indicator of a normal biological process, pathogenic process, or pharmacological response to a therapeutic intervention. Biomarkers are used for (1) diagnosis of patients with a disease or an abnormal organ function, (2) severity of disease, (3) prognosis of a disease, and (4) monitoring of a response to a medical procedure [https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9167141/].
Recent advances in biomarkers for the diagnosis and prognosis the rejection processes of allograft organs have enabled the definition of molecules with a very narrow area of application, for example in identifying AB MR from other types of rejection mechanisms or even to discriminate rejection from other causes of graft dysfunction.
The ideal biomarkers should enable reproducible, accurate, sensitive and specific outputs concerning anticipation of allograft rejection, diagnosis of allograft rejection, define the underneath biologic mechanisms, prognosis and define the general immune status. The biomarkers should be acquired from non or minimal invasive modes (e.g., urine and serum) and should be stable while in storage and in transit, all that while presenting minimum interference by other molecules. They should be analyzed by economical and simple analytical techniques towards its routine clinical application [C. Johnson and B. Kaplan, “Biomarkers and Kidney Transplant,” Transplantation, vol. 102, no. 4, pp. 552- 553, 2018], [T. R. Fleming and J. H. Powers, “Biomarkers and Surrogate Endpoints In Clinical Trials,” Stat. Med., vol. 31, no. 25, pp. 2973-2984, 2013], [S. M. Kurian et al., “Biomarker guidelines for highdimensional genomic studies in transplantation: Adding method to the madness,” Transplantation, vol. 101, no. 3, pp. 457-463, 2017], Due to the high complexity and interrelationships of the diverse biological processes, including variability response associated to each individual, underneath graft dysfunction and loss, as the rejection mechanism, the specific response to immunosuppressive drugs and even the graft repair system, a set of biomarkers would be preferable over a single albeit unique biomarker.
For instance, document EP1905846B1 describes biomarkers for detection of renal allograft rejection and other forms of renal damage and methods for the diagnosis and/or prognosis of renal damage, including renal transplant allograft rejection, using the biomarkers. The method of diagnosis and/or prognosis relies on the determination of the presence or of the amount of a nucleic acid encoding N-acetylmuramoyl-L-alanine amidase precursor, in a sample from a patient. Additionally, EP2527473A2 provides methods for predicting the development of tolerance to a transplant, such as a kidney, using molecular markers that have different expression patterns in tolerant transplant recipients, as compared to non -tolerant or healthy, non-recipient controls. The methods for predicting tolerance in a transplant recipient for the transplant relies on the expression level of the biomarker IgKV 1 D- 13.
Several other biomarkers present in biofluids have been pointed out as resources for the detection of allograft rejection. However, continues the need for the discovery of new biomarkers that can lead to the detection, as early as possible, of allograft rejection processes.
Fourier Transform Infrared Spectroscopy (FTIRS), captures the whole molecular fingerprint of a biological sample in a highly specific and sensitive mode, as explain in the following chapters. This technique monitors the vibration between molecular bonds, as from macromolecules as proteins, nucleic acids, lipids and polysaccharides as well from small molecules as metabolites, and therefore represents an integrated mode of the previous referred omics science.
In infrared spectroscopy, when a sample is irradiated with infrared radiation, the molecules present in the sample absorb energy at specific frequencies (corresponding to their own frequencies). The energy prevenient from the IR region is not strong enough to conduct their electron excitation, inducing vibrational excitation of covalent binding molecules [D. Duygu, T. Baykal, I. Acikgoz. and K. Yildiz, “Fourier transform infrared (FT-IR) spectroscopy for biological studies,” Gazi Univ. J. Set., vol. 22, no. 3, pp. 117-121, 2009], Consequently, symmetrical molecules like monoatomic (e.g., O2, N2, H2 - formed by only one atom) and homopolar diatomic molecules (e.g., He, Ne, Ar - formed by only one type of atoms), present negligible vibrations and are therefore considered “infrared inactive” [P. R. Griffiths, “Introduction to Vibrational Spectroscopy,” in Handbook of Vibrational Spectroscopy, P. R. Griffiths, Ed. Chichester, UK: John Wiley & Sons, Ltd, 2006, pp. 1-18], [B. H. Stuart, Infrared Spectroscopy: Fundamentals and Applications, vol. 8. 2004], The final spectrum that is visualized is then a representation of the vibration of the molecules that make up a sample [D. Duygu, T. Baykal, I. Acikgoz. and K. Yildiz, “Fourier transform infrared (FT-IR) spectroscopy for biological studies,” Gazi Univ. J. Sci., vol. 22, no. 3, pp. 117-121, 2009],
The main advantage of infrared (IR) spectroscopy is that it enables the analysis of a large diversity of samples in different physical states, with a high efficiency in providing information on the purity of different compounds and identifying their structures, thus enabling the knowledge of functional groups, bonding types and molecular conformations [B. H. Stuart, Infrared Spectroscopy: Fundamentals and Applications, vol. 8. 2004], [A. A. Bunaciu, §. Fleschin, V. D. Hoang, and H. Y. Aboul-Enein, “Vibrational Spectroscopy in Body Fluids Analysis,” Crit. Rev. Anal. Chem., vol. 47, no. 1, pp. 67-75, 2017],
In the end the spectrum obtained is called the interferogram (a spectrum consisting of intensity versus acquisition time), which is then translated to the final IR spectrum (intensity versus frequency or wavelength) by the mathematical operation called Fourier-Transform [B. H. Stuart, Infrared Spectroscopy: Fundamentals and Applications, vol. 8. 2004], [A. A. Ismail, F. R. van de Voort, and J. Sedman, “Chapter 4 Fourier transform infrared spectroscopy: Principles and applications,” Tech. Instrum. Anal. Chem., vol. 18, no. C, pp. 93-139, 1997], The result of the technological step from the dispersive system into the spectrometers FT-IR led to an increased reproducibility, a reduction of acquisition time (all frequencies are evaluated simultaneously) and a better signal to noise ratio (higher thus enabling measurements with greater sensitivity).
One example of the application of FTIRS in diagnosis or prognosis is that of prostate cancer. One of the first articles on the subject, published at the end of the last century by Malins and colleagues [D. C. Malins, N. L. Polissar, and S. J. Gunselman, “Models of DNA structure achieve almost perfect discrimination between normal prostate, benign prostatic hyperplasia (BPH), and adenocarcinoma and have a high potential for predicting BPH and prostate cancer,” Proc. Natl. Acad. Sci., vol. 94, no. 1, pp. 259-264, Jan. 1997], used models based on logistic regression of infrared spectrum data and was able to distinguish normal vs. cancer and normal versus benign prostatic hyperplasia (BPH) with a sensitivity and specificity of 100%, and close to 100% for HPB vs. cancer.
Spectroscopic techniques are only as powerful as the information that can be extracted from the resulting spectral data. Chemometrics is the chemical discipline that is responsible for the information retrieval. The purpose of chemometrics is to: explore and identify patterns in the data, track properties of materials, prepare multivariate classification models and analyze model classification and prediction. Chemometrics is a field that combines statistical, mathematical and computational methods to analyze chemical data and to extract relevant information from certain chemical systems [N. D. Lourcnco. J. A. Lopes, C. F. Almeida, M. C. Sarraguca. and H. M. Pinheiro, “Bioreactor monitoring with spectroscopy and chemometrics: A review,” Anal. Bioanal. Chem., vol. 404, no. 4, pp. 1211-1237, 2012], [Y. Roggo, P. Chalus, L. Maurer, C. Lerna-Martinez, A. Edmond, and N. Jent, “A review of near infrared spectroscopy and chemometrics in pharmaceutical technologies,” J. Pharm. Biomed. Anal., vol. 44, no. 3, pp. 683-700, Jul. 2007],
Machine Learning (ML) is defined as a subset of the Artificial Intelligence (Al) domain capable of automatically learning and continuously adapting the interpretation or prediction algorithms. Robust mathematical procedures are applied by computer systems to achieve these complex tasks. Machine Learning it is a data analysis method that automates the process of creating models (with minimal human intervention). These models have the ability to interactively learn from the data provided and find patterns in it, without extensive prompts from a human as it was the norm before.
Basically, there are 3 types of machine learning algorithms: Supervised learning, Unsupervised Learning and Reinforced learning. In the case of supervised learning, labeled parameters are provided and then used to build a model that will try to predict the other labels. Unsupervised Learning, only parameters without any type of label are provided, in which case the model will try to find subgroups within the data that have some similarity between them. Finally, Reinforced learning uses algorithms to learn how to perform actions based on the environment, that is, the system interacts with the environment and begins a series of optimizations.
There are a multitude of algorithms used in machine learning, each with a specific purpose, with different algorithms best serving specific models, as there is no single algorithm that will perform consistently well for all datasets. Regardless, they are all dependent on high computational power, speed or accuracy. A few of the most popular algorithms are: Linear Regression, Logistic regression, Decision Trees, Naive Bayes classification, AdaBoost, SVM, PCA, k-NN, Random Forest. Every single one of these algorithms can be classified according to the type of learning supervised and unsupervised subcategories.
Of the many applications of Machine Learning, medicine has been one the fields in which the introduction of these algorithms as in recent decades observed an exponential increase. Several factors have contributed to this: the introduction of wearable devices and sensors that allow healthcare professionals to access patient data in real time, the growing amount of health data, which in combination with this technology has helped medical experts analyze data to identify trends or alerts, leading to improved diagnostics and treatments [M. Hajiloo, H. R. Rabiee, and M. Anooshahpour, “Fuzzy support vector machine: an efficient rule-based classification technique for microarrays.,” BMC bioinformatics, vol. 14 Suppl 1. p. S4, 2013],
For instance, document WO2022229303A1 describes a computer-implemented diagnostic method learning for determining clinical interpretation of renal graft alterations by applying a trained machine-learning model and a method for training a model for determining clinical categories based on renal transplant lesions, clinical signs and routine laboratory test results.
As explained above, the field of organ transplantation is of utmost importance in medicine, increasing life expectancies, improving the quality of life and it remains the best therapy for terminal and irreversible organ failure. However, despite immunosuppression medications, a significant percentage of patients will experience at least one episode of allograft rejection. Unfortunately, the existing prior art conventional monitoring techniques of allograft rejection processes do not allow an effective anticipation of the rejection processes and an early diagnose and prognose of rejection. Additionally, although biomarkers cannot anticipate rejection processes, i.e. only detects processes already running, or evaluate the patient immune overall status regarding sub or over-immunosuppressive therapies, they are also associated with high costs and do not preclude the need for a conventional histological analysis. Besides these limitations, the need for the surveillance biopsies (with the consequent associated morbidity, sampling error, and cost) reinforces the need for new biomarkers and methods for the diagnosis and/or prognosis of allograft rejection.
Despite significant advances in the epidemiology of organ transplantation, prognostication remains a major clinical challenge. Unfortunately, there is no reliable method to predict allograft rejection. The discovery of methods to aid in clinical risk prediction for recovery after organ transplantation would represent a significant advance over current practice.
Therefore, the nowadays existing limitations emphasize the critical need to discover new biomarkers and methods that can lead to the prediction, as early as possible, of allograft rejection processes and based on minimal or non-invasive methods.
BRIEF DESCRIPTION OF THE DRAWINGS
The invention will now be described in further detail with reference to the accompanying drawings in which:
Figure 1 illustrates a t-SNE based on 2nd derivative of top ranked spectral bands. Samples marked with light gray are samples prior to rejection, and samples marked with dark gray are samples after rejection diagnosis (biopsy proven).
Figure 2 represents the average spectra of serum of transplanted patients presenting rejection (n=369, represented in dark gray) versus average spectra of serum of transplanted patients without rejection process and non -transplanted patients (n=213 represented in light gray), and the corresponding 2nd derivative spectra, with Savitzky-Golay filter (9 smoothing points). X axis represents spectral band (cm1), Y axis represents absorbance.
Figure 3 illustrates a t-SNE plot of 2nd derivative of spectral regions 400 to 1000 and 3300 to 3600 cm'1 from serum of transplanted patients presenting rejection (n=369, and represented as Yes and in dark gray) versus average spectra of serum of transplanted patients without rejection (n= 213, and represented as No and in mid gray) and non -transplanted and healthy controls (n= 38, and represented as He and in light gray).
Figure 4 represents t-SNE of 2nd derivative with Savitzky-Golay fdter (9 smoothing points) of serum spectra. Including non- transplanted and transplanted patients classified as non -rejection in light gray (N)) and rejection Yes in dark gray (Y)).
SUMMARY OF THE INVENTION
The present invention is directed to a prognostic method for determining a probability of allograft rejection comprising the steps of: i) providing an isolated biological fluid sample to be tested; ii) diluting the biological sample in water in a range from 1 :5 to 1: 15; iii) analysing the diluted sample in a Fourier-transform infrared spectrometer; iv) collecting spectral data from the spectrometer; v) submitting said spectral data in a spectral data machine -learning model; v) comparing said spectral data with machine -learning model spectral data; vi) determining the probability of allograft rejection by finding a match between the spectral data and the machine-learning model spectral data.
The present invention also relates to a method for determining a probability of efficiency of an allograft rejection rescue therapy comprising the steps of: i) providing an isolated biological fluid sample to be tested; ii) diluting the biological sample in water in a range from 1 :5 to 1: 15; iii) analysing the diluted sample in a Fourier-transform infrared spectrometer; iv) collecting spectral data from the spectrometer; v) submitting said spectral data in a spectral data machine -learning model; v) comparing said spectral data with machine -learning model spectral data; vi) determining the probability of allograft rejection rescue therapy by finding a match between the spectral data and the machine-learning model spectral data.
In one aspect, the said isolated biological fluid sample is selected from the group comprising: blood sample, serum sample, plasma sample or urine sample.
In another aspect, said isolated biological fluid sample is a serum sample.
In a further aspect, the biological sample is diluted at 1: 10 in water. In another aspect of the present invention, the allograft rejection is a hyperacute, acute or chronic rejection.
In further aspect, the allograft is selected from the group comprising: kidney allograft, liver allograft, heart allograft, lungs allograft, or pancreas allograft.
In another aspect, the allograft is a kidney allograft.
In further aspect of the present invention, the method is implemented before allograft transplantation or after allograft transplantation.
DETAILED DESCRIPTION OF THE INVENTION
The present invention refers to a prognostic method for determining a probability of allograft rejection using an infrared spectroscopy-based method associated with a multivariate data analysis and machine learning techniques.
The Fourier Transform Infrared Spectroscopy (FTIRS) was applied as an analytical technique to retrieve the whole molecular profile of biofluids, such as serum, that associated to multivariate data analysis and a machine -learning model, could enable to correlate the spectral data with diagnosis and prognosis outputs.
The diagnostic and/or prognostic methods that currently exist are based on DNA analysis, contrary to the method of the invention, where the cellular metabolism is analysed. This cellular metabolism generates huge amounts of data which is highly complex to analyze and extract relevant information. In order to address this problem, the inventors applied machine learning methods to biological data sets.
FTIR spectroscopy of a biofluid sample coupled with unsupervised and supervised processing multivariate data analysis enabled to develop good predictive models of the rejection diagnosis and prognosis, the risk of rejection before transplantation and the efficiency of the organ rescue treatments.
In the context of the present invention, the expression "spectral data machine -learning model" as used herein refers to a machine learning model designed to recognize spectral patterns based on a previous spectral database (training data). By definition, a machine learning (ML) is the study of different algorithms that can improve automatically through experience and old data and build a model. The ML learning algorithm discovers patterns within the training data of the spectral database, and it outputs a machine-learning model which captures these patterns and makes predictions on new data and provides an output in response.
The term “biomarker" as used herein refers to spectral patterns discovered and identified by the machine-learning model after a FTIRS analysis of biological fluid samples, which are indicative of allograft rejection process. The identification of these spectral patterns is correlated with the incidence or risk of incidence of allograft rejection.
The term "subject" as used herein refers to a human or non-human organism. Thus, the methods described herein are applicable to both human and veterinary disease. Preferred subjects are humans, and most preferably "patients," which as used herein refers to living humans that are receiving medical care for a disease or condition. This includes persons with no defined illness who are being investigated for signs of pathology.
The term "biological fluid sample” or "biofluid sample" as used herein refers to a sample of bodily fluid obtained for the purpose of diagnosis, prognosis, classification or evaluation of a subject of interest, such as a patient or transplant donor. In certain embodiments, such a sample may be obtained for the purpose of determining the outcome of an ongoing condition or the effect of a treatment regimen on a condition. Preferred biofluid samples include blood, serum, plasma, cerebrospinal fluid, urine, saliva, sputum, and pleural effusions. In addition, one of skill in the art would realize that certain biofluid samples would be more readily analyzed following a fractionation or purification procedure, for example, separation of whole blood into serum or plasma components.
The term "diagnosis" as used herein, refers to methods by which trained medical personnel can estimate and/or determine the probability (i.e., for example, a likelihood) of whether or not a patient is suffering from a given disease or condition.
The term "prognosis" as used herein refers to a probability that a specific clinical outcome will occur, for example, a poor or good outcome. For example, a negative prognosis, or poor outcome, in a patient is associated with an increased probability of poorer graft survival and a positive prognosis, or good outcome, in a patient is associated with an increased probability of a long-term good graft survival.
As above-mentioned, organ transplantation, when possible, is the only mode of therapy for most end-stage organ failure affecting kidneys, liver, heart, lungs, and pancreas. However, one of its major problems is allograft rejection. The present invention aims to promote new biomarkers discovery for diagnosis and prognosis the rejection processes of allograft organs, in a rapid, economic but also sensitive and specific mode. It was also aimed to develop a new methodology for an earlier determination of the risk of allograft rejection before and after transplantation. Based on Fourier Transformed Infrared (FTIR) spectroscopy of biological samples associated to multivariate data analysis and machine learning techniques, the inventors developed a new prognosis methodology, which allows an increased monitorization of organ transplanted patients, the identification of critical patients with an increased risk of rejection processes, and allows the adjustment of immunotherapies for organ rescue. These will lead to disrupt modes of management of these type of patients towards a significant increase of quality of life and even of life expectancy, and in a highly economic mode.
The inventors developed a prognostic method for the determination of the probability of allograft rejection comprising the following steps:
- providing an isolated biological fluid sample to be tested;
- diluting the biological sample in water in a range from 1 :5 to 1: 15;
- analysing the diluted sample in a Fourier-transform infrared spectrometer;
- collecting spectral data from the spectrometer;
- submitting said spectral data in a spectral data machine-learning model;
- comparing said spectral data with machine -learning model spectral data; and
- determining the probability of allograft rejection by finding a match between the spectral data and the machine -learning model spectral data.
The isolated biological fluid sample is selected from the group comprising: blood sample, serum sample, plasma sample or urine sample. In a preferred embodiment of the invention, the isolated biological fluid sample is a serum sample.
These types of samples allow the study of a much higher quantity of analytes (complete sets of genes, transcripts, proteins and/or metabolites at a given state) and it will be possible to discover specific sets of biomarkers that will enable to discriminate graft rejection from other diseases and allow at the same time to define the rejection mechanisms and be able to anticipate clinical rejection. Due to the high complexity of the processes associated to graft rejection, the approach is to consider this multidimensional data. Therefore, to search for suitable biomarkers for diagnosis and prognosis, this untargeted approach is applied directly to biofluids, as for example, serum. The release of analytes from cells to biofluids can be a result from its natural secretions (e.g. as it happens with chemokines and cytokines), as the case of exosomes or even due to graft injury, apoptosis and necrosis from the graft or recipient’s cells.
When working with biological samples, especially those with a high degree of intra-individual variability, it’s always important to further test the dilution effect of the biofluid (e.g. serum), on the overall test performance. Therefore, the dilution degree of the biological fluid sample was optimized before analysing the sample in the Fourier-transform infrared spectrometer. The optimal dilution ratio of the biological sample in water is in the range from 1:5 to 1: 15, preferably, 1: 10.
The FTIRS was applied as an analytical technique to retrieve the whole molecular profde of the biofluid sample, as for example serum, that associated to multivariate data analysis and machine -learning techniques, enable to correlate the spectral data with diagnosis and prognosis outputs.
After performing the analysis of the biological fluid samples in the Fourier-transform infrared spectrometer, the spectral data collect from the spectrometer was submitted in a spectral data machinelearning model.
Since the spectral data collected from the spectrometer is highly complex to analyze and difficult to extract the relevant information, in order to address this problem, the inventors applied a machine - learning model with the aim to find patterns among the spectral data retrieved from the spectrometer. These patterns that the spectral data machine -learning model will search and identify are the biomarkers that will allow the determination of the probability of allograft rejection.
This spectral data machine -learning model was developed by the inventors based on an Artificial Intelligence and with a previous database construction with biofluid samples spectra associated with the patient data. The aim of this spectral data machine -learning model is to find patterns among the spectral data provided to it. This model has the ability to interactively learn from the spectral data provided and find patterns in it.
The development of the spectral data machine -learning model was made following the next steps:
1. Data collection of biofluid samples, and the correlated patient data concerning demography and clinical;
2. Optimization of biofluid samples processing for the acquisition of the infrared spectrum and spectral acquisition conditions;
3. Database construction with biofluid spectra associated with patient data;
4. Training process of the machine learning (ML) using the spectra database as training data to develop the models to predict diagnosis and prognosis of rejection and efficiency of drugs treatment.
In the first step the data collected was from:
- healthy non-transplanted participants; - transplanted patients without rejection processes; and
- transplanted patients with rejection processes.
The biofluid samples of those participants/patients were used for the FTIRS analysis with a previous step of optimization of the biofluid samples processing, namely the dilution rate optimization of the samples. The biofluid spectra obtained by the FTIRS analysis were then used for the construction of a database, where the FTIRS results were associated with patient data.
This spectra database is the training data for the training process of the machine learning (ML). The data from the spectra database was introduced into the ML and used for the training process of the ML. The machine learning model of the invention is therefore trained over a set of spectral data, and then is provided an algorithm to reason over data, extract the pattern from feed data and learn from those data. Once the model gets trained, it can can be used to predict the unseen dataset.
Since it’s known the outcome of the participants/patients, whose biofluid samples were used for the FTIRS analysis and following spectral database construction, the Machine -learning model identifies the common spectral patterns among the healthy non-transplanted participants, the transplanted patients without rejection processes and among the transplanted patients with rejection processes. Thus, the spectral patterns are identified by the Machine -learning model for the 3 different groups of participants/patients .
Therefore, through the spectral database, the Machine -learning model is taught to identify the spectral patterns for the 3 different groups of participants/patients, and afterwards the spectral data machine-learning model is ready to identify these spectral patterns in other unseen biofluid samples, making predictions on new data and providing an output in response.
In the prognosis method of the invention, after submitting the spectral data, of the unseen biofluid sample, into the spectral data machine -learning model, it will be compared with the machinelearning model spectral data, i.e. the machine -learning model will search for similarities between both spectral data (the spectral data of the unseen biofluid sample and the spectral data of the database) in order to identify spectral patterns that relate to an allograft rejection output, and thus determining the probability of allograft rejection.
The prognostic method of the invention could be applied for determining the probability of hyperacute, acute or chronic allograft rejection.
The allograft is selected from the group comprising: kidney allograft, liver allograft, heart allograft, lungs allograft, or pancreas allograft.
The described invention allows to predict which patients will develop a rejection process, and also allows to predict the risk of rejection process, as early as 120 days before it was detected in biopsies. Even before organ transplanting it is also possible to predict the risk of rejection using the method of the invention.
Additionally, patients under immunotherapy to minimize the organ lost, the method of the invention allows to predict the efficiency of the organ rescue treatment. The invention also relates to a method for determining a probability of efficiency of an allograft rejection rescue therapy comprising the steps of: i) providing an isolated biological fluid sample to be tested; ii) diluting the biological sample in water in a range from 1 :5 to 1: 15; iii) analysing the diluted sample in a Fourier-transform infrared spectrometer; iv) collecting spectral data from the spectrometer; v) submitting said spectral data in a spectral data machine -learning model; v) comparing said spectral data with machine -learning model spectral data; vi) determining the probability of allograft rejection rescue therapy by finding a match between the spectral data and the machine-learning model spectral data.
Therefore, the method of the invention can be applied to the following clinical situations:
- Diagnosis of the rejection process;
- Prognosis of rejection post-transplant;
- Predict the risk of rejection before transplantation; and
- Predict the efficacy of pharmacological rescue therapies.
The results achieved by the inventors clearly point FTIRS Analysis as a new tool for the identification of patients at risk of having a rejection event. FTIRS analysis of biological fluid samples associated with a machine -learning model enabled to predict if a patient will develop a rejection process. Therefore, those patients present a cellular precondition detectable in their biofluid FTIR spectrum. With the method of the invention, it’s possible to identify patients at risk of rejection previously to transplant, based on their biofluid FTIR spectrum.
The proposed solution enables an early prediction of allograft rejection processes in a sensitive, specific and accurate mode, based on a non-invasive methodology. The method proposed, due to simplicity, speed and economics, could increase the monitoring of the patients, identify critical patients with an increased risk of rejection processes, and to promoting the adjustment of immunotherapies for organ rescue. These could lead to disrupt modes of management of the patients towards a significant increase of quality of life and even of life expectancy, and in a highly economic mode.
The invention will now be described in more detail with reference to the following examples, which are provided for illustrative purposes only, and are not intended to limit the scope of the invention.
EXAMPLES
Despite the high complexity associated with the mechanisms of rejection in solid organ transplantation, with the present invention it is possible to apply FTIRS associated to uni and multivariate data analysis to monitor and predict allograft rejection process.
1. Materials and Methods
The inventors performed a study that comprehended 620 serum samples from non-transplanted and renal transplanted patients with and without rejection, as follows:
• Group A: 369 serum samples from 29 renal transplanted patients with rejection processes;
• Group B: 213 serum samples from 30 renal transplanted patients without rejection processes;
• Group C: 38 serum samples from 38 non-transplanted and healthy controls.
All the transplanted patients, were transplanted across a negative crossmatch performed by complement dependent cytotoxicity as described by Kumar et al. [A. Kumar, A. Mohiuddin, A. Sharma, M. El Kosi, and A. Halawa, “An Update on Crossmatch Techniques in Transplantation,” J. Kidney, vol. 03, no. 04, 2017] and the lymphocytes originated from lymph node in the case of deceased donor and peripheral blood in the case of living donor. To be included in the rejection group, patients had to have at least one biopsy with a proven rejection.
All health controls used had at least, one medical interview attesting their good health.
1.1 FTIRS and spectra data analysis
The serum from patients were analyzed by FTIRS. 25 pL of serum diluted at 1:4 or 1: 10 in milli-Q- Water were transferred to a 96-wells Si micro plate (Bruker) and then dehydrated for about 2.5 h, in a desiccator under vacuum. Spectral data was collected using a FTIRS (Vertex 70, Bruker) equipped with an HTS-XT (Bruker) accessory. Each spectrum represented 64 coadded scans, with a 2cm1 resolution and was collected in transmission mode, between 400 and 4000 cm The first well of the 96-wells plate did not contain a sample and the corresponding spectra was acquired and used as background, according to the HTS-XT manufacturer.
OPUS software was used to transfer the spectral data acquired from the spectrometer to other software. Pre-processing and processing of the collected data was done by several specialized software, as from OPUS (Bruker), Unscrambler® X (CAMO), Matlab R2012b (MathWorks, Natick, MA, USA) or Orange software (Ujubljana University).
1.2 Ratios of spectral peaks determination
Ratios between spectral peaks were determined as follows:
- The second derivative spectra were determined from the average spectra of a defined set of a population. This was conducted by the Orange software;
- It was compared the negative peaks of second derivative spectra of different populations based on Visual Basic for Applications script in Microsoft Excel. It was considered the peaks to determine ratios of spectral peaks, based on peaks that were different on the spectral derivative;
- Diverse ratios between the peaks identified in the previous points were determined based on all data of the second derivative spectra.
- The ratios of peaks between two samples were evaluated by Mann Whitney test, based on Graphpad Prism v8.
1.3 Statistical Evaluation
The statistics of human population analysis was conducted on GraphPad Prism v8 (USA) and Orange software. Chi-square, t- student and Mann-Whitney U test for spectral band significance was conducted in the open source Orange software. This last one was also used to identify the top 100 spectral bands, i.e. the ones that were statistically more relevant into the clustering of stable versus transplanted patients with signs of rejection.
2. Results and Discussion
2.1 Preliminary study to detect rejection and optimization of serum sample processing and spectral preprocessing methods
The first step of the study was the optimization of the sample processing and the optimization of spectral pre-processing methods. First was performed a pilot study based on 8 patient’s samples (4 without rejection and 4 with rejection processes). All 8 patients received a kidney allograft, where one also received a liver allograft. Next, it was also performed an evaluation of the effect of the dilution degree of serum and spectra pre-processing methods. Only by increasing the dimension of serum samples it was possible to subsequently optimize the dilution degree of serum and the spectra pre-processing methods. Based on the results obtained with this part of the study, all remaining analysis concerning the use of spectra from serum, were based on serum diluted at 1/10 in water, and based on a pre-processing of an 2nd derivative with a 2nd order polynomial, and a Savitzky-Golay filter with a 9-points window.
2.1.1 Pilot Study
Serum was diluted at 1:4 in water, and spectra were obtained from 64 coadded scans, with a 2cm1 resolution in transmission mode, between 400 and 4000 Were conducted quintuplicate of dilutions for each serum sample, and consequently were obtained quintuplicate spectra for each sample. In this pilot study, 8 patients with renal allograft (4 without rejection and 4 with rejection processes) were included. It was possible to discriminate from the serum spectrum, the patients with and without rejection, as highlighted by a region of the 2nd derivative spectra, where e.g. at 630 cm'1 wavenumber it was possible to discriminate all samples associated to rejection from samples not associated with rejection. For this spectral peak, an 100% sensitivity (i.e. a ratio of true positive responses) and a 100% specificity (i.e., a ratio of true negative responses) were observed.
From a Principal component analysis (PCA) score-plot, a separation is clearly visible between the group of samples from patients with and without rejection. Interestingly, samples from the single patient with a double (liver and kidney) transplant (and with indication for rejection) is adequately classified in the group of patients with rejection but in a very distinct subgroup according to the clinical specificity of this patient.
On a Hierarchical Cluster Analysis HCA it was also possible to discriminate the samples of the non-rejection patients from the rejection patients. It was also observed that the patient with multiorgan transplant was classified into a subgroup of the remaining rejection patients according to the PCA. In the cluster analysis it was also observed that one of the non-rejection patients, although separated from the rejection patient samples, is in a group very close to one of the rejection patients. Out of curiosity, this patient presented DSA in spite of no detection of rejection on biopsy, probably predicting a future rejection.
Considering the results obtained, it is clear that FTIRS associated to multivariate data analysis presents a high potential for a diagnosis and prognosis of rejection processes. 2.1.2 Optimization of serum sample processing
The spectra obtained with serum samples diluted at 1:4 from the 8 previous transplanted patients, presents maximum absorbances values lower than 1. However, using spectra from 59 transplanted patients, it was observed diverse spectra with absorbance maximum higher than 1, and some even with absorbance values of 5. This high variability of spectra from transplanted patients is probably due to diversity of the therapeutics (e.g. plasmapheresis removes several components from serum while immunoglobulin therapy increases the amount of serum proteins). All serum samples were also diluted at 1 : 10 in water and the spectra was subsequently acquired.
The score plot of PCA of 2nd derivative spectra of serum samples from transplanted patients diluted at 1: 10 showed more distinct clusters for rejection patients in relation to the non-rejection patients, in contrary to what was observed with serum diluted at 1:4. Therefore, in the subsequent analysis, serum was diluted at 1: 10 in water.
2.1.3 Optimization of spectral pre-processing methods
Usually in spectroscopy, pre-processing methods are critical, as to minimize physical interferences as scattering phenomena due to dehydrated fdms while increasing chemical information, by e.g. resolving overlapping peaks by derivatives. The following pre-processing methods were applied to spectra of serum diluted at 1:4 and 1: 10 baseline correction, normalization and first and second derivatives. The effect of these pre-processing methods was evaluated in the PCA score plot. From the pre-processing methods evaluated, the spectra pre-processed by second derivative of serum samples diluted at 1: 10 resulted in the best separation of samples from patients with and without rejection processes. These results corroborate the observations of the previous study.
A subsequent study based on quintuplicates of serum diluted at 1: 10 of 21 patients with a rejection diagnosed by biopsy and 18 patients without clinical evidence of rejection was conducted. It was observed a clear discrimination between samples from patients without rejection (stable) in relation to samples from patients with rejection processes. Only one replicate of one “rejection” sample was misclassified near a rejection sample. However, this replicate was an outlier as the remaining 4 replicas were well classified.
2.2 Rejection Diagnosis
To evaluate the feasibility to identify when the rejection occurred, the following 360 serum samples from transplanted patients with rejection were used: 151 serum samples prior to the detection of rejection in biopsy, and classified as prerejection (P);
- 209 serum samples taken after the rejection process was detected in biopsy and designated as post-rejection (Y).
From the study with the above-mentioned serum samples, the following four spectral regions were highlighted as presenting high molecular information: 400 to 1000 cm1, 1000 to 1800 cm1, 2700 to 3200 cm'1 and 3200 to 4000 cm1.
An ANOVA was conducted to identify the top 100 spectral bands of the second derivative that discriminate between the sample of pre-rejection (n=151) from the sample of post-rejection (n=209).
The 2nd derivative top 100 ranked spectral bands identified were subsequently used to discriminate the two types of samples (pre and post-rejection), by:
- Multivariate data analysis by /-distributed Stochastic Neighbor Embedding (t-SNE) and Decision Tree analysis;
- Univariate data analysis.
In the /-SNE plot based on 2nd derivative of 100 selected spectral bands, 4 main data clusters were observed (Figure 1), where the cluster designated by D, presents the majority of pre-rejection samples. The remaining 3 clusters present mostly post -rejection samples. In cluster A (n=79) and C (n=88), most samples were obtained from days to 3 months before the rejection diagnosis. In cluster B (n=71), most samples were obtained from 6 months to 3 years before the rejection diagnosis. Interestingly, most misclassified samples in this group (n=27) are from patients that will develop chronical rejection. It was observed that many samples from cluster B are from patients with AR (n=41).
With this study, it was possible to predict the rejection among transplanted patients with rejection processes, by Support vector machine (SVM), an algorithm used in machine learning, with an area under the curve (AUC) 0.804, a positive predictive value of 94.26%, and a sensitivity and a specificity of 71.90% and 86.05% respectively.
2.3 Prognosis of rejection after transplantation
The following 620 serum samples from non-transplanted and transplanted patients with and without rejection, respectively, were used:
- 38 serum samples from 38 non-transplanted and healthy controls (He);
- 213 serum samples from 30 transplanted patients without rejection processes (N); - 369 serum samples from 29 transplanted patients, that will develop a rejection processes (Y).
Two parallel studies were conducted, one based on a control group including transplanted patients without rejection processes, another where the control group included this last group plus healthy non transplanted volunteers.
For the 369 serum samples from patients presenting rejection processes, it was not considered the time from which the samples were taken in relation to the period were the rejection was detected. That is, these samples could be in fact associated with periods in time in which patients did not presented rejection process.
The average spectra of all samples of stable transplanted patients presented differences in relation to the average spectrum of transplanted patients with rejection process. These differences are highlighted in the 2nd derivative spectra of these averaged spectra (Figure 2).
In the t-SNE plot of 2nd derivative of spectra regions, including 400 to 1000 cm'1 and 3300 to 3600 cm1, from serum of transplanted patients presenting rejection versus average spectra of serum of transplanted patients without rejection and non-transplanted and healthy controls (Figure 3), it was observed several defined data clusters in the 2nd derivative spectra, as 2 clusters mostly containing samples from transplanted stable patients, and a cluster presenting mostly non-transplanted patients.
With this study, was possible to show that supervisioned machine learning models associated to serum FTIRS, enabled to predict if a patient will develop a rejection process, as obtained with a k-nearest neighbors algorithm (k-NN) model, an algorithm used in machine learning, that led to an AUC of 0.916 and a sensitivity of 84.7% and a specificity of 95.1%.
2.4 Predict the risk of rejection before transplantation
The following 97 serum samples from non-transplanted and transplanted patients with and without rejection, were used:
- 38 serum samples from 38 non transplanted and healthy controls (He);
- 30 serum samples from 30 transplanted patients without rejection processes (N);
- 29 serum samples from 29 transplanted patients with rejection processes (Y).
The zero-hour time or day zero sample is a sample collected at the time when the patient arrives in the transplant unit, i.e. before transplantation. This sample is neutral in terms of therapeutics, as at this moment the patient has no induction therapy as no transplant was yet performed.
In the t-SNE plot based on serum 2nd derivative spectra, including non-transplanted volunteers or only transplanted patients (Figure 4), it was observed the formation of clusters associated to samples from patients with rejection processes.
With this study, the inventors concluded that it’s possible to identify patients at risk of rejection previously to transplant, based on their serum FTIR spectrum. The best predictive model, based on Neural Network of septrum 2nd derivative spectra with transplanted patients only, an AUC of 0.97 and a sensitivity and specificity of 93.33% and 96.55%, respectively, were achieved.
2.5 Therapies Prognosis of Organ Rescue
In order to identify the outcome of rejection treatment in a premature manner, 20 samples were used out of the 29-rejection group of patients. The 9 remaining patients were excluded mainly due to inconclusive treatment results or unfinished treatment. The 20 samples included were all selected from the follow-up samples of these patients, and they were immediately prior to the initiation of salvage treatment of the rejected organ. Of the overall samples, 9 were from patients that positively responded to treatment and 11 were from patients in which the treatment did not work. In order to evaluate the impact of the introduction of variability, 20 health controls were also included. The overall distribution included 11 acute rejection (AR) and 9 chronic rejection (CR), and can be divided into: 5 acute cellular, 6 acute humoral and 9 chronical humoral.
In order to achieve our objective, identifying biomarkers for the therapeutics response in patients with rejection events, the data was analyzed in 4 groups:
- group 1, unprocessed spectra from patients and health controls;
- group 2, 2nd derivative spectra from patients and healthy controls;
- group 3, unprocessed spectra from patients;
- group 4, 2nd derivative spectra from patients.
In this study, it was considered the whole spectra and the 100 best classification bands. Due to the low number of samples when compared to the number of variables, a lower number of bands (100) in relation to the total 3724 of the whole spectra was also considered. The Rank widget included in the Orange software considers class-labeled datasets (classification or regression) and scores the attributes according to their correlation with the class. The spectra were ranked by importance/contribution to the target by ANOVA. Interesting to note that only 7 spectral bands are shared in the top 100 spectral bands (patients only versus patients and healthy controls). FTIRS coupled with machine learning algorithms can be used as a tool to predict the efficiency of the rejection rescue therapies, with a Naive Bayes (an algorithm used in machine learning) performing with a classification accuracy of 0.95.