Machine learning for differentiating multiple diseases

Abstract

The present disclosure provides a personalized predictive risk assessment tool for determining the risk of a variety of diseases in a subject using microbiome. At present, only one disease or health condition can be detected by a risk prediction test using microbiome once. By simultaneously determining multiple diseases, the disclosed techniques may provide a cost-effective approach to support clinical decision making and thus help improve disease prevention and management.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application is a PCT application of U.S. Provisional Patent Application No. 63 / 405,311, filed on September 9, 2022, entitled “Machine Learning For Differentiating Among Multiple Diseases,” and claims the benefit of the same, which is incorporated herein by reference in its entirety for all purposes.

[0003] background

[0004] Fecal microbiome-based analyses can serve as non-invasive methods for detecting human disease, but may be limited by shared microbial signals between different disease phenotypes. Existing risk prediction models using metagenomics data can only distinguish diseases dichotomously. Even when reference data for multiple diseases are available, the risk of each disease can only be determined one at a time. Such a risk prediction model may involve extracting data from a database for subjects with a first disease and data for healthy controls. The database data can then be compared to data from a test subject. This process needs to be repeated by extracting data from subjects with a second disease and data from healthy controls, which are then compared to data from the test subject. Therefore, a method is needed that can determine the risk of multiple diseases simultaneously.

[0005] Brief Overview

[0006] The present disclosure provides technology for personalized predictive risk assessment tools that use microbiome to determine the risk of multiple diseases in a subject. Current risk prediction tests using microbiome can only detect one disease or health condition at a time. By determining multiple diseases simultaneously, the disclosed technology can provide a cost-effective method to support clinical decision making and thus help improve disease prevention and management.

[0007] The technology includes generating a training data set of objects with multiple known health classifications. For each cohort of objects: the relative abundance of DNA fragments of each bacterial species can be measured, and the bacterial species can correspond to the bacterial species in the stool sample of each object. Each object in the cohort can have a health classification, wherein each cohort corresponds to a health classification including health and multiple conditions. At least 10 bacterial species are present in objects greater than a specified percentage (e.g., at least 5%). For each object, a feature vector containing the relative abundance of bacterial species can be generated. The training data can be used to train a multi-class machine learning model. The training data can include the known health classifications of the object and the feature vectors of the object. The multi-class machine learning model provides the probability of each of the M health classifications. Training optimizes the sensitivity and specificity for determining the correct health status by achieving the highest average AUC for the M health classifications.

[0008] These and other embodiments of the present disclosure are described in detail below.For example, other embodiments relate to systems, devices, and computer-readable media related to the methods described herein.

[0009] BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a diagram of the framework for dataset partitioning, model training, and independent validation according to the implementation scheme.

[0011] Figure 2A is a flowchart showing a summary of generating a multi-class machine learning model according to an implementation scheme.

[0012] Figure 2B is a flow chart showing a technique for applying a model to predict risk of a predetermined health condition in a subject, according to an embodiment.

[0013] Figure 3A is a table showing a demographic summary of individuals who provided stool samples according to an embodiment.

[0014] Figure 3B is a graph showing the alpha diversity of bacteria in a cohort according to an embodiment.

[0015] Figure 3C is a graph showing the richness of fecal biome bacteria in a cohort according to an embodiment.

[0016] Figure 3D is a graph showing associations of fecal biome bacterial species with health status phenotypes according to an embodiment.

[0017] Figures 4A-4C Differences in fecal microbiome between health conditions according to an embodiment are shown.

[0018] Figure 5A A graph showing a receiver operating characteristic (ROC) curve of a random forest (RF) multi-class classifier according to an embodiment.

[0019] Figure 5B A graph showing the threshold of a random forest (RF) multi-class classifier at the highest Youden index according to an embodiment.

[0020] Figures 6A-6D Shown are graphs of the area under the receiver operating characteristic curve (AUC) and the area under the precision-recall curve (AUPR) for a model trained to classify health conditions according to an embodiment.

[0021] Figures 7A-7I A graph showing a probability distribution resulting from a trained random forest (RF) multi-class classifier, according to an embodiment.

[0022] Figures 8A-8B The performance of a random forest (RF) multi-class classifier according to an embodiment in one-to-one differentiation of multiple health conditions is shown.

[0023] Figures 9A-9B is a graph showing the performance of a random forest (RF) multi-class classifier stratified by age.

[0024] Figures 10A-10D The performance of a random forest (RF) multi-class classifier trained and tested on a balanced cohort size according to an embodiment is shown.

[0025] Figures 11A-11B The performance of the Random Forest (RF) multi-class classifier using different numbers of features is shown.

[0026] Figures 12A-12D Independent validation of the performance of a random forest (RF) multi-class classifier according to an embodiment is shown.

[0027] Figures 13A-13C The performance of a random forest (RF) multi-class classifier using the first 50 features according to an embodiment is shown.

[0028] Fig.14 is a chart showing microbial species associated with health states or different health conditions according to an embodiment.

[0029] Fig.15 is a diagram of a method for training a model to predict a health condition according to an implementation scheme.

[0030] Fig.16 is a diagram of a method for predicting a health condition according to an embodiment.

[0031] Fig.17 A block diagram of a computer system according to an embodiment of the present disclosure is shown.

[0032] Fig.18 A measurement system according to an embodiment of the present disclosure is shown.

[0033] The appendix shows the following table.

[0034] Table 1 shows a list of 325 bacterial species according to an embodiment.

[0035] Table 2 shows a list of 50 bacterial species according to an embodiment.

[0036] Table 3 shows a list of 10 bacterial species according to embodiments.

[0037] Table 4 shows a list of bacteria that can be used to treat diseases according to embodiments.

[0038] the term

[0039] The term "sample" or "biological sample" refers to any sample collected from an object suspected of having a health condition and containing one or more target nucleic acid molecules. The biological sample can be a body fluid, such as blood, plasma, serum, urine, vaginal fluid, fluid from hydrocele (e.g., testicular), vaginal washings, pleural fluid, ascites, cerebrospinal fluid, saliva, sweat, tears, sputum, bronchoalveolar lavage fluid, discharge from the nipple, aspirate from different parts of the body (e.g., thyroid, breast), intraocular fluid (e.g., aqueous humor), etc. Fecal (feces) samples can also be used. In various embodiments, most of the DNA in a biological sample (e.g., a plasma sample obtained via a centrifugation protocol) that has been enriched for cell-free DNA can be cell-free, for example, greater than 50%, 60%, 70%, 80%, 90%, 95%, or 99% of the DNA can be cell-free. Centrifugation protocols for enriching cell-free DNA from biological samples may include, for example, centrifuging the biological sample at 3,000 g x 10 minutes, obtaining a fluid portion of the centrifuged sample, and re-centrifuging at, for example, 30,000 g for 10 minutes to remove residual cells. As part of the biological sample analysis, a statistically significant number of cell-free DNA molecules (e.g., to provide accurate measurements) may be analyzed for the biological sample. In some embodiments, at least 1,000 cell-free DNA molecules are analyzed. In other embodiments, at least 10,000 or 50,000 or 100,000 or 500,000 or 1,000,000 or 5,000,000 cell-free DNA molecules or more may be analyzed. At least the same number of sequence reads may be analyzed.

[0040] The terms "reference" and "reference genome" refer to generally refer to a haploid or diploid genome, with which the sequence reads from a biological sample can be compared and compared. For a haploid genome, there is only one nucleotide at each locus. For a diploid genome, a heterozygous locus can be identified, such a locus has two alleles, wherein either allele can be allowed to match for comparison with the locus. A reference genome can be a reference microbial genome, a genome assembled by a metagenome, or a species-level genome box corresponding to a specific microbial species, for example, by including one or more microbial genomes, a genome assembled by a metagenome, or a genome box at the species level.

[0041] The phrases "healthy," "control," and "normal" are used interchangeably and generally refer to a subject in good health. Such a subject exhibits the absence of any health condition or disease. A "healthy individual" may suffer from other diseases or conditions unrelated to the condition being measured and may not generally be considered "healthy."

[0042] As used herein, the term "fragment" (e.g., DNA or RNA fragment) may refer to a portion of a polynucleotide or polypeptide sequence comprising at least 3 consecutive nucleotides. Nucleic acid fragments may retain the biological activity and / or some characteristics of the parent polypeptide. Nucleic acid fragments may be double-stranded or single-stranded, methylated or unmethylated, complete or nicked, complexed or uncomplexed with other macromolecules (e.g., lipid particles, proteins). Nucleic acid fragments may be linear fragments or circular fragments. Bacterial nucleic acid may refer to any nucleic acid of bacteria. Such bacterial nucleic acid may be released from a microorganism. As part of a biological sample analysis, a statistically significant number of fragments may be analyzed, for example, at least 1,000 fragments may be analyzed. As other examples, at least 5,000, 10,000, or 50,000, or 100,000, or 500,000, or 1,000,000, or 5,000,000 fragments or more may be analyzed.

[0043] The term "determination" generally refers to a technique for determining the properties of a nucleic acid or nucleic acid sample (e.g., a statistically significant number of nucleic acids or microorganisms of relative abundance) and the properties of an object from which a sample is obtained. Determination may include techniques for determining the amount of nucleic acid in a sample, the genomic identity of nucleic acid in a sample, the copy number variation of nucleic acid in a sample, the methylation state of nucleic acid in a sample, the fragment size distribution of nucleic acid in a sample, the mutation state of nucleic acid in a sample, or the fragmentation pattern of nucleic acid in a sample. Any determination known to those of ordinary skill in the art can be used to detect any properties of nucleic acid mentioned herein. The properties of nucleic acid include sequence, amount, genomic identity, copy number, methylation state at one or more nucleotide positions, size of nucleic acid, nucleic acid mutation at one or more nucleotide positions, and fragmentation pattern of nucleic acid (e.g., nucleotide positions of nucleic acid fragmentation). The term "determination" can be used interchangeably with the term "method". Determinations or methods can have specific sensitivity and / or specificity (e.g., based on the selection of one or more cutoff values), and their relative usefulness as diagnostic tools can be measured using the receiver operating characteristic (ROC) curve area under (AUC) statistics.

[0044] "Sequence read" refers to a string of nucleotides obtained from any part or all of a nucleic acid molecule. For example, a sequence read can be a short nucleotide string (e.g., 20-150 nucleotides) sequenced from a nucleic acid fragment, a short nucleotide string at one or both ends of a nucleic acid fragment, or the sequencing of the entire nucleic acid fragment present in a biological sample. Sequence reads can, for example, be obtained in a variety of ways using sequencing techniques or using probes such as in a hybridization array or capture probes such as can be used in a microarray, or amplification techniques such as polymerase chain reaction (PCR) or linear amplification or isothermal amplification using a single primer. Exemplary sequencing techniques include shotgun metagenomic sequencing, 16S rRNA sequencing, massively parallel sequencing, targeted sequencing, Sanger sequencing, sequencing while connecting, ion semiconductor sequencing, and single molecule sequencing (e.g., using nanopores or single molecule real-time sequencing (e.g., from Pacific Biosciences)). This sequencing can be random sequencing or targeted sequencing (e.g., by using capture probes hybridized to a specific region or by amplifying certain regions, both of which are enriched in such regions). Exemplary PCR techniques include real-time PCR and digital PCR (e.g., droplet digital PCR). As part of the analysis of a biological sample, a statistically significant number of sequence reads can be analyzed, for example, at least 1,000 sequence reads can be analyzed. As other examples, at least 5,000, 10,000, or 50,000, or 100,000, or 500,000, or 1,000,000, or 5,000,000 sequence reads or more can be analyzed.

[0045] As used herein, the term "classification" refers to any number or other character associated with a particular characteristic of a sample. For example, a "+" symbol (or the word "positive") can indicate that the sample has a particular health classification, where a vector of such symbols can be provided, each indicating a classification of a corresponding health condition (e.g., healthy or multiple unhealthy conditions). The classification can be binary (e.g., positive or negative) or have more class levels (e.g., a scale of 1-10 or 0-1), including probabilities.

[0046] The term "parameter" as used herein means a numerical value that characterizes a quantitative data set and / or a numerical relationship between quantitative data sets. For example, the ratio (or a function of the ratio) between a first nucleic acid sequence of a first amount and a second nucleic acid sequence of a second amount is a parameter. Parameters can be used to determine any classification described herein.

[0047] "Relative abundance" can refer to a ratio (e.g., a percentage, a fraction, or a concentration). In particular, the relative abundance of a particular bacterial species can provide the ratio of bacterial DNA fragments from the particular bacterial species (e.g., as determined by aligning sequence reads or via probes specified for the particular bacterial species). The relative abundance of a particular bacterial species can be determined by dividing the raw abundance of the particular species by the total number of species counts per sample.

[0048] The terms "cutoff value" and "threshold value" refer to predetermined numbers used in an operation. For example, a cutoff size may refer to a size beyond which a fragment is excluded. A threshold value may be a value that, when higher or lower than the value, indicates that a particular classification is applicable. Either of these two terms may be used in either of these two contexts. A cutoff value or threshold value may be a "reference value," or may be derived from a reference value representing a particular classification or distinguishing between two or more classifications. A cutoff value may be predetermined with or without reference to the characteristics of a sample or object. For example, a cutoff value may be selected based on the age or sex of a test subject. A cutoff value may be selected after the output of the test data and based on the output. For example, when the sequencing of a sample reaches a certain depth, certain cutoff values ​​may be used. As another example, a reference object having a known classification of one or more conditions and a measured characteristic value (e.g., relative abundance, statistical size value, or count) may be used to determine a reference level to distinguish between different conditions and / or the classification of conditions (e.g., whether the subject suffers from the condition). Reference value can be selected as a value representing a classification (e.g., mean value) or between two clusters of measurement (e.g., selected to obtain desired sensitivity and specificity). As another example, reference value can be determined based on statistical simulation of sample. Any of these terms can be used in any context in these contexts. As the technician will appreciate, such reference value can be determined in various ways. For example, the measurement of two different object queues with different known classifications can be determined, and reference value can be selected as a value representing a classification (e.g., mean value) or between two clusters of measurement (e.g., selected to obtain desired sensitivity and specificity). As another example, reference value can be determined based on statistical simulation of sample. Specific values ​​of cutoff value, threshold value, reference etc. can be determined based on desired accuracy (e.g., sensitivity and specificity).

[0049] "Health classification" refers to a classification of a subject's health status. As an example, a health classification may be provided as whether a subject is healthy or has a condition, such as the conditions mentioned below and herein. A classification may be provided for each of a variety of conditions or whether a subject is healthy. As an example, a classification may be a binary value as to whether a condition exists or a probability value for the existence of a condition.

[0050] "The level of a condition" may refer to whether bacteria are present in a biological sample, or the amount of bacteria present. For example, the level of a condition may indicate the number of sequence reads associated with bacteria obtained from a sample (e.g., a fecal sample) of an object (e.g., a read per million). The presence of bacteria may indicate the amount, degree, or severity of a condition associated with bacteria. In some cases, the amount, degree, or severity of a condition is predicted based on the number of bacteria in a biological sample. The level of a condition may be a number or other mark, such as a symbol, letter, and color. The level may be zero. The level of a condition may be used in various ways. Exemplary conditions include gastrointestinal (GI) diseases, such as colorectal cancer (CRC), colorectal adenocarcinoma (CA), Crohn's disease (CD), ulcerative colitis (UC), and inflammatory bowel disease (IBD); obesity; and cardiovascular disease (CVD).

[0051] The term "true positive" (TP) can refer to a subject having a condition. A true positive generally refers to a subject having a disease (e.g., post-acute COVID-19 syndrome). A true positive generally refers to a subject having a condition and identified as having the condition by an assay or method of the present disclosure.

[0052] The term "true negative" (TN) may refer to a subject who does not have a disease or a detectable condition. A true negative generally refers to a subject who does not have a disease or a detectable disease, including post-acute COVID-19 syndrome. A true negative generally refers to a subject who does not have a disease or a detectable condition, or is identified as not having the condition by an assay or method of the present disclosure.

[0053] The term "false positive" (FP) may refer to a subject that does not have a disease condition. A false positive generally refers to a subject that does not have a disease condition. The term "false positive" generally refers to a subject that does not have a disease condition but is identified as having the condition by an assay or method of the present disclosure.

[0054] The term "false negative" (FN) may refer to a subject having a condition. A false negative generally refers to a subject having a disease. The term "false negative" generally refers to a subject having a condition but identified as not having the condition by an assay or method of the present disclosure.

[0055] The term "sensitivity" or "true positive rate" (TPR) can refer to the number of true positives divided by the sum of the number of true positives and false negatives. Sensitivity can characterize the ability of an assay or method to correctly identify the proportion of a population that actually has a condition. For example, sensitivity can characterize the ability of a method to correctly identify the number of subjects in a population that have a disease. In another example, sensitivity can characterize the ability of a method to correctly identify one or more markers that are indicative of a disease.

[0056] The term "specificity" or "true negative rate" (TNR) can refer to the number of true negatives divided by the sum of the number of true negatives and false positives. Specificity can characterize the ability of an assay or method to correctly identify the proportion of a population that is truly free of a disease. For example, specificity can characterize the ability of a method to correctly identify the number of subjects in a population that are free of a disease. In another example, specificity can characterize the ability of a method to correctly identify one or more markers that are indicative of a disease.

[0057] The term "ROC" or "ROC curve" may refer to a receiver operating characteristic curve. A ROC curve may be a graphical representation of the performance of a binary classifier system. For any given method, a ROC curve may be generated by plotting sensitivity against specificity at various threshold settings. The sensitivity and specificity of the method for predicting the health status level of an object may be determined based on the concentration of microbial DNA in the object's fecal sample by various probabilities generated by a machine learning model. In addition, at least one of the three parameters (e.g., sensitivity, specificity, and threshold settings) is provided, and the ROC curve may determine the value or expected value of any unknown parameter. Unknown parameters may be determined using a curve fitted to the ROC curve. The term "AUC" or "ROC-AUC" generally refers to the area under the receiver operating characteristic curve. The metric may provide a measure of the diagnostic utility of the method, taking into account the sensitivity and specificity of the method. Typically, ROC-AUC ranges from 0.5 to 1.0, where values ​​closer to 0.5 indicate that the method has limited diagnostic utility (e.g., lower sensitivity and / or specificity), and values ​​closer to 1.0 indicate that the method has greater diagnostic utility (e.g., higher sensitivity and / or specificity). See, e.g., Pepe et al., "Limitations of the Odds Ratio in Gauging the Performance of a Diagnostic, Prognostic, or Screening Marker," Am. J. Epidemiol 2004, 159(9):882-890, which is incorporated herein by reference in its entirety. According to Cook, "Use and Misuse of the Receiver Operating Characteristic Curve in Risk Prediction," Circulation 2007, 115:928-935 (which is incorporated herein by reference in its entirety), additional methods for characterizing diagnostic utility using likelihood functions, odds ratios, information theory, predictive values, calibration (including goodness of fit), and reclassification measures are summarized.

[0058] "Negative predictive value" or "NPV" can be calculated by TN / (TN+FN) or the true negative fraction of all negative test results. The negative predictive value is inherently affected by the prevalence of the condition in the population and the pre-test probability of the desired test population. "Positive predictive value" or "PPV" can be calculated by TP / (TP+FP) or the true positive fraction of all positive test results. It is inherently affected by the prevalence of the disease and the pre-test probability of the desired test population. See, e.g., O'Marcaigh AS, Jacobson RM, "Estimating The Predictive Value Of A Diagnostic Test, How ToPrevent Misleading Or Confusing Results," Clin. Ped. 1993, 32(8): 485-491, which is incorporated herein by reference in its entirety.

[0059] A "machine learning model" (ML model) may refer to a software module configured to run on one or more processors to provide a classification or numerical value of a characteristic of one or more samples. An ML model may be generated using sample data (e.g., training data) to make predictions on test data. One example is an unsupervised learning model. Another exemplary type of model is supervised learning that may be used with embodiments of the present disclosure. Exemplary supervised learning models may include different methods and algorithms, including analytical learning, statistical models, artificial neural networks, backpropagation, boosting (meta-algorithm), Bayesian statistics, case-based reasoning, decision tree learning, inductive logic programming, Gaussian process regression, genetic programming, grouping methods for data processing, kernel estimators, learning automata, learning classifier systems, minimum message length (decision trees, decision diagrams, etc.), multilinear subspace learning, naive Bayes classifiers, maximum entropy classifiers, conditional random fields, nearest neighbor algorithms, possibly approximately correct learning (PAC) learning, ripple descent rules, knowledge acquisition methods, symbolic machine learning algorithms, sub-symbolic machine learning algorithms, minimum complexity machines (MCM), random forests, ensembles of classifiers, ordered classification, data preprocessing, handling unbalanced data sets, statistical relational learning, or Proaftn - a multi-criteria classification algorithm. The model may include linear regression, logistic regression, deep recurrent neural networks (e.g., long short-term memory, LSTM), hidden Markov models (HMM), linear discriminant analysis (LDA), k-means clustering, density-based spatial clustering with noisy applications (DBSCAN), random forest algorithms, support vector machines (SVM), or any model described herein. The supervised learning model may be trained in various ways using various cost / loss functions defining the error from known labels (e.g., least squares and absolute differences from known classifications) and various optimization techniques, such as back propagation, steepest descent, conjugate gradients, and Newton and quasi-Newton techniques.

[0060] The term "about" or "approximately" can mean within an acceptable error range of a particular value determined by a person of ordinary skill in the art, which will depend in part on how the value is measured or determined, i.e., the limitations of the measurement system. For example, according to the practice in the art, "about" can mean within 1 or more than 1 standard deviation. Alternatively, "about" can mean up to 20%, up to 10%, up to 5%, or up to 1% of a given value. Alternatively, particularly with respect to biological systems or processes, the term "about" or "approximately" can mean within an order of magnitude of a value, within 5 times, and more preferably within 2 times. Where specific values ​​are described in the application and claims, unless otherwise indicated, the term "about" should be considered to mean within an acceptable error range of a specific value. The term "about" can have a meaning as commonly understood by a person of ordinary skill in the art. The term "about" can mean ±10%. The term "about" can mean ±5%.

[0061] In the case of providing a range of values, it should be understood that, unless the context clearly provides otherwise, each intervening value between the upper and lower limits of the range is also specifically disclosed, to one tenth of the lower limit unit. Each smaller range between any of the values ​​or intervening values ​​in the range and any other values ​​or intervening values ​​in the range is included in the embodiments of the present disclosure. The upper and lower limits of these smaller ranges can be independently included in the range or excluded in the range, and each range in which any one of the limiting values ​​is included in the smaller range, both limiting values ​​are not included in the smaller range, or both limiting values ​​are included in the smaller range is also included in the present disclosure, subject to any specific exclusion of the limitation in the range. In the case where the range includes one or two limiting values, the scope excluding any one or two of the limiting values ​​included is also included in the present disclosure.

[0062] Standard abbreviations may be used, eg, bp, base pair; kb, kilobase; pi, picoliter; s or sec, second; min, minute; h or hr, hour; aa, amino acid; nt, nucleotide; and the like.

[0063] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs. Although any methods and materials similar or equivalent to those described herein can be used in the practice or testing of embodiments of the present disclosure, some potential and exemplary methods and materials can now be described.

[0064] Details

[0065] The present disclosure provides technology for personalized predictive risk assessment tools that use microbiome to determine the risk of multiple diseases in a subject. Current risk prediction tests using microbiome can only detect one disease or health condition at a time. By determining multiple diseases simultaneously, the disclosed technology can provide a cost-effective method to support clinical decision making and thus help improve disease prevention and management.

[0066] Recent studies have shown that an unbalanced gut microbiota, known as “dysbiosis,” contributes to a variety of human diseases [1], but the potential role of dysbiosis in disease prediction has remained largely unexplored due to limited sample sizes, insufficient validation, and extensive heterogeneity between studies [2,3]. Furthermore, current development of microbiome biomarkers has primarily used binary classifiers (disease presence or absence) and focused on gastrointestinal (GI) diseases such as colorectal cancer (CRC) and inflammatory bowel disease (IBD) [4-6].

[0067] However, emerging evidence suggests that the gut microbiome is involved in the pathogenesis of non-GI diseases, including obesity and cardiovascular disease (CVD) [7,8]. Recently, coronavirus disease 2019 (COVID-19) and its long-term sequelae, known as post-acute COVID-19 syndrome (PACS), have also been associated with significant and persistent gut dysbiosis [9,10,11]. Given that most health status phenotypes (i.e., corresponding to health categories) exhibit overlapping gut microbiome signatures

[12] , single disease diagnostic models (single health status phenotype versus healthy controls) may be confounded by signals shared between unrelated health status phenotypes and may lead to misclassification.

[0068] The present disclosure provides a technique for simultaneously predicting the phenotypic risk of multiple health conditions in an object using a multi-class machine learning model. Health conditions may include post-acute COVID-19 syndrome (PACS or long COVID), Crohn's disease (CD), colorectal cancer (CRC), colorectal adenocarcinoma (CA), obesity (Ob), irritable bowel syndrome (IBS), ulcerative colitis (UC), and cardiovascular disease (CVD). Metagenomics data, such as fecal microbiome datasets of multiple human health conditions, can be used to train machine learning multi-class models. Existing risk prediction models using metagenomics data can distinguish diseases in two categories. However, using a two-part technique, even when reference data for multiple diseases are available, it may be necessary to determine the risk of each disease one at a time. The technology in the present disclosure provides a method for simultaneously determining the risk of multiple diseases.

[0069] Metagenomics can include the study of genetic material from communities of organisms. Communities can be found in samples including stool samples, intestinal mucosal samples, saliva, skin swabs, soil samples, water samples, etc. For example, species characteristics of stool samples can be used to identify health status phenotypes. Various techniques, such as shotgun metagenomic sequencing, can be used to characterize the species of a sample. In shotgun metagenomic sequencing, DNA is obtained from a heterogeneous cell sample and is fragmented into DNA fragments that can be aligned to the genomes of multiple microbial species to identify the species in the sample.

[0070] The technology in the present disclosure includes a machine learning model trained using a data set derived from a cohort of 2,320 individuals with 9 well-characterized health status phenotypes. Fecal samples from the cohort provide DNA for characterizing the microbial species in the data set. Sequencing produced 14.6 terabytes of deoxyribonucleic acid (DNA) sequencing for the cohort of 2,320 individuals. Health status phenotypes (e.g., health status) include CRC, CA, CD, Ob, UC, IBS, PCAS, and healthy control phenotypes. Fecal samples from the cohort are used to generate a processed data set that contains microbial species present in more than 5% of the individuals in the cohort. The processed data set can be used for training data for a machine learning model that can use the fecal samples after sequencing to classify the individual based on the health status phenotype.

[0071] A classifier can be a model that outputs the probability of a given entity being in a class (an example of a healthy classification). The probability can be a value from 0.0-1.0, and a value called a discrimination threshold can be the dividing line between classes. For example, if the discrimination threshold is 0.6, then entities with probabilities of 0.0-0.6 can be classified into class A, and if the probability is 0.6-1.0, then the entity can be sorted into class B.

[0072] The performance of a classification model (e.g., a classifier) ​​can be evaluated using a graph called a receiver operating characteristic (ROC) curve. The graph shows the ability of a classifier to classify entities into categories as the discrimination threshold is varied. The y-axis of the graph is the true positive rate or sensitivity, and the x-axis is the false positive rate (e.g., 1 – sensitivity). The area under the ROC (AUC) is a measure of the performance of a classification model. The AUC value indicates the separability of the classifier, and the value can range from 0.0-1.0, with higher values ​​indicating higher separability. For example, a classifier with an AUC of 1.0 is able to classify all entities into the correct category.

[0073] Using the techniques in the present disclosure, a trained machine learning multi-class classifier achieved a receiver operating characteristic curve area (AUC) of 0.90-0.99 (interquartile range, IQR, 0.91-0.94) for health status phenotype prediction, with high sensitivity (IQR, 0.87-0.93) and specificity (IQR 0.83-0.95). The classifier remained predictive in a cross-regional public dataset (mean AUC 0.82, IQR 0.79-0.87, Fig. 12B We associated the top 50 microbial species contributing to the model and disease phenotypes, thereby identifying health status phenotype-specific microbial signatures.

[0074] I. Multi-class machine learning models

[0075] Multi-class classification is the problem of classifying entities into multiple categories. An algorithm can be trained to perform multi-class classification using training data with known classifications. During training, training data is input into the algorithm and the algorithm's parameters are modified until the classification output by the algorithm matches the known classifications.

[0076] A. Model Selection

[0077] Classification tasks involving more than two categories in machine learning can be called "multi-class classification", which can effectively account for the confounding effects of co-existing categories

[16] . Due to the complexity of the gut biome, multi-class classification is more suitable for developing microbiome-based diagnostic tools than binary classifiers (single disease versus control). For example, based on our cohort of 2,320 individuals representing 9 health conditions (including 8 diseases and 1 healthy group), we trained 5 machine learning multi-class classifiers to classify different diseases using standardized data at the species level (325 species). The dataset was divided into training and test sets. The training set was used to train the model, and the performance of the trained model was evaluated using the retained test set.

[0078] Multi-class classification can be performed by a machine learning model, which can be trained by an algorithm. Training data can be input into the algorithm as a feature vector, and the algorithm can output a classification based on the input. A feature vector can be an n-dimensional vector containing numerical features of an entity. The training data can have a known classification, and during training, the parameters of the algorithm can be modified until the output classification matches the known classification. For example, the training data can include the microbiome of stool samples from cohort members, and the classification can be a phenotype (e.g., whether a health condition is present).

[0079] Multi-class classification models can be trained by several algorithms. For example, K-nearest neighbors is an algorithm that can maintain existing instances and classify new instances using a similarity measure (e.g., a distance function). Support vector machines, or SVMs, are a common supervised learning technique, and the algorithms used to create SVMs can be used to solve classification and regression problems. The SVM algorithm classifies data points by finding an optimal line or decision boundary in an n-dimensional space. The decision boundary divides the data points into categories so that additional data points can be easily placed in the appropriate category in the future. A hyperplane can be the name of the optimal choice boundary. Random forest is a supervised learning method used for classification and regression in machine learning. Random forest is a supervised machine learning algorithm that averages the results of many decision trees applied to different subsets of a data set to improve the projection accuracy of the data set. MLP classifier stands for multilayer perceptron classifier, which is a feedforward artificial neural network. MLP classifiers consist of nodes connected to nodes arranged in layers by edges. Like the SVM algorithm, the MLP algorithm can classify data by finding an optimized decision boundary in an n-dimensional space. Graph Convolutional Neural Networks (GCNs) are a method for semi-supervised learning on graph-structured data. GCNs are based on a variation of convolutional neural networks that can operate directly on graphs. The choice of convolutional structure can be motivated by a local first-order approximation of spectral graph convolutions. The number of graph edges can be scaled linearly, and the model can learn hidden layer representations that can encode local graph structure and node features.

[0080] B. Model training and validation

[0081] Figure 1 It is a framework diagram for data set partitioning, model training and independent validation according to an embodiment. The machine learning model can be trained using a metagenomics data set derived from an object with one or more known health conditions. The health condition may include healthy controls that have not yet been diagnosed with a known condition.

[0082] Metagenomics data set can be a collection of metagenomics data, which can obtain metagenomics data by sequencing samples containing genetic material from an organism community. For example, a metagenomics data set can be a group of fecal samples collected or downloaded from a public database. Metagenomics data sets can include sequencing data generated by the same standardized scheme, which includes steps from fecal DNA extraction to sequencing, raw data quality filtering, host read purification to microbiome interpretation. The metagenomics data of individuals whose samples are tested (e.g., classified) by training models can use the same standardized scheme for generating model training data.

[0083] The health status of the subject, whose sample is included in the training data set of the multi-class classifier, can be determined by a formal diagnosis specific to the condition. For example, (1) PACS is defined as at least one persistent symptom or long-term complication of SARS-CoV-2 infection for more than 4 weeks from the onset of symptoms (appearing only after COVID-19), which cannot be explained by alternative diagnoses; (2) CVD is confirmed by carotid ultrasound examination of cardiovascular stenosis and plaques; (3) IBS subjects are diagnosed according to the ROMEⅢ criteria. Colonoscopy can be performed to rule out any GI disorders that manifest as changes in bowel habits, such as inflammatory bowel disease, celiac disease, parasitic infections, or other organic disorders; (4) CD is diagnosed by endoscopy, radiology, and histology; (5) UC is defined by endoscopy, radiology, and histology; (6) CRC and CA are diagnosed by colonoscopy and histology. (7) Obesity is defined by a BMI of more than 28, and the recruited subjects do not suffer from other diseases.

[0084] Subjects with cardiovascular disease (CVD) were recruited from the general public as part of the Hong Kong General Population Cardiovascular Health Survey. Subjects underwent carotid ultrasound to measure the intima-media thickness (IMT) of the common carotid artery, internal carotid artery, external carotid artery (CCA, ICA, and ECA, respectively), and carotid bulb, and subjects with single or multiple vessel stenosis ≥50% were considered to be at risk for CVD.

[0085] Individuals diagnosed with each health condition can provide samples that can be sequenced to generate metagenomic data sets. For example, stool samples or intestinal mucosal samples can be sequenced to determine the relative abundance of all species in the microbiome. Intestinal mucosal samples can be obtained by gastrointestinal biopsy, sampling using lumen brushes, intestinal catheter aspiration, gastrointestinal surgical resection, etc. Although in some cases, not all species need to be used. The number of individuals can be 10, 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1250, 1500, 1750, 2000, 2500, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10000, 15000, 20000, 50000, 100000, 500000, or 1000000 individuals. A feature vector can be created from a metagenomics dataset to generate a training dataset. The feature vector can be an N-dimensional vector, where N is the number of bacterial species. N can be the number of bacterial species detected in the metagenomics dataset, the top 350 bacterial species in the dataset, the top 50 bacterial species in the dataset, or all species with a prevalence greater than 0.15% in the dataset.

[0086] By randomly dividing the feature vector into training data and test data, a nested cross validation procedure can be applied to calculate the accuracy within the cohort. The training data can be randomly divided into k subsets. The subsets can be used to create k folds, where k-1 subsets are used as training sets and the remaining subsets are used as in-training validation sets. The training set can be used to train the model, and the in-training validation set can be used to evaluate the performance of the model during training. Test data that may not be used to train the model can be used to evaluate the trained model.

[0087] For example, the validation procedure can be 20 repetitions of five-fold stratified cross-validation (balancing the class ratio between folds), where k=5. For each split in this example, an L1 regularized (Lasso) logistic regression model can be trained on the training set, which can then be used to predict the classification of the validation set. After the process is repeated 20 times, the model can be evaluated on the test data. The λ hyperparameter can be selected for each model to maximize the AUC-ROC under the constraint that the model contains at least five non-zero coefficients. The λ hyperparameter of the model can be changed until the maximum AUC-ROC for the model is found.

[0088] The area under the receiver operating characteristic curve (AUC) can be used to compare the performance of models with different methods and features. AUC is a widely used metric that considers the trade-off between sensitivity and specificity at all possible thresholds to compare the performance of various classifiers. For a random classifier, a baseline AUC value can be 0.5.

[0089] The area under the precision-recall curve (AUPR) can be provided as a supplementary assessment that considers the trade-off between precision (or positive predictive value) and recall (or sensitivity) over a baseline equal to the proportion of positive disease cases in all samples. Precision is the fraction of relevant instances among the retrieved instances, and recall is the total proportion of relevant instances retrieved. A precision-recall curve is a plot of precision (x-axis) against recall (y-axis). An AUPR value of 1.0 indicates high precision and recall, while a value of 0.0 indicates low precision and recall.

[0090] C. Generate multi-class machine learning models

[0091] Figure 2A is a flowchart showing a summary of generating a multi-class machine learning model according to an embodiment. In some cases, the generation of a multi-class machine learning model is an ongoing process that includes continuous updating, verification, and improvement. The model can be updated and verified as new data is collected. To generate a multi-class machine learning model, the following steps can be performed:

[0092] At block 205, a collection of metagenomic data may be collected from subjects with known health conditions. Health condition phenotypes may include disease and healthy control phenotypes.

[0093] At block 210, the composition of the microbiome of any one or more subjects may be characterized by determining the relative abundance of all species present in those subjects, for example, by performing shotgun metagenomic sequencing of a stool sample.

[0094] At block 215 , the data set is split into a training set (eg, 70%) and a holdout test set (eg, 30%).

[0095] At box 220, a plurality of multi-class machine learning models can be created by the algorithm using the training set data. The algorithm can be trained into a model using a cross validation (CV) method, and after training, the best model can be selected. The random forest algorithm can be trained to produce a random forest classifier. Other algorithms can be trained to create machine learning models, including a multilayer perceptron (MLP) algorithm, a support vector machine (SVM) algorithm, a K nearest neighbor (KNN) algorithm, and a graph convolutional neural network (GCN) algorithm. For example, by dividing the training set into a ratio of 8:2, the CV method can be used to create 100 models with 20 repetitions and 5-fold training. 80% of the training set can be used for training, and 20% for validation within the training set. Then, the model with the highest AUC is selected from these 100 models.

[0096] The k-fold hyperparameter optimization procedure can be nested in the k-fold cross validation of the algorithm selection to further improve the model performance (e.g., as evaluated by the AUC of the model) while reducing the overfitting problem. In the inner loop, the model can be adjusted by changing the hyperparameters (e.g., if a random forest is used, the hyperparameter n_estimators is adjusted, and, for example, the parameter can be set to n_estimators = 2000).

[0097] At block 225 , the model may be evaluated using data from the test set to obtain the model's AUC (eg, the probability that the model can correctly distinguish between randomly selected correctly classified samples and incorrectly classified samples).

[0098] At block 230, if the AUC is not satisfactory, the model may be adjusted by increasing the number of objects in the training set. If the value is below a threshold, the AUC may be unsatisfactory.

[0099] Blocks 205-230 may be repeated repeatedly until the AUC from block 225 is satisfactory and a final model may be obtained. Whether the AUC is acceptable may vary depending on the health condition being tested. A lower AUC score may indicate a higher confidence level, and a higher AUC score may indicate a lower confidence level. An acceptable threshold may be an AUC value greater than or equal to 0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0.

[0100] At block 235, ad hoc thresholds for the final model for each health phenotype can be calculated based on the Youden Index. The Youden Index can integrate sensitivity and specificity information, and by using Youden Index analysis, we can find the optimal cutoff value, such as the value that provides the best trade-off between sensitivity and specificity. Figure 5B The Youden index is discussed in more detail in the description of .

[0101] Figure 2B is a flow chart showing a technique for applying a model to predict a predetermined health condition risk in a subject according to an embodiment. To determine whether an individual is at risk for a predefined health condition, the following steps may be performed:

[0102] At block 240, metagenomic data is collected from the individual subject. The metagenomic data may be determined by shotgun metagenomic sequencing of a stool sample.

[0103] At block 245, the relative abundance of all species in the subject is determined.

[0104] At box 250, if a random forest classifier is generated at box 220, a decision tree will be generated from the training data through the random forest. The relative abundance and values ​​of other clinical factors will be run along the decision tree, and a risk score (probability) can be generated for each predefined health condition. If the risk score is above the threshold calculated at box 235, it can be determined that the test subject is at risk for the predefined health condition.

[0105] II. Exemplary Results: Fecal Microbiome-Based Machine Learning for Multi-Class Disease Diagnosis

[0106] Figure 3A is a table showing a demographic summary of individuals who provided stool samples to one embodiment. We performed a cohort of 2,320 Hong Kong, China individuals (mean age 54.9 years, 48.7% female, Figure 3A) fecal samples were sequenced at an average depth of 6.45 gigabases from a cohort of 9 well-characterized disease phenotypes: CRC (n=174), colorectal adenocarcinoma (CA, n=168), Crohn's disease (CD, n=200), ulcerative colitis (UC, n=147), irritable bowel syndrome (IBS, subtype D, n=145), obesity (n=148), CVD (atherosclerosis, n=143), PACS (n=302), and healthy controls (n=893). A total of 14.6 terabytes of sequences were obtained, and a total of 1,208 bacterial species were detected that were present in at least one metagenomic sample. Of these, 325 bacterial species had a relative abundance higher than 0.15%, and these species were present in more than 5% of the subjects. This sample size and sequencing depth allowed us to capture most microbial signatures by bootstrap analysis, which covered more than 95% of the total expected microbial taxa at the species level (339.6 out of 325, standard error = 2.9). Individuals with CA or CRC were older than subjects in other phenotypes (p < 0.05), but no gender differences were found between different phenotypes.

[0107] Figure 3B is a graph showing alpha diversity of bacteria in a cohort according to an embodiment. Diversity can be the range of different kinds of species in a community, such as single-celled organisms, bacteria, archaea, protists, and fungi. Alpha diversity can be a measure of the number of species in a community and the relative abundance of each species in the community. In this case, the community can be the individuals in the cohort, and the species are bacterial species. Figure 3A The diversity calculated using the Shannon index is shown, which is defined in the following equation:

[0108] H′=-∑[(p i )*ln(p i )]

[0109] Where H' is the Shannon index value, and p i is the proportion of the total population within the community represented by species i. P values ​​were calculated by comparing healthy controls to other health conditions using MaAsLin 2 after adjusting for age and sex.

[0110] Figure 3C is a graph showing the richness of fecal biome bacteria in a cohort according to an embodiment. Richness can be the number of species observed in a community. In this case, the species can be the number of bacterial species, and the community can be the individuals in the cohort. P values ​​were calculated by comparing healthy controls to other health conditions using MaAsLin 2 after adjusting for age and sex.

[0111] Deep shotgun metagenomic sequencing provided the opportunity to explore the common set of microbial species in our cohort (common core, 325 species). We observed different changes in bacterial diversity (Shannon index) and richness (number of bacterial species observed) in different disease-versus-health comparisons. For example, bacterial diversity was significantly higher in stool samples from colorectal adenomas compared with healthy controls (p < 0.05), but lower in stool samples from CD (p < 0.0001). These results are consistent with meta-analyses

[13] , suggesting that ecological indices (e.g., Shannon diversity and richness) may not be robust predictors of health or disease.

[0112] Figure 3D is a diagram showing the association of fecal biome bacterial species with health status phenotypes according to the implementation scheme. The top 50 microbial species with the highest number of associations are shown. The associations are colored by the effect direction (red, positive; blue, negative; p<0.05), and significant associations at FDR<0.05 are marked with "+" (positive correlation) or "-" (negative correlation). After adjusting for age and sex, we used the linear model of MaAsLin2 to explore the association of microbial composition at the species level with the above-mentioned 9 diseases or health phenotypes. We found that at the species level, there were 1,061 significant associations (FDR<0.05) between these 9 phenotypes and 215 bacterial taxa. Different disease phenotypes have different numbers of associations with different bacterial species. Compared with CVD (69 associations), CD showed the strongest association (defined by the total number of bacterial phenotype associations; 143 associations), followed by PACS (140 associations) and CRC (125 associations).

[0113] Of the 215 disease-associated bacterial species identified, more than 94% (203 / 215) were significantly associated with two or more diseases, and approximately 74% (159 / 215) were associated with four or more diseases. For example, in our cohort, Klebsiella pneumoniae, a well-characterized opportunistic pathogen

[14] , was positively associated with CD, CRC, IBS-D, obesity, PACS, and UC, whereas Roseburia intestinalis, a potential probiotic with butyrate-producing properties

[15] , was negatively associated with these six health status phenotypes (Figure 2D). Consistent with previous findings [12,13], our data demonstrated shared microbiome associations between different diseases, suggesting that single disease diagnostic models may be confounded by unrelated diseases

[12] .

[0114] Figures 4A-4C Differences in fecal microbiome between health conditions according to an embodiment are shown. Figure 4Ais a plot showing the principal coordinates analysis (PCoA) of the beta diversity of the fecal biome. Beta diversity can be a measure of the similarity or dissimilarity of different communities. Beta diversity can be the ratio between the diversity of individuals in a cohort and the total diversity of that cohort. PcoA is a type of multidimensional scaling that creates entity maps based on distance. Figure 4B is a graph showing the results of a permutational analysis of variance (PERMANOVA) test of fecal biome beta diversity for one health condition versus another. PERMANOVA is a nonparametric alternative to multivariate analysis of variance (MANOVA), and PERMANOVA compares entity groups to the null hypothesis that there are no differences between entities.

[0115] Figure 4C Figure 2 is a graph showing the area under the receiver operating characteristic curve (AUC) of a random forest (RF) binary classifier for one-to-one discrimination for multiple health conditions. We measured differences in microbiome composition between phenotypes using a permutation test based on beta diversity and machine learning binary classification (random forest, RF). Our data showed that PcoA analysis based on beta diversity separated different disease phenotypes ( Figures 4A-4B ), and these differences in microbiome composition between different phenotypes were also reflected in the area under the receiver operating characteristic curve (AUC) of the random forest (RF) binary classifier (one-to-one, mean 0.87, interquartile range (IQR) 0.83-0.91, Figure 4C ). Overall, despite the presence of many shared microbial markers between diseases, these findings suggest that microbial composition still varies by disease state.

[0116] Figure 5A A graph of a receiver operating characteristic (ROC) curve of a random forest (RF) multi-class classifier according to an embodiment is shown. The figure also lists the AUCs of 9 health conditions evaluated by the RF multi-class classifier using data from sequenced fecal samples. For different disease phenotypes, our RF multi-class classifier achieved an average AUC of 0.90-0.99 (interquartile range, IQR 0.91-0.94, one relative to all others).

[0117] Figure 5B A graph showing the threshold of a random forest (RF) multi-class classifier at the highest Youden Index according to an embodiment. The Youden Index is a statistical value that combines specificity and sensitivity in a single value. The Youden Index can be calculated using the following formula:

[0118] Y = Sensitivity + Specificity - 1

[0119] For different health conditions, Figure 5BThe specificity of the classifiers shown in the range is 0.758-0.982 (IQR 0.83-0.95, one versus all others), highlighting the good diagnostic performance using this multi-class classifier. For example, for CRC, our classifier achieved an average AUC of 0.94, a sensitivity of 0.877, and a specificity of 0.846 (one versus all others, Figure 5A-5B ); this performance exceeds that of our own trained binary classifier (CRC is relatively healthy, with an average AUC of 0.91, Figure 4C ) and the performance of a previously published CRC diagnostic model [4]. One versus all the others could mean that a classifier is trained for each health condition, and the output of the health condition classifier is either the probability that the sample came from someone with that health condition, or the probability that it came from someone with any other health condition.

[0120] Figures 6A-6D A chart of the AUC and AUPR of a model trained to classify health conditions according to an embodiment is shown. The figure lists 9 health conditions and 4 models, including a support vector multi-class classifier (SVC), a k-nearest neighbor (KNN) multi-class classifier, a random forest (RF) multi-class classifier, a multi-layer perceptron (MLP) multi-class classifier, and a graph convolutional neural network (GCN) multi-class classifier. The RF multi-class classifier performs well, with an average area under the precision-recall curve (AUPR) of 0.40-0.93 (IQR 0.53-0.86, FIG. 6), which is generally better than all other models ( FIG. 6 ). Under a threshold based on the highest Youden index, the sensitivity range of our RF multi-class classifier is 0.807-0.946 (IQR 0.87-0.93).

[0121] To fully characterize the RF multi-class model, we compared its performance at different split ratios and obtained similar results, indicating high stability and good predictive power without the risk of overfitting ( Figure 6C ). These models achieved an average AUROC (e.g., AUC) of 0.67-0.99 (interquartile range, IQR 0.81-0.92), indicating that multi-class disease classification based on fecal microbiome is feasible. Among them, the RF multi-class model achieved an average AUROC of 0.90-0.99 (IQR 0.91-0.94, one relative to all others) for different disease phenotypes in the test set. The performance of RF models in the test set was generally better than that of all other models ( Fig.6D ), and the performance is similar to that of the training set (calculated by 5-fold cross validation, Figure 6B ), indicating that the classifier has high integrity.

[0122] Figures 7A-7I A graph showing probability distributions generated by a trained random forest (RF) multi-class classifier according to an embodiment is shown. Probabilities of multiple health conditions during independent testing were calculated. In independent testing, the multi-class classifier evaluated a data set containing a single health condition. Probabilities were calculated for the health conditions in the test set. Fig. 7A shows the probability of colorectal adenoma (CA); 7B shows the probability of Crohn's disease (CD); 7C shows the probability of colorectal cancer (CRC); 7D shows the probability of cardiovascular disease (CVD); 7E shows the probability of health conditions; 7F shows the probability of irritable bowel syndrome with diarrhea (IBS-D); 7G shows the probability of obesity; 7H shows the probability of post-acute COVID-19 syndrome (PACS); and 7I shows the probability of ulcerative colitis (UC).

[0123] Figures 8A-8B The performance of a random forest (RF) multi-class classifier according to an embodiment in one-to-one differentiation of multiple health conditions is shown. Fig. 8A The area under the receiver operating characteristic curve (AUC) and the highest Youden index of the random forest multiclass classifier using species-level fecal microbiome data are shown. Figure 8B The sensitivity and specificity of the highest Youden index selection based on one-to-one discrimination of multiple phenotypes are shown. The performance was evaluated using the predicted probability in the test set ( Figures 7A-7I ), and further evaluation showed that the trained classifier achieved an average AUC of 0.94 for all one-to-one classifications (IQR 0.92-0.98, Fig. 8A ), with high sensitivity (IQR 0.88-0.95) and specificity (IQR 0.83-0.94, Figure 8B ), which supports multi-class model analysis performance better than binary models ( Figure 5B ).

[0124] Figures 9A-9B is a graph showing the performance of a random forest (RF) multi-class classifier stratified by age. Fig. 9A shows the performance of the RF multi-class classifier for colorectal adenoma (CA), and Fig. 9B The performance of the classifier for colorectal cancer is shown. Considering that the subjects with CRC or CA are older ( Figure 3A ), we evaluated the performance of our model stratified by age and found consistent results.

[0125] Figures 10A-10D The performance of a random forest (RF) multi-class classifier trained and tested on a balanced cohort size according to an embodiment is shown. Fig. 10A The area under the receiver operating characteristic curve (AUC) is shown. Fig. 10B Details of model performance metrics of the random forest multi-class classifier for diagnosing one phenotype relative to all other phenotypes using species-level fecal microbiome data are shown. Fig. 10C The AUC and highest Youden index of the random forest multi-class classifier for one-to-one discrimination of multiple phenotypes are shown.

[0126] Fig. 10D The sensitivity and specificity of the classifier selected based on the highest Youden index for one-to-one discrimination of multiple phenotypes are shown. To examine whether uneven cohort sizes among the nine phenotypes affected classification performance, we trained the RF multiclass classifier by randomly pooling 143 subjects from each disease phenotype (a total of 1,287 subjects, 70% training, 30% testing) following the same procedure. Figure 5A ). We observed that the performance of the classifier was (AUC IQR 0.91-0.94, one versus all others; IQR 0.89-0.96, one versus all; Figures 10A-10D ) versus a non-subset cohort of 2,320 individuals (AUCIQR 0.89-0.96, one versus all others; IQR 0.92-0.98; Figure 5B , Figures 8A-8B ), indicating that the impact of unbalanced sample size on performance is negligible.

[0127] Figures 11A-11B The performance of the Random Forest (RF) multi-class classifier using different numbers of features is shown. Fig.11A Shown are the areas under the receiver operating characteristic curves (AUCs) of the random forest multi-class classifier for diagnosing one phenotype versus all other phenotypes using the top 10, 50, 100, and all 325 features in a balanced sample size cohort (n=1287). Fig. 11B The area under the receiver operating characteristic curve (AUC) of the random forest multi-class classifier for diagnosing one phenotype relative to all other phenotypes using the top 10, 50, 100, and all 325 features in the full cohort (n = 2320) is shown. The AUC values ​​of the models increased with the number of features, which ruled out the possibility of overfitting based on 325 features ( Figures 11A-11B ).

[0128] Figures 12A-12CAn independent validation of the performance of a random forest (RF) multi-class classifier according to an embodiment is shown. 12A shows a summary of the health status, sample size, and country of origin of the publicly available cross-regional datasets used in the independent validation. 12B shows the area under the receiver operating characteristic curve (AUC) of the trained RF multi-class classifier for diagnosing one phenotype relative to all other phenotypes in the cross-regional dataset. 12C shows the probabilities generated by the trained random forest multi-class classifier for 60 subjects post-COVID-19 but without acute post-COVID-19 syndrome. The independent validation was determined by integrating 1,597 shotgun fecal metagenomic data from 12 published studies covering 11 countries in Asia, Europe, and North America ( Fig. 12A ). The RF multi-class classifier showed an average AUC of 0.82 (IQR 0.79-0.87) for classifying different diseases. Fig. 12B ).

[0129] Specifically, the AUCs for diagnosing CRC, IBS-D, and CVD were 0.82, 0.86, and 0.88, respectively. This performance from independent cross-regional validation further confirms the robustness and generalizability of our model across populations and geographies. In addition, to further validate the accuracy of the RF model, we selected 60 patients who had fully recovered from COVID-19, as it has been shown that the gut microbiome of COVID-19 survivors without long-term symptoms is similar to that of controls

[10] . We found that the trained model had a specificity of 83.3% (50 / 60) in diagnosing these subjects as healthy ( Fig. 12C ), which validated that fully recovered COVID-19 survivors have similar gut microbiome characteristics to healthy people.

[0130] We integrated 1,597 shotgun feces metagenomic data from 12 public datasets from Asia, Europe, and North America ( Fig. 12A Our RF multi-class classifier showed an average AUROC of 0.69-0.91 (IQR 0.79-0.87) when classifying different diseases. Fig. 12B ). This performance from an independent validation cohort demonstrates the robustness and generalizability of the model across different populations and geographic locations. To further validate the accuracy of the model, we selected 60 patients who had fully recovered from COVID-19 infection. Our trained model showed 83.3% (50 / 60) accuracy in classifying these subjects as healthy ( Fig. 12C). These data confirm that fully recovered COVID-19 survivors (and without PACS) share similar gut microbiome features with healthy individuals. In addition, we also tested our RF model on unrelated cirrhosis and constipation-type IBS datasets (n=60). The results showed that 12 subjects were predicted to be healthy (n=1), CD (n=3), CRC (n=3), PACS (n=3), and UC (n=2), while the other 48 subjects were not predicted because their highest prediction probability failed to exceed the corresponding threshold ( Fig.12D ). These results indicate that our model has high specificity and accuracy, with a low probability of misclassifying unrelated diseases into the nine phenotypes analyzed in this study.

[0131] Figures 13A-13C The performance of a random forest (RF) multi-class classifier using the first 50 features according to an embodiment is shown. Fig.13A Shown is the area under the receiver operating characteristic curve (AUC) of the RF multi-class classifier (using the top 50 features) for diagnosing one phenotype versus all other phenotypes in an independent test set. Fig. 13B Validation of the performance of the RF multi-class classifier (using the top 50 features) in a publicly available dataset is shown. Considering the strong discrimination of the microbial features employed in the multi-class classifier, we associated the top model contributors (top 50 bacterial species contributing to the model, sum of importance: 0.41) with different phenotypes. These top 50 bacterial species achieved about 98.6% of the performance of the full model using 325 features (comparing the average AUC values ​​across phenotypes, Fig.13A Compared with healthy controls, almost all disease states were associated with a significant decrease in the abundance of microbiota from the Firmicutes or Actinobacteria phyla (FDR < 0.05) and a significant increase in the Bacteroidetes phylum (FDR < 0.05).

[0132] The top 50 bacterial species contributing to the model were associated with different disease phenotypes to identify clues to model interpretability. In our test set, these top 50 bacterial species achieved an average AUROC of 0.88-0.99 for different diseases (IQR 0.90-0.93, Fig.13A ), and achieved an average AUROC of 0.67-0.90 (IQR 0.78-0.86, Fig. 13B). A total of 363 significant associations were found between these 50 species and different disease phenotypes. Compared with healthy controls, almost all disease states were associated with a decrease in the abundance of microbiota bacteria from the Firmicutes or Actinobacteria phyla (FDR<0.05) and an increase in the Bacteroidetes phylum (FDR<0.05). An imbalance in the Firmicutes / Bacteroidetes ratio was previously reported to occur primarily in patients with obesity and IBD, but its association with other diseases has not been reported. Nevertheless, this shared microbial signature can be used as a basis for distinguishing between health and disease. We then identified microbial signatures that can classify different diseases. Specifically, the abundance of several bacterial species in the Bacteroidetes phylum was significantly different between patients with PACS, UC, and CD. Compared with healthy controls, subjects with PACS showed a significant increase in the abundance of Bacteroides vulgatus and Bacteroides xylanisolvens, whereas those with UC were enriched in Bacteroides ovatus, and subjects with CD showed a significant decrease in Bacteroides uniformis, Bacteroides vulgatus, and Bacteroides xylanisolvens.

[0133] Although patients with CRC and colorectal adenomas share a relatively similar gut bacterial composition, the abundance of Parvimonas micra was significantly higher in patients with CRC but not in patients with colorectal adenomas compared to healthy controls, which is consistent with previous findings showing that Parvimonas micra can be used as a marker to distinguish CRC from colorectal adenocarcinoma. For other diseases, differences in the microbiome were primarily driven by Actinomyces. Subjects with obesity showed an increase in Actinomyces naeslundii, Actinomycesodontolyticus, and Actinomyces oris, and subjects with IBS-D showed an increase in Collinsella aerofaciens and Collinsella stercoris. We further associated bacteria and phenotypes in compiled public datasets and found that many disease-specific biomarkers were stable across datasets, such as Bacteroidetes for UC, Micromonas parvus for CRC, and Actinobacteria for obesity ( Fig. 13C ). Overall, these results suggest that our model can capture various disease-specific microbial signatures, which may explain the robust diagnostic performance of this multi-class classifier.

[0134] Overall, our data show that fecal microbiome-based multi-class models for disease diagnosis are feasible. The novelty lies in the high-quality dataset, superior and reproducible machine learning methods that are highly clinically relevant, and our classifier can be used as a rapid screening method for various diseases in daily medical examinations or disease risk assessment. Our results also have implications for the potential development of biomarkers to predict drug response and common treatment strategies using identified common or specific markers for multiple diseases. This work has some limitations. Since our model predicts the probability of multiple diseases simultaneously, it can also be applied to the diagnosis of multiple diseases in a single patient.

[0135] Fig.14 It is a chart showing microbial species associated with health states or different health conditions according to the implementation scheme. The top 50 microbial species with the highest contribution to the random forest multi-class classifier are clustered by taxonomy, and different phenotypes are clustered using hierarchical clustering. Associations are colored by effect direction (red, positive; blue, negative; p<0.05), and significant associations at FDR<0.05 are marked with "+" (positive correlation) or "-" (negative correlation). CA, colorectal adenoma; CD, Crohn's disease; CRC, colorectal cancer; CVD, cardiovascular disease; IBS-D, diarrhea-predominant irritable bowel syndrome; PACS, acute post-COVID-19 syndrome; UC, ulcerative colitis.

[0136] An imbalance in the Firmicutes / Bacteroidetes ratio has been previously reported primarily in patients with obesity and IBD

[17] , but its association with other diseases remains limited. Nevertheless, these shared microbial signatures may serve as a basis for distinguishing health from disease ( Fig.14 ). In addition, we observed microbial signatures that could classify distinct disease phenotypes ( Fig.14). The abundance of several bacterial species in the Bacteroidetes phylum differed significantly between patients with PACS, UC, and CD. Compared with healthy controls, subjects with PACS showed a significant increase in Bacteroides vulgaris and Bacteroides xylanisolvens, while those with UC were enriched in Bacteroides ovatus, and subjects with CD showed a significant decrease in Bacteroides monomorpha, Bacteroides vulgaris, and Bacteroides xylanisolvens. Although the gut microbiomes of patients with CRC and colorectal adenomas shared relatively similar microbiome taxa, we found that the abundance of Micromonas parvus was significantly higher in patients with CRC, but not in patients with colorectal adenocarcinoma, when compared with healthy controls, which is consistent with previous findings showing that Micromonas parvus can be used as a marker to distinguish CRC from colorectal adenomas [18,19]. For the other diseases, the differences were mainly from Actinobacteria. Subjects with obesity showed an increase in Actinomyces naeslundii, Actinomyces cariogenicus, and Actinomyces viscosus, and those with IBS-D had a specific increase in Bifidobacterium pseudocatenulatum. Overall, these results suggest that our model captures a variety of disease-specific microbial signatures, which may explain the good diagnostic performance of this multi-class classifier.

[0137] The shared microbial signatures between different disease phenotypes are one of the biggest challenges in developing microbiome-based biomarkers for disease diagnosis

[12] . Our work demonstrates that multi-class disease diagnosis based on fecal microbiome is feasible and that some microbial taxa are unique to specific disease phenotypes. Machine learning models using a multi-class approach can achieve higher performance than traditional single disease models, likely because they can better focus on disease-specific signals and are less sensitive to overfitting. All of these results reveal that our fecal microbiome-based machine learning for multi-class disease diagnosis can be more widely applied.

[0138] Our results also have implications for the development of potential therapeutic options for a variety of diseases by providing potential microbial targets with diagnostic and therapeutic value. A list of bacteria that could be used to treat diseases is included in Table 4. For example, we observed a negative association between enteric Roseburia rosea

[15] and six diseases, so this species may provide a range of health benefits and has the potential to be developed as a universal probiotic. However, a limitation is that there is little biological evidence associated with the microbiome-phenotype associations we identified. Future work to explain the causal relationship of these associations can help understand the role of the gut microbiome in disease pathogenesis. As for other limitations, the pooled public dataset may have various confounding factors, such as comorbidities and antibiotics, so the validation performance may vary with appropriate participant exclusions.

[0139] Overall, we provide the largest fecal microbiome dataset from a mixed disease cohort to date, and we trained a machine learning multi-class classifier that achieved accurate disease classification. This fecal microbiome-based noninvasive diagnostic tool may be applicable to current clinical practice, as the combination of different methods tends to further improve the efficiency of disease screening. For example, the combination of fecal microbiome testing and immunochemical testing showed higher sensitivity for CRC [4]. Therefore, this type of novel noninvasive diagnostic tool may provide substantial future value to the public.

[0140] III. Exemplary Implementation Details

[0141] Study population

[0142] All participants were recruited and diagnosed at the Prince of Wales Hospital, Hong Kong, from January 2017 to March 2022. Subjects with CRC and CA were diagnosed by colonoscopy and confirmed based on histological examination; subjects with CD and UC were diagnosed based on the criteria of endoscopy, radiology, and histological examination. Subjects with IBS were diagnosed according to the ROME III criteria, and endoscopy and colonoscopy were performed to rule out other GI conditions such as IBD, celiac disease, parasitic infection, or other organic conditions. Obesity was defined as a subject with a body mass index (BMI) of more than 28 and no other medical complications. Subjects with cardiovascular disease (CVD) were recruited from the public as part of the General Population Cardiovascular Health Survey in Hong Kong, China. Subjects underwent carotid ultrasound to measure the intima-media thickness (IMT) of the common carotid artery, internal carotid artery, external carotid artery (CCA, ICA, and ECA, respectively), and carotid bulb, and subjects with single or multiple vessel stenosis ≥50% were considered to be at risk for CVD. Subjects with post-acute COVID-19 syndrome (PACS) were defined as those with at least one persistent symptom or long-term complication of SARS-CoV-2 infection that lasted more than 4 weeks from viral clearance (which could not be explained by alternative diagnoses), and we assessed the presence of the 30 most commonly reported symptoms after COVID onset [10,20]. All subjects with other medical conditions (except the obesity group) had a normal BMI range of 18.5-22.9. All subjects were on a stable traditional Chinese type diet and were of Han Chinese ethnicity. Patients were excluded if they had: age below 18 years or above 80 years; self-reported comorbidities of other diseases; infection with enteric pathogens; acquired immunodeficiency syndrome; known history of organ dysfunction or failure and abdominal surgery; active malignancy or receiving radiochemotherapy; short bowel syndrome; taking medications generally known to affect the gut microbiome, including proton pump inhibitors, oral antidiabetic drugs, nonsteroidal anti-inflammatory drugs, corticosteroids, laxatives, or selective serotonin response inhibitors, and use of antibiotics or probiotics within three months of sample collection; pregnancy or breastfeeding; and special diets such as vegetarians.

[0143] During the same recruitment period, healthy controls were recruited from the community and the endoscopy center of Prince of Wales Hospital by advertisement among subjects with normal colonoscopies (stool was collected before bowel preparation). Exclusion criteria for healthy controls were known complicated infection or sepsis; known history of severe organ failure (including decompensated cirrhosis, malignant disease, renal failure, epilepsy, active severe infection, acquired immunodeficiency syndrome); bowel surgery within the last 6 months (excluding colonoscopies / surgeries related to perianal disease); presence of ileostomy / stoma; and current pregnancy; any long-term medication for chronic disease; use of antibiotics within the last 3 months; use of laxatives or antidiarrheal medications within the last 3 months or recent dietary changes (e.g., becoming vegetarian / vegan). Finally, a total of 2,320 subjects were recruited. Clinical metadata and dietary data were collected during clinical interviews.

[0144] In addition, an additional 60 subjects (mean age 53.5 years, 48.3% female) were prospectively followed up for up to two years after COVID-19 infection and were confirmed to have fully recovered from the initial infection without any PACS symptoms. These subjects served as an independent validation cohort and provided stool samples after SARS-CoV-2 clearance. All subjects were over 18 years of age and provided written informed consent.

[0145] Stool samples

[0146] Human stool samples were collected, immediately frozen, and DNA was purified by standard methods. Stool samples from participants were collected by hospital staff after completion of diagnostic and recruitment procedures. All samples were collected in tubes containing preservative medium (cat. 63700, Norgen Biotek Corp, Ontario Canada) and immediately stored at −80°C until processing. We have previously shown that data from gut microbiome cohorts generated from stool collected using this preservative medium are comparable to those obtained from samples immediately stored at −80°C

[21] .

[0147] Fecal DNA extraction and sequencing

[0148] After removing the preservative medium, the fecal pellet (100 mg) was added to 1 mL of CTAB buffer and vortexed for 30 seconds, and then the sample was heated at 95 ° C for 5 minutes. After that, the sample was thoroughly vortexed with beads at maximum speed for 15 minutes. Then, 40 μL of proteinase K and 20 μL of RNase A were added to the sample, and the mixture was incubated at 70 ° C for 10 minutes. The supernatant was then obtained by centrifugation at 13,000g for 5 minutes and added to the MaxwellRSC machine for DNA extraction according to the instructions (Cat: AS1700, Maxwell, USA). After quality control procedures were performed by Qubit 2.0, agarose gel electrophoresis and Agilent 2100, the extracted DNA was subjected to DNA library construction using the Nextera DNA Flex library preparation kit (Illumina, San Diego, CA), which was completed by the process of end repair, adding A to the tail, purification and PCR amplification. The libraries were subsequently sequenced on our in-house sequencer, Illumina NextSeq 550 (150 base pair paired end), at the Microbiome Research Center of the Chinese University of Hong Kong. All samples were subjected to DNA extraction, library construction, and sequencing in a random order. ZymoBIOMICS Spike-in Control I (high microbial load, Cat: D6320-10, ZYMO Research, USA) and ZymoBIOMICS Microbial Community DNA Standard (Cat: D6306-A) were used as positive controls during DNA extraction, library construction, sequencing, and quality assessment.

[0149] Microbiome analysis

[0150] The raw sequence data were quality filtered using Trimmomatic V.39 to remove adapters, low-quality sequences (quality score <20), and reads shorter than 50 base pairs. Contaminated human reads were filtered using Kneaddata (V.0.7.2, genome reference database: GRCh38 p12) using default parameters. After that, microbial cohorts were inferred into features from quality-filtered forward reads using MetaPhlAn3, version 3.0.5. GNU parallel was used to parallelize the analysis work to speed up data processing. Alpha diversity metrics (Shannon diversity, Chao1 richness) were calculated using the phyloseq package, version 1.26.0. Species with average abundance and prevalence below 0.15% and 5% were filtered out.

[0151] Random Forest Binary Classifier

[0152] Binary subcohorts consisted of two phenotypes taken from the entire cohort. A total of 36 binary subcohorts were generated. Machine learning binary classifiers used random forests with the Sklearn

[22] library under Python 3.6.7, as this algorithm has been shown to outperform other learning tools on average for microbiome data

[23] . Models were trained using standardized abundance tables from each binary subcohort. The machine learning model was first trained on a randomly selected training set (70%, 5-fold cross-validation) and then applied to a held-out validation set (30%) to obtain the final performance. This process was repeated 20 times to obtain the distribution of random forest prediction scores on the validation set, and the average AUC value was calculated accordingly for visualization of the results.

[0153] Machine learning for diagnosing multiple diseases

[0154] All classifier models were implemented using Python 3.6.7 using the following publicly available standard libraries: pandas (0.23.4), numpy (1.14.5), scikit-learn (1.1), and matplotlib (2.2.3). Classifier models were trained and evaluated on the microbiome dataset using the cross-validation (CV) method. Machine learning-based classifiers were implemented using the python library Sklearn

[22] . For each phenotype, samples were randomly divided into a training set (70% of the samples, n = 1,724) and a test set for independent evaluation (the remaining 30%, n = 696). Random forest (RF), K nearest neighbor (KNN), SVM multilayer perceptron (MLP), support vector machine (SVM), and graph convolutional neural network (GCN) were used as classifier models for diagnosing different phenotypes by using the classification features of the fecal microbiome. We implemented the RF multi-class classifier with the following modifications to the default SciKit-learn settings: n_estimaters = 2000 and class_weight = balanced. KNN, SVM, and MLP were implemented from SciKit-learn using default settings. The best model selected based on the cross-validation results was evaluated in the held-out evaluation dataset as the final performance for predicting sudden illness. Highly ranked and frequently selected microbial features were considered as predictive features for further interpretation. We retrieved the prediction performance using the same training dataset.

[0155] Cross Validation

[0156] A nested cross-validation procedure was applied to calculate the accuracy within the cohort by partitioning the data into training and test sets for 20 repetitions of five-fold stratified cross-validation (balancing the class ratio between folds). For each partition, an L1 regularized (Lasso) logistic regression model was trained on the training set and then used to predict the test set. The lambda parameter for each model was chosen to maximize the AUC-ROC under the constraint that the model contained at least five non-zero coefficients.

[0157] Binary subcohorts consisted of two phenotypes taken from the entire cohort. A total of 36 binary subcohorts were generated. Machine learning binary classifiers used random forests with the Sklearn

[22] library under Python 3.6.7, as this algorithm has been shown to outperform other learning tools on average for microbiome data

[23] . Models were trained using standardized abundance tables from each binary subcohort. The machine learning model was first trained on a randomly selected training set (70%, 5-fold cross-validation) and then applied to a held-out validation set (30%) to obtain the final performance. This process was repeated 20 times to obtain the distribution of random forest prediction scores on the validation set, and the average AUC value was calculated accordingly for visualization of the results.

[0158] Statistical analysis

[0159] Statistical analyses were performed using R, version 3.4.3. Nonparametric statistical tests were performed between groups using the ggpubr package (https: / / github.com / kassambara / ggpubr), and corrections for multiple hypothesis testing were considered when necessary. Principal coordinate analysis (PCoA) was used to visualize the clustering of samples based on their species-level compositional characteristics. Multivariate analysis was used to identify associations of specific microbial species with phenotypes using the linear model (MaAsLin2) statistical framework implemented in the Huttenhower Lab Galaxy instance (http: / / huttenhower.sph.harvard.edu / galaxy / ). PCoA, PERMANOVA, and MaAsLin2 analyses were implemented in the vegan R package V.2.5–7. The Benjamini-Hochberg correction was used to control for multiple testing, and results were considered significant at FDR < 0.05.

[0160] Public data download and processing

[0161] A total of 1597 raw shotgun fecal metagenomes were obtained from 12 independently published studies in 11 countries (UC, 114 [24-26]; CD, 102 [25-27]; CVD, 218 [8]; CRC, 177 [28-30]; adenoma, 86 [29,30]; IBS-D, 83 [31-33]; obesity, 81

[34] and corresponding 736 healthy controls). Publicly available raw sequencing data were downloaded through the Sequence Read Archive using accession numbers retrieved from the cited papers. After downloading, quality filtering and species-level taxonomic analysis were performed according to the above process.

[0162] Data availability

[0163] Raw microbiome sequencing data are publicly available in the Sequence Read Archive under BioProject accession number: PRJNA841786. All analytical or phenotypic data can be accessed by making an appropriate request to the corresponding author (SCN), after verifying that the request complies with any patient confidentiality obligations. Analytical and phenotypic data can only be requested for research / scientific purposes in accordance with the informed consent signed by the study participants, which specifies that the collected data will not be used for commercial purposes. CU-Med Biobank will additionally evaluate the submitted request and will respond to the request within four weeks. All other data supporting the findings of this study are available in the paper and its supplementary files.

[0164] Code availability

[0165] Open source codes and scripts for microbiome analysis or plots are available at the GitHub repository (https: / / github.com / qsu123 / multi-class-classificer). All machine learning data processing and modeling were performed on Python 3.6.7 using the following publicly available standard libraries: pandas (0.23.4), numpy (1.14.5), scikit-learn (1.1), and matplotlib (2.2.3). The developed multi-classification algorithm and trained models are available upon appropriate request to the corresponding author (SCN) and a written statement of non-commercial use.

[0166] IV. Predicting Health Status

[0167] A. Train a model to predict health status

[0168] Fig.151 is a diagram of a method 1500 for training a model to predict a health condition according to an embodiment. Method 1500 can train a multi-class machine learning model to determine the risk (e.g., probability) of multiple conditions in a subject. Method 1500 and any method described herein can be performed in whole or in part using a computer system including one or more processors that can be configured to perform steps. Thus, embodiments relate to a computer system configured to perform any method steps described herein, which may have different components that perform corresponding steps or corresponding groups of steps.

[0169] At block 1510, a training data set of subjects having a plurality of known health classifications is generated. A training data set for each cohort of M subjects may be generated. As examples, health classifications (e.g., health conditions) may include post-acute COVID-19 syndrome (PACS), Crohn's disease (CD), colorectal cancer (CRC), colorectal adenocarcinoma (CA), obesity (Ob), and cardiovascular disease (CVD).

[0170] For each cohort in the M cohorts of subjects, the relative abundance of DNA fragments corresponding to each of the N bacterial species can be measured in the sample of each subject. N can be 10 or greater, for example, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 25, 30, 35, 40, 45, 50, 60, 70, 80, 90 or 100 or other numbers described herein. Other values, such as 5, 6, 7, 8 or 9, can also be used. Each subject in the cohort has a health classification, so that the M cohorts correspond to M health classifications. M can be 3 or greater, for example, 3, 4, 5, 6, 7, 8, 9, 10 or other values ​​described herein. The M health classifications include health and multiple conditions. At least 10 of the N bacterial species are present in subjects greater than a specified percentage, for example, as described below.

[0171] Training data set can be generated by measuring the relative abundance of deoxyribonucleic acid (DNA) fragments in the sample. DNA can correspond to the bacterial species in the sample, and the sample can be a fecal sample from each object in the queue. Other samples can be used, such as intestinal mucosal samples. There can be 3 or more queues, wherein each queue contains 10 or more objects, and the queue corresponds to an object (e.g., individual) with a health classification (e.g., health status). At least 20 bacterial species in the sample can be present in the object of at least a specified percentage (e.g., 5%, 10%, 15%, 20%, 25%, 50%, 75% or more), and if the relative abundance of the species is greater than 0.01%, 0.05%, 0.10%, 0.15%, 0.20%, 0.50%, 1.00% or more, the species can exist.

[0172] The bacterial species can be selected from the species lists, such as Table 1, Table 2, Table 3, Figure 3D , Fig. 13C or Fig.14 In one embodiment, the bacterial species include Blautia wexlerae, Fusicatenibacter saccharivorans, Bacteroides vulgaris, Agathobaculum butyriciproducens and Dorea longicatena. The bacterial species may include bacterial species with the top 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 25, 30, 35, 40, 45, 50, 60, 70, 80, 90 or 100 points from Table 1.

[0173] Measuring the relative abundance of DNA fragments for each subject can include receiving a series of reads from sequencing a stool sample, and can include sequencing itself. In some cases, intestinal mucosal samples can be sequenced. The sequence reads can be compared with a human reference genome and a bacterial genome database, each bacterial genome corresponding to a different bacterial species. The relative abundance of each bacterial species can be determined using the sequence reads corresponding to the bacterial reference genome.

[0174] As another example, the relative abundance of DNA fragments of each object can be measured and can include the use of probe-based technology. Probe-based technology can be used to measure the quantity of DNA fragments of selected features and the total number of DNA fragments. The signal of a specific probe can be provided, wherein the signal (e.g., an intensity signal or a digital signal indicating the presence or absence) can provide a plurality of DNA fragments corresponding to the specific probe. The relative abundance of each feature can be calculated by the above numbers.

[0175] At block 1520, a feature vector may be generated for each subject. The feature vector may contain the relative abundance of bacterial species in the subject's sample. The feature vector may be an N-dimensional vector, where N is the number of bacterial species. The relative abundance of a species may be the number of that bacterial species divided by the total number of bacteria in the sample (e.g., the percentage of a particular bacterium compared to the total number of bacteria in the sample).

[0176] At box 1530, a multi-class machine learning model can be trained using a training data set. The machine learning model can be a random forest (RF) model, a k-nearest neighbor model (KNN), a multi-layer perceptron model (MLP), or a support vector machine (SVM). The multi-class machine learning model can provide a probability for each of M health classifications (e.g., health conditions), and the training data set can include a known health classification of the object and a feature vector of the object. The model can be trained by an algorithm by optimizing the sensitivity and specificity for determining the correct condition by comparing the highest probability to a threshold corresponding to the correct health condition. Training can optimize the sensitivity and specificity for determining the correct condition (i.e., the output of the model matches the known health classification) by achieving the highest average AUC for the M health classifications.

[0177] If the probability of the healthy classification is the highest probability or if the probability is above a threshold, the individual can be classified as having the health condition. An individual that has been classified using the healthy classification can be treated for the condition. Treatment can include changing one or more existing treatment courses to change the fecal microbiome (e.g., fecal transplants, probiotics, etc.).

[0178] B. Use the trained model to predict health status

[0179] Fig.16 1 is a diagram of a method 1600 for predicting a health condition according to an embodiment. Method 1600 may implement a multi-class machine learning model to distinguish between multiple possible diseases of an object. Method 1600 and any method described herein may be performed in whole or in part using a computer system including one or more processors that may be configured to perform steps. Thus, embodiments relate to a computer system configured to perform any method steps described herein, which may have different components that perform corresponding steps or corresponding groups of steps. Method 1600 may be performed using the techniques described herein.

[0180] At block 1610, the relative abundance of the DNA fragments can be measured in the sample of the subject. The DNA fragments can correspond to bacterial species in the sample, and the relative abundance of each of the N bacterial species can be measured. N can be 10 or greater, for example, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 25, 30, 35, 40, 45, 50, 60, 70, 80, 90 or 100 or other numbers described herein. Other values, such as 5, 6, 7, 8 or 9, can also be used.

[0181] DNA can correspond to the bacterial species in the sample, and the sample can be a fecal sample from each object in the queue. Other samples can be used, such as intestinal mucosal samples. There can be 3 or more queues, wherein each queue comprises 10 or more objects, and the queue corresponds to an object (e.g., individual) with a health classification (e.g., health status). At least 20 bacterial species in the sample can be present in the object of at least a specified percentage (e.g., 5%, 10%, 15%, 20%, 25%, 50%, 75% or more), and if the relative abundance of the species is greater than 0.01%, 0.05%, 0.10%, 0.15%, 0.20%, 0.50%, 1.00% or more, the species can exist.

[0182] The bacterial species can be selected from the species lists, such as Table 1, Table 2, Table 3, Figure 3D , Fig. 13C or Fig.14 In one embodiment, the bacterial species include Blautia wesleyi, Streptobacillus sucroseus, Bacteroides vulgaris, Bacteroides butyricum, and Dorea_longicatena. The bacterial species may include bacterial species with the top 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 25, 30, 35, 40, 45, 50, 60, 70, 80, 90 or 100 points from Table 1.

[0183] Measuring the relative abundance of DNA fragments for each subject can include receiving a series of reads from sequencing a stool sample, and can include sequencing itself. In some cases, intestinal mucosal samples can be sequenced. The sequence reads can be compared with a human reference genome and a bacterial genome database, each bacterial genome corresponding to a different bacterial species. The relative abundance of each bacterial species can be determined using the sequence reads corresponding to the bacterial reference genome.

[0184] As another example, the relative abundance of DNA fragments of each object can be measured and can include the use of probe-based technology. Probe-based technology can be used to measure the quantity of DNA fragments of selected features and the total number of DNA fragments. The signal of a specific probe can be provided, wherein the signal (e.g., an intensity signal or a digital signal indicating the presence or absence) can provide a plurality of DNA fragments corresponding to the specific probe. The relative abundance of each feature can be calculated by the above numbers.

[0185] At block 1620, a feature vector may be generated using the relative abundance of N bacterial species of the subject. The N bacterial species of the subject may be bacterial species identified in a sample derived from the subject. The feature vector may contain the relative abundance of bacterial species in the subject sample. The feature vector may be an N-dimensional vector, where N is the number of bacterial species. The relative abundance of a species may be the number of the bacterial species divided by the total number of bacteria in the sample (e.g., the percentage of a particular bacterium compared to the total number of bacteria in the sample).

[0186] At block 1630, M probabilities of M health classifications may be determined by operating the feature vector using a multi-class machine learning model. M may be 3 or greater, e.g., 3, 4, 5, 6, 7, 8, 9, 10, or other values ​​described herein. As examples, health classifications (e.g., health conditions) may include post-acute COVID-19 syndrome (PACS), Crohn's disease (CD), colorectal cancer (CRC), colorectal adenocarcinoma (CA), obesity (Ob), and cardiovascular disease (CVD).

[0187] The machine learning model can be a random forest (RF) model, a k-nearest neighbor model (KNN), a multi-layer perceptron model (MLP), or a support vector machine (SVM). The multi-class machine learning model can provide a probability for each health classification (e.g., health condition), and the training data set can include the known health classification of the object and the feature vector of the object. The model can be trained by the algorithm by comparing the highest probability with the threshold corresponding to the correct health condition, by optimizing the sensitivity and specificity for determining the correct health condition.

[0188] At block 1640, the highest probability of the M probabilities may be identified. If the probability of the healthy classification is the highest probability or if the probability is above a threshold, the individual may be classified as having a health condition. An individual that has been classified using a healthy classification may be treated for the condition. Treatment may include altering one or more existing treatment courses to alter the fecal microbiome (e.g., fecal transplants, probiotics, etc.).

[0189] At block 1650, the highest probability may be compared to a corresponding threshold value corresponding to a first condition in the plurality of conditions. The M probabilities may be compared to corresponding threshold values ​​corresponding to the plurality of conditions. It may be determined whether the subject suffers from the plurality of conditions based on the probabilities exceeding the corresponding threshold values.

[0190] At block 1660, it can be determined whether the subject has a first condition based on the highest priority exceeding the corresponding threshold. The subject can be treated for the condition. For example, the subject can be treated with a microbial therapy (e.g., a microbial transplant). The microbial therapy can be selected from the list disclosed in Table 4.

[0191] V. Exemplary Systems

[0192] Any computer system mentioned herein may utilize any suitable number of subsystems. Examples of such subsystems are Fig.17 1710. In some embodiments, the computer system includes a single computer device, wherein the subsystem can be a component of the computer device. In other embodiments, the computer system can include multiple computer devices, each of which is a subsystem with internal components. The computer system can include desktop and laptop computers, tablet computers, mobile phones and other mobile devices.

[0193] Fig.17 The subsystems shown in FIG. 17 are connected to each other via a system bus 1775. Additional subsystems are shown, such as a printer 1774, a keyboard 1778, a storage device 1779, a monitor 1776 coupled to a display adapter 1782 (e.g., a display screen such as an LED), etc. Peripheral devices and input / output (I / O) devices coupled to the I / O controller 1771 can be connected to the computer system by any number of means known in the art, such as input / output (I / O) ports 1777 (e.g., USB, ). For example, I / O port 1777 or external interface 1781 (e.g., Ethernet, Wi-Fi, etc.) can be used to connect computer system 1710 to a wide area network, such as the Internet, a mouse input device, or a scanner. The interconnection via system bus 1775 allows central processor 1773 to communicate with each subsystem and control the execution of multiple instructions from system memory 1772 or storage device 1779 (e.g., a fixed disk, such as a hard drive or optical disk), as well as the exchange of information between subsystems. System memory 1772 and / or storage device 1779 can be embodied as a computer-readable medium. Another subsystem is a data collection device 1785, such as a camera, microphone, accelerometer, etc. Any data mentioned herein can be output from one component to another and can be output to a user.

[0194] The computer system may include, for example, multiple identical components or subsystems connected together through external interfaces 1781, through internal interfaces, or via removable storage devices that can be connected from one component to another and removed therefrom. In some embodiments, the computer system, subsystem, or device may communicate over a network. In this case, one computer may be considered a client and another computer may be considered a server, where each computer may be part of the same computer system. The client and server may each include multiple systems, subsystems, or components.

[0195] Aspects of the embodiments may be implemented in the form of control logic using hardware circuits (e.g., application specific integrated circuits or field programmable gate arrays) and / or using computer software stored in a memory with a general programmable processor in a modular or integrated manner, so the processor may include a memory storing software instructions for configuring the hardware circuits and an FPGA or ASIC with configuration instructions. As used herein, the processor may include a single-core processor, a multi-core processor on the same integrated chip, or multiple processing units on a single circuit board or networked, and dedicated hardware. Based on the disclosure and teachings provided herein, those of ordinary skill in the art will know and understand other ways and / or methods of implementing the embodiments of the present disclosure using hardware and a combination of hardware and software.

[0196] Any software component or function described in this application can be implemented as a software code executed by a processor using any suitable computer language, such as, for example, Java, C, C++, C#, Objective-C, Swift, or a scripting language using, for example, conventional or object-oriented techniques, such as Perl or Python. The software code can be stored as a series of instructions or commands on a computer-readable medium for storage and / or transmission. Suitable non-transitory computer-readable media may include random access memory (RAM), read-only memory (ROM), magnetic media such as a hard drive or floppy disk, or optical media such as a compact disk (CD) or DVD (digital versatile disk) or Blu-ray disc, flash memory, etc. The computer-readable medium can be any combination of such devices. In addition, the order of operations can be rearranged. The process can terminate when its operation is completed, but may have additional steps not included in the figure. The process may correspond to a method, a function, a program, a subroutine, a subprogram, etc. When the process corresponds to a function, its termination may correspond to the function returning to the calling function or the main function.

[0197] Such program can also be encoded and transmitted using a carrier signal suitable for transmission via wired, optical and / or wireless network transmissions that meet the various protocols including the Internet. Thus, a data signal utilizing such program encoding can be used to create a computer-readable medium. The computer-readable medium encoded with program code can be packaged with a compatible device, or provided separately from other devices (e.g., downloaded via the Internet). Any such computer-readable medium can be located on or inside a single computer product (e.g., a hard drive, a CD or an entire computer system), and can be present on or inside different computer products in a system or network. A computer system can include a monitor, a printer or other suitable displays, for providing any result mentioned herein to a user.

[0198] Fig.18A measurement system 1800 according to an embodiment of the present disclosure is shown. The system shown includes a sample 1805, such as a cell-free DNA molecule in an assay device 1810, wherein the sample 1805 can be assayed 1808. For example, the sample 1805 can be contacted with a reagent for assay 1808 to provide a signal of a physical feature 1815. An example of an assay device can be a flow cell, which includes a probe and / or primer for an assay or a tube through which a droplet moves (wherein the droplet includes an assay). Physical features 1815 (e.g., fluorescence intensity, voltage, or current) from the sample are detected by a detector 1820. The detector 1820 can be measured at intervals (e.g., periodic intervals) to obtain data points that constitute a data signal. In one embodiment, an analog-to-digital converter converts the analog signal from the detector into a digital form multiple times. The assay device 1810 and the detector 1820 can form an assay system, for example, a sequencing system for performing sequencing according to embodiments described herein. The data signal 1825 is sent from the detector 1820 to a logic system 1830. As an example, data signal 1825 can be used to determine the sequence and / or position of a reference genome of a DNA molecule. Data signal 1825 can include various measurements made simultaneously, for example, different colored fluorescent dyes or different electrical signals for different molecules of sample 1805, so data signal 1825 can correspond to multiple signals. Data signal 1825 can be stored in local memory 1835, external memory 1840, or storage device 1845.

[0199] The logic system 1830 may be or may include a computer system, an ASIC, a microprocessor, a graphics processing unit (GPU), etc. It may also include a display (e.g., a monitor, an LED display, etc.) and a user input device (e.g., a mouse, a keyboard, a button, etc.) or be coupled to the display and the user input device. The logic system 1830 and other components may be part of an independent or network-connected computer system, or they may be directly attached to or integrated in a device (e.g., a sequencing device) including a detector 1820 and / or a determination device 1810. The logic system 1830 may also include software executed in the processor 1850. The logic system 1830 may include a computer-readable medium that stores instructions for controlling the measurement system 1800 to perform any method described herein. For example, the logic system 1830 may provide a command to a system including the determination device 1810 so as to perform sequencing or other physical operations. This physical operation may be performed in a specific order, for example, adding and removing reagents in a specific order. This physical operation may be performed by a robotic system, which may include, for example, a mechanical arm that may be used to obtain a sample and perform a determination.

[0200] System 1800 can also include a treatment device 1860 that can provide a treatment to a subject, such as, for example, a treatment selected from Table 4. Treatment device 1860 can determine a treatment and / or be used to perform a treatment. Examples of such treatments can include histamine antagonist physical therapy, counseling, breathing exercises, pulmonary and cardiac rehabilitation, memory exercises, olfactory training for acute post-COVID-19 syndrome; anti-inflammatory drugs (e.g., 5-aminosalicylic acid), immunosuppressants (e.g., azathioprine, mercaptopurine, methotrexate, cyclosporine), biologics (e.g., tumor necrosis factor-α blockers, integrin blockers, interleukin blockers), and surgery for Crohn's disease and ulcerative colitis; surgery, radiation therapy, chemotherapy, immunotherapy, targeted therapy, hormone therapy, and stem cell transplantation for colorectal cancer and colorectal adenomas; calorie-controlled diet, exercise, naltrexone / bupropion, phentermine / topiramate ER, glucagon-like Peptide-1 receptor agonists, sodium-glucose cotransporter-2 inhibitors, and bariatric surgery for obesity; low-dose aspirin, clopidogrel, rivaroxaban, ticagrelor, prasugrel, statins (e.g., atorvastatin, simvastatin, rosuvastatin, pravastatin), beta blockers (e.g., atenolol, bisoprolol, metoprolol), nitrates, angiotensin converting enzyme inhibitors (e.g., ramipril, lisinopril), angiotensin-2 receptor blockers, calcium channel blockers (e.g., amlodipine, verapamil), diuretics, coronary angioplasty, coronary artery bypass grafting for cardiovascular disease; dietary modification, smooth muscle relaxants, low-dose antidepressants, psychotherapy for IBS. Such treatment may also include regulating the microbiome by supplementing with probiotics, prebiotics, or synbiotics or by changing the diet. Logic system 1830 may be connected to treatment device 1860, for example, to provide the results of the methods described herein. The treatment device may receive input from other devices, such as imaging devices and user input (eg, to control the treatment, such as control of a robotic system).

[0201] Any method described herein can be performed in whole or in part using a computer system including one or more processors, which can be configured to perform steps. Any operation performed using a processor (e.g., comparison, determination, comparison, calculation, calculation) can be performed in real time. The term "real time" can refer to a computational operation or process completed within a certain time limit. The time limit can be 1 minute, 1 hour, 1 day or 7 days. Therefore, the embodiment can be related to a computer system configured to perform any method steps described herein, which may have different components for performing corresponding steps or corresponding groups of steps. Although presented as numbered steps, the method steps herein can be performed at the same time or at different times or in different orders. In addition, the parts of these steps can be used together with the parts of other steps from other methods. Moreover, all or part of the steps can be optional. In addition, any step of any method can be performed using the module, unit, circuit or other means of the system for performing these steps.

[0202] Without departing from the spirit and scope of the embodiments of the present disclosure, the specific details of a particular embodiment may be combined in any suitable manner. However, other embodiments of the present disclosure may be directed to specific embodiments related to each individual aspect, or specific combinations of these individual aspects.

[0203] The above description of exemplary embodiments of the present disclosure has been presented for purposes of illustration and description. It is not intended to be exhaustive or to limit the disclosure to the precise form described, and many modifications and variations are possible in light of the above teachings.

[0204] Unless expressly stated otherwise, the recitation of "a", "an", or "the" is intended to mean "one or more". Unless expressly stated otherwise, the use of "or" is intended to mean an "inclusive or", and not an "exclusive or". Reference to a "first" component does not necessarily require that a second component be provided. Furthermore, unless expressly stated, reference to a "first" or "second" component does not limit the referenced components to a particular location. The term "based on" is intended to mean "based, at least in part, on".

[0205] All patents, patent applications, publications and descriptions mentioned herein are incorporated herein by reference in their entirety for all purposes. None is acknowledged to be prior art. In the event of a conflict between the present application and the references provided herein, the present application shall prevail.

[0206] Appendix A - Table 1: List of 325 bacterial species and their importance scores

[0207]

[0208]

[0209]

[0210]

[0211]

[0212]

[0213]

[0214]

[0215] Appendix B - Table 2: List of 50 bacterial species and their importance scores

[0216]

[0217]

[0218]

[0219] Appendix C - Table 3: List of 10 bacterial species and their importance scores

[0220] sample score Blautia westernii 0.013072 Streptobacillus sucrosevorans 0.012885 Bacteroides vulgaris 0.011013 Bacillus butyrate-producing bacteria 0.010856 Dorea_longicatena 0.010583 Eubacterium hallii 0.010097 Faecalibacterium prausnitzii 0.009772 Bacteroides thetaiotaomicron 0.009741 Bacteroides monomorpha 0.009704 Escherichia coli 0.009689

[0221] Appendix D-Table 4: List of bacteria that can be used to treat corresponding diseases

[0222]

[0223]

[0224]

[0225] References

[0226] 1. Lynch, SV & Pedersen, O. The Human Intestinal Microbiome in Health and Disease. New England Journal of Medicine 375, 2369-2379 (2016).

[0227] 2.Metwaly,A.,Reitmeier,S.&Haller,D.Microbiome risk profiles asbiomarkers for inflammatory and metabolic disorders.Nat Rev GastroenterolHepatol(2022).

[0228] 3.Lin,Y.,Wang,G.,Yu,J.&Sung,J.J.Y.Artificial intelligence andmetagenomics in intestinal diseases.J Gastroenterol Hepatol 36,841-847(2021).

[0229] 4.Liang,J.Q.,et al.Anovel fecal Lachnoclostridium marker for the non-invasive diagnosis of colorectal adenoma and cancer.Gut 69,1248-1257(2020).5.Vila,A.V.,et al.Gut microbiota composition and functional changesininflammatory bowel disease and irritable bowelsyndrome.ScienceTranslational Medicine 10,eaap8914(2018).

[0230] 6.Shaukat,A.&Levin,T.R.Current and future colorectal cancerscreeningstrategies.Nat Rev Gastroenterol Hepatol(2022).

[0231] 7.Turnbaugh,P.J.,et al.An obesity-associated gut microbiome withincreasedcapacity for energy harvest.Nature 444,1027-1031(2006).

[0232] 8.Jie,Z.,et al.The gut microbiome in atherosclerotic cardiovasculardisease.Nat Commun 8,845(2017).

[0233] 9.Yeoh,Y.K.,et al.Gut microbiota composition reflects diseaseseverity anddysfunctional immune responses in patients with COVID-19.Gut 70,698-706(2021).

[0234] 10.Liu,Q.,et al.Gut microbiota dynamics in a prospective cohort ofpatientswith post-acute COVID-19 syndrome.Gut 71,544-552(2022).

[0235] 11.Vestad,B.,et al.Respiratory dysfunction three months after severeCOVID-19 is associated with gut microbiota alterations.J Intern Med 291,801-812(2022).

[0236] 12.Gacesa,R.,et al.Environmental factors shaping the gut microbiomein aDutch population.Nature 604,732-739(2022).

[0237] 13.Gupta,V.K.,et al.Apredictive index for health status usingspecies-levelgut microbiome profiling.Nat Commun 11,4635(2020).

[0238] 14.Wyres,K.L.,Lam,M.M.C.&Holt,K.E.Population genomics ofKlebsiellapneumoniae.Nat Rev Microbiol 18,344-359(2020).

[0239] 15.Nie,K.,et al.Roseburia intestinalis:A Beneficial Gut Organism FromtheDiscoveries in Genus and Species.Front Cell Infect Microbiol 11,757718(2021).

[0240] 16.Margherita Grandini,E.B.,Giorgio Visani.Metrics for Multi-ClassClassification:an Overview.arXiv(2020).

[0241] 17.Stojanov,S.,Berlec,A.&Strukelj,B.The Influence of Probiotics ontheFirmicutes / Bacteroidetes Ratio in the Treatment of Obesity andInflammatoryBowel disease.Microorganisms 8(2020).

[0242] 18.Xu,J.,et al.Alteration of the abundance of Parvimonas micra in thegutalong the adenoma-carcinoma sequence.Oncol Lett 20,106(2020).

[0243] 19.Lowenmark,T.,et al.Parvimonas micra as a putative non-invasivefecalbiomarker for colorectal cancer.Sci Rep 10,15250(2020).

[0244] 20.Nalbandian,A.,et al.Post-acute COVID-19 syndrome.Nat Med 27,601-615(2021).

[0245] 21.Chen,Z.,et al.Impact of Preservation Method and 16SrRNAHypervariable Region on Gut Microbiota Profiling.mSystems 4,e00271-00218(2019).

[0246] 22.Pedregosa,F.,et al.Scikit-learn:Machine learning in Python.theJournalof machine Learning research 12,2825-2830(2011).

[0247] 23.Pasolli,E.,Truong,D.T.,Malik,F.,Waldron,L.&Segata,N.MachineLearning Meta-analysis of Large Metagenomic Datasets:Tools andBiologicalInsights.PLoS Comput Biol 12,e1004977(2016).

[0248] 24.Franzosa,E.A.,et al.Gut microbiome structure and metabolicactivity ininflammatory bowel disease.Nat Microbiol 4,293-305(2019).

[0249] 25.Nielsen,H.B.,et al.Identification and assembly of genomes andgeneticelements in complex metagenomic samples without using referencegenomes.Nat Biotechnol 32,822-828(2014).

[0250] 26.Weng,Y.J.,et al.Correlation of diet,microbiota and metabolitenetworksin inflammatory bowel disease.J Dig Dis 20,447-459(2019).

[0251] 27.He,Q.,et al.Two distinct metacommunities characterize the gutmicrobiotain Crohn's disease patients.Gigascience 6,1-11(2017).

[0252] 28.Yachida,S.,et al.Metagenomic and metabolomic analyses revealdistinctstage-specific phenotypes of the gut microbiota in colorectalcancer.Nat Med25,968-976(2019).

[0253] 29.Feng,Q.,et al.Gut microbiome development along the colorectaladenoma-carcinoma sequence.Nat Commun 6,6528(2015).

[0254] 30.Zeller,G.,et al.Potential of fecal microbiota for early-stagedetection ofcolorectal cancer.Mol Syst Biol 10,766(2014).

[0255] 31.Vervier,K.,et al.Two microbiota subtypes identified in irritablebowelsyndrome with distinct responses to the low FODMAP diet.Gut(2021).

[0256] 32.Goll,R.,et al.Effects of fecal microbiota transplantation insubjects withirritable bowel syndrome are mirrored by changes in gutmicrobiome.GutMicrobes 12,1794263(2020).

[0257] 33.Mars,R.A.T.,et al.Longitudinal Multi-omics Reveals Subset-SpecificMechanisms Underlying Irritable Bowel Syndrome.Cell 182,1460-1473e1417(2020).

[0258] 34.Meslier,V.,et al.Mediterranean diet intervention in overweight andobesesubjects lowers plasma cholesterol and causes changes in the gutmicrobiomeand metabolome independently of energy intake.Gut 69,1258-1268(2020).

Claims

1. A method for using a multi-class machine learning model to distinguish between multiple possible diseases of a subject, the method comprising: for each bacterial species of N bacterial species, measuring the relative abundance of DNA fragments corresponding to the bacterial species in the sample of the subject, wherein N is 10 or greater; generating a feature vector using the relative abundances of N bacterial species of the subject; Determining M probabilities of M health classifications by operating the feature vector using the multi-class machine learning model, M being 3 or greater, wherein the M health classifications include health and multiple conditions; Identifying the highest probability among the M probabilities; comparing the highest probability to a corresponding threshold corresponding to a first condition in the plurality of conditions; as well as Based on the highest probability exceeding the corresponding threshold, the subject is determined to suffer from the first condition.

2. The method of claim 1, further comprising: The subject is treated for the first condition.

3. The method of claim 2, wherein treating the subject comprises microbial therapy.

4. The method of claim 3, wherein the microbial therapy is selected from Table 4.

5. The method of claim 1, further comprising: Comparing the M probabilities with corresponding thresholds corresponding to the multiple conditions; as well as Based on the M probabilities exceeding the corresponding thresholds, it is determined that the subject suffers from multiple conditions. The method of claim 1 , wherein the sample is a stool sample or an intestinal mucosa sample.

7. A method of training a multi-class machine learning model to determine risk of multiple conditions in a subject, the method comprising: A training dataset of objects with multiple known healthy classifications was generated by: For each of the M sets of object queues: For each bacterial species among the N bacterial species, measuring the relative abundance of DNA fragments corresponding to the bacterial species in the sample of each of the subjects, wherein each subject in the cohort has a health classification, such that the M cohorts correspond to M health classifications, M is 3 or greater, wherein the M health classifications include health and multiple conditions, wherein N is 10 or greater, and wherein at least 10 of said N bacterial species are present in greater than a specified percentage of said subjects; For each subject, a feature vector is generated using the relative abundance of N bacterial species in the subject; as well as The multi-class machine learning model is trained using a training data set including known health classifications of the object and feature vectors of the object, wherein the multi-class machine learning model provides a probability of each of the M health classifications, and wherein the training optimizes the sensitivity and specificity for determining the correct condition by achieving the highest average AUC for the M health classifications.

8. The method of claim 7, wherein the specified percentage is at least 5%.

9. The method of claim 7, wherein the sample from each of the subjects is a stool sample or an intestinal mucosal sample.

10. The method of claim 1 or claim 7, wherein the N bacterial species are selected from Table 1.

11. The method of claim 1 or claim 7, wherein the N bacterial species are selected from Table 2.

12. The method of claim 1 or claim 7, wherein the N bacterial species are selected from Table 3.

13. The method of claim 1 or claim 7, wherein measuring the relative abundance of DNA fragments comprises: For each subject: receiving a set of sequence reads obtained from sequencing of the sample; as well as aligning the set of sequence reads to a database of a human reference genome and a bacterial reference genome; as well as Using sequence reads corresponding to the bacterial reference genome, the relative abundance of each of the N bacterial species is determined.

14. The method of claim 1 or claim 7, wherein measuring the relative abundance of DNA fragments uses probes.

15. The method of claim 1 or claim 7, wherein the multiple conditions include post-acute COVID-19 syndrome (PACS), Crohn's disease (CD), ulcerative colitis (UC), colorectal cancer (CRC), colorectal adenocarcinoma (CA), obesity (Ob), irritable bowel syndrome with diarrhea (IBS-D), and cardiovascular disease (CVD).

16. The method of claim 1 or claim 7, wherein the multi-category machine learning model comprises a random forest, a K-nearest neighbor, a multi-layer perceptron, a graph convolutional neural network, or a support vector machine.

17. A computer product comprising a non-transitory computer readable medium storing a plurality of instructions which, when executed, control a computer system to perform the method of any one of the preceding claims.

18. A system comprising: The computer product of claim 17; as well as One or more processors for executing instructions stored on the non-transitory computer readable medium.

19. A system comprising means for performing any of the above methods.

20. A system comprising one or more processors configured to perform any of the above methods.

21. A system comprising modules respectively performing the steps of any of the above methods.

Images