Systems and methods for deriving and optimizing classifiers from multiple data sets

Abstract

Systems and methods are provided for object clinical condition assessment using a plurality of modules. The modules include their respective feature values associated with features of a phenotype related to the clinical condition, absence, presence or stage. For at least a first of the plurality of modules, a first data set having feature values is obtained, the feature values being obtained from respective objects in a transcriptomic, proteomic or metabolomic form by a first technical context. A second training data set having feature values is obtained, the feature values being obtained from training objects of the second data set in the same form as the first data set of the at least first module by a technical context other than the first technical context. Batch effects between the data sets are removed by co-normalizing the feature values across the training data sets, thereby computing co-normalized feature values for training a classifier for clinical condition assessment of test objects.

Description

[0001] Cross-reference to related applications This application claims priority to U.S. Provisional Patent Application 62 / 822,730, filed March 22, 2019, the entire contents of which are incorporated herein by reference for all purposes. Technical Field

[0002] This disclosure relates to the training and implementation of a machine learning classifier for assessing the clinical condition of a subject. Background Technology

[0003] Biological modeling approaches relying on transcriptomics and / or other 'omics'-based data (e.g., genomics, proteomics, metabolomics, lipidomics, glycomics, etc.) can be used to provide meaningful and actionable diagnoses and prognoses for medical conditions. For example, some commercial genomic diagnostic tests are used to guide cancer treatment decisions. The Oncotype IQ test suite (Genomic Health) is an example of such genomic-based testing that can provide diagnostic information to guide treatment for various cancers. For instance, one of these tests, ONCOTYPE DX® (Genomic Health) for breast cancer, queries 21 genomic alleles in a patient's tumor to provide diagnostic information to guide treatment for early invasive breast cancer, for example, by providing prognostic information such as potential benefit from chemotherapy and the likelihood of recurrence. See, for example, Paik. et al ., 2004, N Engl J Med. 351, pp. 2817-2825 and Paik et al ., 2016, J Clin Oncol. 24(23), pp. 3726-3734.

[0004] High-throughput omics techniques, such as gene expression microarrays, are often used to discover small panels of targeted biomarkers. However, such datasets always have more variables than samples, making them prone to non-reproducible overfitting. See, for example, Shi... et al ., 2008, BMC Bioinformatics, 9(9), p. S10 and Ioannidis et al ., 2001, Nat Genet. 29(3), pp. 306-09. Furthermore, to increase statistical power, biomarker discovery is often performed using a single type of assay, such as a single type of microarray, in a homogeneous clinical cohort. While this homogeneous design does produce greater statistical power, results are unlikely to remain true across different clinical cohorts using different laboratory techniques. Therefore, any new classifier derived from high-throughput studies requires multiple independent validations.

[0005] Fortunately, technological advancements have led to the development of many different types of high-throughput biological data analysis. This, in turn, has enabled large-scale clinical studies on the biological effects of various medical diseases. Numerous omics-based datasets are available online, for example, in the Gene Expression Omnibus (GEO) hosted by the National Center for Biotechnology Information (NCBI) and the Array Express Archive of Functional Genomics hosted by the European Institute of Bioinformatics (EMBL-EBI). These datasets, along with others (many of which are publicly available), are excellent sources for training machine learning classifiers to distinguish between various disease states and anticipated treatment outcomes, especially because they utilize diverse clinical cohorts and different laboratory techniques. Theoretically, better classifiers can be trained using these diverse datasets because individual patient cohort and detection-specific and batch-specific effects of detection techniques can be identified and ignored, while phenotypic effects induced by underlying biology can be emphasized.

[0006] However, training classifiers on heterogeneous datasets, such as those collected from multiple studies and / or using multiple analytics platforms, is problematic because feature values, such as expression levels, are not comparable across different studies and analytics platforms. In other words, including multiple datasets from different technological and biological backgrounds leads to significant heterogeneity among the included datasets. If not removed, this heterogeneity can obscure the construction of classifiers across datasets. Traditional approaches to training classifiers on heterogeneous datasets simply optimize the parameterized classifier in a single cohort and then apply it externally. However, the diverse technological backgrounds hinder direct application to external datasets, so classifiers are often retrained locally, leading to severely biased estimates of performance. See Tsalik. et al , 2016; and Sci Transl Med 8, 322ra311. In another approach, the non-parametric classifier is optimized on multiple un-co-normalized datasets because it is impossible to optimize these classifiers also in the pooling setting. See Sweeney et al , 2015, Sci Transl Med 7(287), pp.287ra71; and Sweeney et al ., 2016, Sci Transl Med 8(346), pp. 346ra91. Finally, in a recently published work, a group from Sage Bionetworks attempted to learn parameterized models across multiple improperly co-normalized aggregate datasets. However, these models reportedly performed poorly in validation. See Sweeney et al .,2018, Nature Communications 9, 694. Summary of the Invention

[0007] In light of the foregoing background, there is a need in the art for improved methods and systems for developing and implementing more robust and scalable machine learning classifiers. Advantageously, the present invention provides technical solutions (e.g., computational systems, methods, and non-transitory computer-readable storage media) to address these problems and other issues in the field of medical diagnostics. For example, in some embodiments, this disclosure provides methods and systems for generating machine learning classifiers using heterogeneous libraries of input molecules (e.g., genomics, transcriptomics, proteomics, metabolomics) and / or clinical data with relevant clinical phenotypes, for example for diagnosis, prognosis, or clinical prediction, which are more robust and scalable than conventional classifiers.

[0008] Importantly, as described in this paper, non-traditional co-normalization techniques have been developed that reduce the effects of dataset variability and bring data into a single aggregate format. Appropriately co-normalized heterogeneous datasets unlock the potential of machine learning by integrating and overcoming clinical heterogeneity to produce generalizable, accurate classifiers. Therefore, the methods and systems described in this paper enable breakthroughs in the development of novel classifiers using multiple datasets.

[0009] The following is an overview of the invention to provide a basic understanding of some aspects of it. This overview is not a broad summary of the invention. It is not intended to identify key / critical elements of the invention or to depict its scope. The sole purpose of this overview is to present some of the concepts of the invention in a simplified form as a prelude to the more detailed description that follows.

[0010] In some embodiments, this disclosure provides methods and systems for implementing those methods of training neural network classifiers based on a heterogeneous library of clinical data with relevant clinical phenotypes, including input molecules (e.g., genomes, transcriptomes, proteomes, metabolomes). In some embodiments, the method includes a priori identifying biomarkers with statistically significant differential eigenvalues ​​(e.g., gene expression values) in a clinical condition of interest, and determining the sign or direction of the eigenvalues ​​of each biomarker in that clinical condition, e.g., positive or negative. In some embodiments, multiple datasets are collected that typically examine the same clinical condition, such as the presence of an acute infection, for example, a medical condition. The raw data from each of these datasets are then normalized using a study-specific procedure, e.g., using a robust multiarray averaging (RMA) algorithm to normalize gene expression microarray data or Bowtie and Tophat algorithms to normalize RNA sequencing (RNA-Seq) data. The normalized data from each of these datasets are then mapped to a common variable and co-normalized with other datasets. Finally, the co-normalized and mapped datasets are then used to build and train a neural network classifier, where each input unit corresponding to an identified biomarker (whose statistically significant differential eigenvalues ​​have a shared effect sign, e.g., positive or negative, for clinical condition status) is grouped into 'modules' using a uniform sign coefficient to preserve the orientation of the module's genetic effect.

[0011] For example, in one aspect, this disclosure provides methods and systems for performing such methods to assess the clinical condition of test subjects of a species using prior feature grouping, wherein the prior feature grouping comprises multiple modules. Each of the multiple modules comprises independent multiple features, the corresponding feature values ​​of which are each associated with the absence, presence, or stage of an independent phenotype associated with the clinical condition. The method includes obtaining a first training dataset in electronic form, wherein for each of a first plurality of training subjects of a species, the first training dataset comprises: (i) a first plurality of feature values ​​obtained in a first form for the independent multiple features using biological samples of the respective training subject, the first form being one of transcriptomics, proteomics, or metabolomics of at least a first module of the plurality of modules; and (ii) an indication of the absence, presence, or stage of a first independent phenotype corresponding to the first module in the respective training subject. The method then includes obtaining a second training dataset in electronic form, wherein for each corresponding training object in a second plurality of training objects of a species, the second training dataset includes: (i) a second plurality of feature values ​​obtained, for independent plurality of features, in a second form identical to the first form of at least the first module, using biological samples of the corresponding training object from a second technical background other than the first technical background, and (ii) an indication of the absence, presence, or stage of a first independent phenotype in the corresponding training object. The method then includes co-normalizing the feature values ​​of features present in at least the first and second training datasets across at least the first and second training datasets to remove batch effects between datasets, thereby calculating co-normalized feature values ​​of at least the first module of the corresponding training object for each corresponding training object of the first plurality of training objects and for each corresponding training object of the second plurality of training objects. The method then includes training a master classifier on a composite training set to assess the clinical condition of test subjects, wherein for each corresponding training object in the first plurality of training objects and for each corresponding training object in the second plurality of training objects, the composite training set includes: (i) a summary of the co-normalized feature values ​​of the first module, and (ii) an indication of the absence, presence, or stage of a first independent phenotype in the corresponding training object.

[0012] On the other hand, this disclosure provides methods and systems for performing such methods to assess the clinical condition of test subjects of a species. The method includes obtaining a first training dataset electronically, wherein for each corresponding training subject in a first plurality of training subjects of a species, the first training dataset includes: (i) first plurality of feature values ​​obtained for a plurality of features using biological samples of the corresponding training subject, and (ii) an indication of the absence, presence, or stage of a first independent phenotype in the corresponding training subject. The first independent phenotype represents a disease condition, and a first subset of the first training dataset consists of subjects without a disease condition. The method then includes obtaining a second training dataset electronically, wherein for each corresponding training subject in a second plurality of training subjects of a species, the second training dataset includes: (i) second plurality of feature values ​​obtained for the plurality of features using biological samples of the corresponding training subject, and (ii) an indication of the absence, presence, or stage of the first independent phenotype in the corresponding training subject. A first subset of the second training dataset consists of subjects without a disease condition. The method then includes co-normalizing the feature values ​​of subsets of a plurality of features from at least the first and second training datasets to remove batch effects between datasets, wherein the feature subsets exist in at least the first and second training datasets. Common normalization involves estimating the inter-dataset batch effect between the first and second training datasets using only a first subset of the respective first and second training datasets. The inter-dataset batch effect includes additive and multiplicative components, and common normalization is performed by solving an ordinary least squares model for the feature values ​​across the first subset of the respective first and second training datasets and reducing the resulting parameters representing the additive and multiplicative components using an empirical Bayesian estimator, thereby using the resulting parameters to compute: for each corresponding training object in the first plurality of training objects and for each corresponding training object in the second plurality of training objects, the common normalized feature values ​​of a subset of multiple features. The method then includes training a master classifier against a composite training set to assess the clinical status of test subjects, for each corresponding training object in the first plurality of training objects and for each corresponding training object in the second plurality of training objects, the composite training set comprising: (i) the common normalized feature values ​​of a subset of multiple features, and (ii) an indication of the absence, presence, or stage of a first independent phenotype in the corresponding training object.

[0013] Each embodiment of the system, method, and apparatus within the scope of the appended claims has several aspects, and the desired properties described herein are not solely attributed to any single aspect. Without limiting the scope of the appended claims, some prominent features are described herein. After considering this discussion, and particularly after reading the section entitled "Detailed Description," one will understand how to use the features of the various embodiments.

[0014] By combining references All publications, patents and patent applications mentioned in this specification are incorporated herein by reference in their entirety as if each individual publication, patent or patent application were explicitly and individually identified by reference. Attached Figure Description

[0015] The embodiments disclosed herein are illustrated by way of example and not limitation in the accompanying drawings. Similar reference numerals denote corresponding parts throughout the various views of the drawings.

[0016] Figure 1A and 1B Example block diagrams of computing devices according to some embodiments of the present disclosure are shown together.

[0017] Figure 2A , 2B Figures 2C, 2D, 2E, 2F, 2G, 2H and 2I illustrate example flowcharts of methods for classifying objects according to some embodiments of the present disclosure, wherein optional steps are indicated by dashed boxes.

[0018] Figure 3 The diagram illustrates a network topology, where multiple modules at the bottom each contribute... Prior Given the geometric mean of the genes, all of them move in the same direction on average under the clinical condition of interest. According to some embodiments of this disclosure, the output at the top of the network is the clinical condition of interest (bacterial infection - I). bac Viral infection I vira No infection - I non ).

[0019] Figure 4 The diagram illustrates a network topology where each module uses a minispoke network (one of which is shown in more detail on the right side of the diagram). Individual biomarkers are aggregated by local networks (rather than by their geometric mean) and then fed into the main classification network.

[0020] Figure 5A and 5B The diagram illustrates an iterative COCONUT alignment according to an embodiment of this disclosure, where "Reference" is the microarray data and "Target" is the NanoString data. The graph shows the distribution of NanoString gene expression and microarray gene expression in healthy samples, targeting two genes (5A - HK3, 5B - IFI27) from a group of 29 genes. The microarray distribution is shown in three different iterations of the co-normalization-based alignment process. Dashed lines represent the distribution at intermediate iterations, and solid lines represent the distribution at program termination.

[0021] Figure 6A and 6BThe figure illustrates the distribution of co-normalized expression values ​​of selected genes (6A - fever markers) (6B - severity markers) in the training dataset used in the examples of this disclosure, across bacterial, viral, and non-infected training set samples.

[0022] Figure 7A and 7B The two-dimensional (7A) and three-dimensional (7B) t-SNE projections of the co-normalized expression values ​​of 29 genes in a training dataset according to an embodiment of the present disclosure are illustrated, respectively, wherein each object is labeled as bacteria, virus or uninfected.

[0023] Figure 8A and 8B Two-dimensional (8A) and three-dimensional (8B) principal component analysis plots of the co-normalized expression values ​​of 29 genes in a training dataset according to an embodiment of the present disclosure are illustrated, respectively, where each object is labeled as bacteria, virus, or uninfected.

[0024] Figure 9 The illustration shows a two-dimensional principal component analysis plot of co-normalized expression values ​​of 29 genes across a training dataset according to an embodiment of this disclosure, where each object is labeled by a source study.

[0025] Figure 10A and 10B The figures illustrate the validation performance bias analysis using six geometric mean scores instead of the direct expression values ​​of 29 genes according to the embodiments of this disclosure, whereby... Figure 10A The smaller image above shows logistic regression. Figure 10A The smaller image below shows XGBoost. Figure 10B The smaller image above shows a support vector machine with an RBF kernel. Figure 10B The smaller plot below shows a multilayer perceptron. The x-axis represents the difference between the average area under the pairwise ROC (APA) curves of the outer and inner folds for the top 10 models of each model type, sorted by cross-validation APA. Each point corresponds to one model. The y-axis corresponds to the outer fold APA. The vertical dashed line indicates that there is no difference in APA between the inner and outer loops.

[0026] Figure 11A and 11B The illustrations separately illustrate the validation performance bias analysis using the direct expression values ​​of 29 genes according to the embodiments of this disclosure, wherein... Figure 11A The smaller image above shows logistic regression. Figure 11A The smaller image below shows XGBoost. Figure 11B The smaller image above shows a support vector machine with an RBF kernel. Figure 11BThe smaller plot below shows a multilayer perceptron. The x-axis represents the difference between the average area under the pairwise ROC (APA) curves of the outer and inner folds for the top 10 models of each model type, sorted by cross-validation APA. Each point corresponds to one model. The y-axis corresponds to the outer fold APA. The vertical dashed line indicates that there is no difference in APA between the inner and outer loops.

[0027] Figure 12 The illustration shows pseudocode for an iterative application of the COCONUT algorithm according to some embodiments of this disclosure.

[0028] Figure 13 An example flowchart of a method for training a classifier to assess the clinical condition of a subject, according to some embodiments of the present disclosure, is illustrated.

[0029] Figure 14 An example flowchart illustrating a method for assessing the clinical condition of a subject according to some embodiments of this disclosure is shown. Detailed Implementation

[0030] Examples of the embodiments will now be shown in detail with reference to the accompanying drawings. Numerous specific details are set forth in the following detailed description in order to provide a thorough understanding of this disclosure. However, it will be apparent to those skilled in the art that this disclosure can be practiced without these specific details. In other instances, well-known methods, procedures, components, circuits, and networks have not been described in detail to avoid unnecessarily obscuring various aspects of the embodiments.

[0031] The implementations described herein provide various technical solutions for generating and using machine learning classifiers to diagnose, provide prognoses, or provide clinical predictions of medical conditions. In particular, the methods and systems provided herein facilitate the use of heterogeneous reservoirs of molecular (e.g., genomic, transcriptomic, proteomic, metabolomic) and / or clinical data with relevant clinical phenotypes to train machine learning classifiers with improved performance.

[0032] In some implementations, as described herein, the disclosed methods and systems achieve machine learning classifiers with improved performance by estimating inter-dataset batch effects between heterogeneous training datasets.

[0033] In some implementations, the systems and methods described herein utilize developed co-normalization methods to bring multiple discrete datasets into a single aggregated data frame. These methods improve classifier performance on some average function of overall aggregate accuracy, the accuracy of individual datasets within the aggregated frame, or both. Those skilled in the art will recognize that this capability requires improved co-normalization of heterogeneous datasets, which is not a feature of traditional omics-based data science pipelines.

[0034] In some implementations, the initial step in the classifier training method described herein is to priori identify the biomarkers to be trained on. Biomarkers of interest can be identified using a literature search or in a 'discovery' dataset, where statistical tests are used to select biomarkers relevant to the clinical condition of interest. In some implementations, the biomarkers of interest are then grouped according to the sign of their changing direction in clinical decisions of interest.

[0035] In some implementations, a subset of variables used to train these classifiers is selected from known molecular variables (e.g., genomic, transcriptomic, proteomic, metabolomic data) present in heterogeneous datasets. In some implementations, these variables are selected using statistical thresholding of differentially expressed data using tools such as microarray significance analysis (SAM), meta-analysis between datasets, or correlation with categories or other methods. In some implementations, the available data is expanded by designing novel features based on molecular map patterns. These novel features can be discovered using unsupervised analysis (e.g., denoising autoencoders) or supervised methods (e.g., path analysis using existing ontology or path databases (e.g., KEGG)).

[0036] In some implementations, the datasets used to train the classifier are obtained from public or private sources. In the public domain, repositories such as NBCI GEO or ArrayExpress (if transcriptome data is used) can be used. The datasets must contain at least one class of interest, and if a co-normalization function requiring healthy controls is used, they must have healthy controls. In some implementations, only data of a single biological type is collected (e.g., transcriptional data only, but not proteome data), but can come from a wide variety of technological backgrounds (e.g., both RNAseq and DNA microarrays).

[0037] In some implementations, the input data is stratified to ensure that each class is represented in approximately equal proportions across each input dataset. This step avoids the confusion caused by heterogeneous data sources when learning a single classifier across aggregate datasets. Stratification can be performed once, multiple times, or not at all.

[0038] In some implementations, when raw data from the original technical format is obtained, a standardized intra-dataset normalization procedure is performed to minimize the impact of different normalization methods on the final classifier. Data from the same type of technical platform is preferably normalized in the same way, typically using general procedures such as background correction, logarithmic normalization, etc. 2 Transformation and quantile normalization. Platform-specific normalization procedures are also common (e.g., gcRMA for the Affymetrix platform for positive-match controls). The result is a single file or other data structure for each dataset.

[0039] In some implementations, common normalization is then performed in two steps: optional inter-platform common variable mapping, followed by the necessary common normalization.

[0040] Inter-platform common variable mapping is necessary in cases where the platforms involving datasets do not follow the same naming conventions and / or measure the same target as multiple variants (e.g., many RNA microarrays have degenerate probes for a single gene). A common reference (e.g., mapped to RefSeq genes) should be chosen, and variables should be relabeled (either in a single case or in a summary case; e.g., by taking a measure of central tendency, such as median, mean, etc., or in a fixed-effects meta-analysis of degenerate probes for the same gene).

[0041] Common normalization is necessary because, after identifying variables with common names across the datasets, these variables often have significantly different distributions across the datasets. Therefore, these values ​​are transformed to match the same distribution (e.g., mean and variance) across the datasets. Various methods can be used for common normalization, such as COCONUT (Sweeney...) et al. , 2016, Sci Transl Med 8(346), pp. 346ra91; and Abouelhoda et al (BMC Bioinformatics 9, p. 476, 2008), bit normalization, ComBat, pooled RMA, pooled gcRMA, or normalization of invariant genes (e.g., housekeeping genes), etc.

[0042] In some implementations, data co-normalized using the improved methods described herein undergoes machine learning to train a master classifier for a category of clinical condition of interest, such as disease diagnosis or prognosis classification. In non-limiting instances, this can utilize linear regression, penalized linear regression, support vector machines, tree-based methods (e.g., random forests or decision trees), ensemble methods (e.g., AdaBoost, XGBoost) or ensembles of other weak or strong classifiers, neural network methods (e.g., multilayer perceptrons), or other methods or variations thereof. In some implementations, the master classifier can be learned directly from selected variables, engineered features, or both. In some implementations, the master classifier is an ensemble of classifiers.

[0043] In some implementations, these methods and systems are further extended by generating new samples from the aggregated data using generative functions. In some implementations, this involves adding random noise to each sample. In some implementations, this involves more sophisticated generative models, such as Boltzmann machines, deep belief networks, generatively unfavorable networks, adversarial autoencoders, other methods, or variations thereof.

[0044] In some implementations, methods and systems for classifier development include cross-validation, model selection, model evaluation, and calibration. Initial cross-validation estimates the performance of a fixed classifier. Model selection uses hyperparameter search and cross-validation to identify the most accurate classifier. Model evaluation assesses the performance of the selected model on independent data and can be performed using methods such as leave-one-out-of-one-data (LODO) cross-validation, nested cross-validation, or bootstrap-corrected performance evaluation. Calibration adjusts the classifier scores based on phenotypic distributions observed in clinical practice, aiming to convert the scores into intuitive, human-interpretable values. It can be performed using methods such as the Hosmer-Lemeshow test and calibration slope.

[0045] In some implementations, neural network classifiers, such as multilayer perceptrons, are used for supervised classification of outcomes of interest (e.g., the presence of infection) in co-normalized data. Variables that move together on average under the clinical condition of interest are known to be grouped into 'modules,' and the neural network architectures that interpret these grouping modules are learned above.

[0046] In some implementations, the 'module' is constructed in one of two ways. In the first way, biomarkers within the module are measured by taking their central tendency, such as a geometric mean, and feeding it into the main taxonomic grouping (e.g., as in...). Figure 3 (As shown in the diagram). In another embodiment, a 'spoke' network is constructed, wherein the inputs in the modules are biomarkers, and they are interpreted via component classifiers fed into the main classifier (e.g., as shown in the diagram). Figure 4 (as shown in the image).

[0047] definition The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in the specification and appended claims, the singular forms “a,” “an,” and “described” are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and covers any and all possible combinations of one or more of the associated listed items. It should be further understood that, when used in this specification, the term “comprises and / or comprising” specifies the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0048] As used herein, depending on the context, the term “if” can be interpreted as “when” or “in response to determination” or “in response to detection”. Similarly, depending on the context, the phrase “if determination” or “if [the stated condition or event] is detected” can be interpreted as “when determination” or “in response to determination” or “when [the stated condition or event] is detected” or “in response to detection”.

[0049] It will also be understood that while the terms “first,” “second,” etc., may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used only to distinguish one element from another. For example, without departing from the scope of this disclosure, a first subject may be referred to as a second subject, and similarly, a second subject may be referred to as a first subject. Although both the first and second objects are objects, they are not the same object. Furthermore, the terms “object,” “user,” and “patient” are used interchangeably herein.

[0050] As disclosed herein, the terms “nucleic acid” and “nucleic acid molecule” are used interchangeably. The term refers to nucleic acids in any constituent form, such as deoxyribonucleic acid (DNA), ribonucleic acid (RNA), and / or DNA or RNA analogs (e.g., containing base analogs, sugar analogs, and / or non-natural backbones, etc.), all of which can be single-stranded or double-stranded. Unless otherwise limited, nucleic acids may contain known natural nucleotide analogs, some of which can function in a manner similar to naturally occurring nucleotides. Nucleic acids can be any form used to carry out the processes described herein (e.g., ，Nucleic acids can be linear, circular, supercoiled, single-stranded, double-stranded, etc. In some embodiments, nucleic acids may be derived from a single chromosome or a fragment thereof (e.g., a nucleic acid sample may be derived from a chromosome sample obtained from a diploid organism). In some embodiments, nucleic acids include nucleosomes, fragments or portions of nucleosomes, or nucleosome-like structures. Nucleic acids sometimes contain proteins (e.g., histones, DNA-binding proteins, etc.). Nucleic acids analyzed by the methods described herein are sometimes substantially isolated and substantially unassociated with proteins or other molecules. Nucleic acids also include derivatives, variants, and analogs (“sense” or “antisense”, “positive” or “negative” strand, “positive” or “reverse” reading frame) and double-stranded polynucleotides from single-stranded DNA that has been synthesized, replicated, or amplified. Deoxyribonucleotides include deoxyadenosine, deoxycytidine, deoxyguanosine, and deoxythymidine. Nucleic acids may be prepared using nucleic acids obtained from a subject as templates.

[0051] As disclosed herein, the term "subject" refers to any living or non-living organism, including but not limited to humans (e.g., ， The object can be any human or non-human animal, including but not limited to mammals, reptiles, birds, amphibians, fish, ungulates, ruminants, bovines (e.g., cattle), equines (e.g., horses), goats and sheep (e.g., sheep, goats), pigs (e.g., pigs), camels (e.g., camels, llamas, alpacas), monkeys, apes (e.g., gorillas, chimpanzees), bears (e.g., bears), poultry, dogs, cats, rodents, fish, dolphins, whales, and sharks. In some embodiments, the object is male or female at any stage (e.g., men, women, fetuses, pregnant women, children, etc.). ， (Male, female, or child).

[0052] As used herein, the terms “control,” “control sample,” “reference,” “reference sample,” “normal,” and “normal sample” describe samples from subjects that do not have a specific condition or are otherwise healthy. In one example, the methods disclosed herein can be applied to subjects with tumors, where the reference sample is a sample taken from healthy tissue of the subject. The reference sample can be obtained from the subject or from a database. The reference can be, for example, a reference genome, which is used to map sequence reads obtained from sequencing samples of the subject.

[0053] As used herein, the terms “sequencing”, “sequence determination”, etc., generally refer to any and all biochemical methods that can be used to determine the order of biological macromolecules such as nucleic acids or proteins. For example, sequencing data may include all or part of the nucleotide bases in nucleic acid molecules, such as mRNA transcripts or genomic loci.

[0054] Exemplary System Implementation Having provided an overview of some aspects of this disclosure and some definitions used herein, details of an exemplary system are now described in conjunction with Figure 1. Figure 1 is a block diagram illustrating a system 100 according to some embodiments. In some embodiments, device 100 includes one or more processing units, CPU 102 (also referred to as a processor), one or more network interfaces 104, user interface 106, non-persistent memory 111, persistent memory 112, and one or more communication buses 114 for interconnecting these components. The one or more communication buses 114 may optionally include a circuit system (sometimes referred to as a chipset) interconnecting and controlling communication between system components. Non-persistent memory 111 typically includes high-speed random access memory such as DRAM, SRAM, DDR RAM, ROM, EEPROM, flash memory, while persistent memory 112 typically includes CD-ROM, digital versatile disc (DVD) or other optical storage devices, magnetic tape cassettes, magnetic tape, disk storage devices or other magnetic storage devices, optical disc storage devices, flash memory devices or other non-volatile solid-state storage devices. Persistent memory 112 may optionally include one or more storage devices arranged remotely from CPU 102. One or more non-volatile storage devices within the persistent memory 112 and the non-persistent memory 112 include non-transitory computer-readable storage media. In some embodiments, the non-persistent memory 111 or (alternatively) the non-transitory computer-readable storage media (sometimes in combination with the persistent memory 112) stores programs, modules, and data structures or subsets thereof: • Operating system 116, which includes programs for handling various basic system services and for performing hardware-related tasks; • Network communication module (or instructions) 118, which is used to connect the visualization system 100 to other devices or to a communication network; • Variable selection module 120 is used to identify information features of phenotypes of interest; • Raw data normalization module 122, used to normalize the raw feature data 136 within each raw training dataset 132; • Data co-normalization module 124, which is used to co-normalize feature data across heterogeneous training datasets, such as internal normalized data construction 138, such as normalized feature data 142; • Classifier training module 126, used to train a machine learning classifier based on co-normalized feature data 148 across heterogeneous datasets; • Training dataset storage 130, which is used to store one or more data structures, such as original data construction 132, internal normalized data construction 138 and / or co-normalized data construction 144 for one or more samples of training objects, each such data construction including multiple feature values ​​for each corresponding training object among multiple training objects, such as original feature values ​​136, internal normalized feature values ​​142 and / or co-normalized feature values ​​148. • A data module set memory 150 is used to store one or more modules 152 for training a classifier, each such corresponding module 150 including (i) identifiers of multiple independent differential modulating features 154, (ii) corresponding summarization algorithms or component classifiers 156, and (iii) independent phenotypes 157 related to the clinical condition under study (e.g., the clinical condition itself or a deterministic or clinical condition-related phenotype); and • Test dataset storage 160, which stores one or more data constructs 162 for one or more samples of test object 164, each such data construct including multiple feature values ​​166.

[0055] In some embodiments, one or more of the aforementioned elements are stored in one or more of the previously mentioned storage devices and correspond to sets of instructions for performing the aforementioned functions. The aforementioned modules, data, or programs (e.g., instruction sets) do not need to be implemented as separate software programs, processes, datasets, or modules, and therefore, subsets of these modules and data can be combined or otherwise rearranged in various embodiments. In some embodiments, non-persistent memory 111 optionally stores a subset of the aforementioned modules and data structures. Furthermore, in some embodiments, the memory stores additional modules and data structures not described above. In some embodiments, one or more of the aforementioned elements are stored in a computer system outside the computer system of the visualization system 100, which is addressable by the visualization system 100, such that the visualization system 100 can retrieve all or part of such data when needed.

[0056] Although Figure 1 depicts "System 100", this figure is intended more as a functional description of the various features that may exist in a computer system than as a structural schematic diagram of the implementation described herein. In practice, and as those skilled in the art will recognize, items shown individually may be combined, and some items may be separated. Furthermore, although Figure 1 depicts some data and modules in non-persistent memory 111, some or all of these data and modules may reside in persistent memory 112.

[0057] Exemplary method implementation Although the system according to this disclosure has been disclosed with reference to FIG1, the method according to this disclosure is now described in detail with reference to FIG2.

[0058] refer to Figure 2A In some embodiments, as indicated by boxes 202-214, a method for assessing the clinical condition of test subjects of a species using prior grouping of features is provided at a computer system (e.g., system 100 of Figure 1), having one or more processors 102 and memory 111 / 112 storing one or more programs, such as variable selection module 120, which are executed by the one or more processors. The prior grouping of features comprises a plurality of modules 152. Each corresponding module 152 comprises a plurality of independent features 154, each corresponding to a feature value associated with the absence, presence, or stage of an independent phenotype 157 associated with the clinical condition. For example, Table 1a provides non-limiting example definitions and compositions of six sepsis-related modules (genomes), each module associated with the absence, presence, or stage of an independent phenotype 157 associated with sepsis. Modules 152-1 and 152-2 of Table 1a target genes with elevated (module 152-1) and decreased (module 152-2) expression, respectively, in strict viral infection. Modules 152-3 and 152-4 in Table 1a represent genes with increased (module 152-3) and decreased (module 152-4) expression in patients with sepsis compared to those with aseptic inflammation, respectively. Modules 152-5 and 152-6 represent genes with increased (module 152-5) and decreased (module 152-6) expression in patients who died within 30 days of admission, respectively.

[0059] Table 1a: Definition and composition of sepsis-related modules Referring to block 204, in some embodiments, the object is a human or mammal. In some embodiments, the object is any living or non-living organism, including but not limited to humans (e.g., humans). ， The target species includes men, women, fetuses, pregnant women, children, non-human animals, plants, bacteria, fungi, or protozoa. In some embodiments, the target species include mammals, reptiles, birds, amphibians, fish, ungulates, ruminants, bovids (e.g., cattle), equines (e.g., horses), goats and sheep (e.g., sheep, goats), pigs (e.g., pigs), camels (e.g., camels, llamas, alpacas), monkeys, apes (e.g., gorillas, chimpanzees), bears (e.g., bears), poultry, dogs, cats, rodents, fish, dolphins, whales, and sharks. In some embodiments, the target species includes males or females at any stage (e.g., men, women, fetuses, pregnant women, children, etc.), non-human animals, plants, bacteria, fungi, or protozoa. ， (Male, female, or child).

[0060] Referring to box 206, in some embodiments, the clinical condition is a binary clinical condition (e.g., having sepsis versus not having sepsis, having cancer versus not having cancer, etc.). Referring to box 208, in some embodiments, the clinical condition is a multi-category clinical condition. For example, referring to box 210, in some embodiments, the clinical condition consists of three categories: (i) severe bacterial infection, (ii) severe viral infection, and (iii) non-infectious inflammation.

[0061] Referring to block 212, in some embodiments, the plurality of modules 152 comprises at least three modules or at least six modules. Table 1a above provides an example where the plurality of modules 152 consists of six modules. In some embodiments, the plurality of modules 152 comprises three to one hundred modules. In some embodiments, the plurality of modules 152 consists of two modules.

[0062] Furthermore, referring to block 214, in some embodiments, each of the multiple features 154 of each module 152 in the plurality of modules includes at least three features or at least five features. Table 1a above provides an example in which the plurality of modules 152 consist of six modules. In some embodiments, the plurality of modules 152 includes three to one hundred modules. In some embodiments, the plurality of modules 152 consists of two modules. Furthermore, it is not required that each module include the same number of features. The examples in Table 1a above demonstrate this. Thus, for example, in some embodiments, one module 152 may have two features 154, while another module may have more than fifty features. In some embodiments, each module 152 has two to fifty features 154. In some embodiments, each module 152 has three to one hundred features. In some embodiments, each module 152 has four to two hundred features. In some embodiments, the features 154 in each module 152 are unique. That is, any given feature appears only in one of the modules 152. In other implementations, it is not required that the features in each module 152 be unique; that is, in such implementations, a given feature 154 may be in more than one module.

[0063] See Figure 2B Box 216, to obtain the first training dataset (e.g., Figure 1AThe original data construction 132-1). For each corresponding training object 134 in the first plurality of training objects of a species, the first training dataset includes: (i) first plurality of feature values ​​136 obtained in a first form for independent plurality of features using biological samples of the corresponding training object, through a first technical background, the first form being one of the transcriptomics, proteomics, or metabolomics of at least a first module 152 of a plurality of modules; and (ii) an indication of the absence, presence, or stage of a first independent phenotype 157 corresponding to the first module in the corresponding training object. In practice, because this is a training dataset, the dataset will provide an indication of the clinical condition of each object. However, in some embodiments, the first independent phenotype and the clinical condition are the same. In their different embodiments, the training set provides both the first independent phenotype and the clinical condition. For example, if the first module is module 152-1 of Table 1a above, the first dataset will provide for each training object in the first dataset: (i) measured expression values ​​of genes IFI27, JUP and LAX1, which are obtained using biological samples of the corresponding training object through the first technical background, (ii) an indication of whether the object has a fever, and (iii) whether the object has sepsis.

[0064] In some implementations, each module 158 is uniquely associated with the absence, presence, or stage of an independent phenotype related to a clinical condition. However, for each training subject, the first training dataset provides only an indication of the clinical condition itself, not the absence, presence, or stage of the independent phenotype 157 for each corresponding module. For example, in the case of Table 1a, in some implementations, the first training dataset includes an indication of the absence, presence, or stage of a clinical condition (sepsis) but does not indicate whether each training subject has the phenotype of fever. That is, in some implementations, this disclosure relies on prior work that has identified which features are up- or down-regulated for a given phenotype (e.g., fever), and therefore does not require an indication of whether each training subject in the training dataset has the phenotype of a module. Instead of providing a phenotype corresponding to a module, an indication of the absence, presence, or stage of a clinical condition is provided for the training subject.

[0065] In some implementations, the first training dataset only provides the presence or absence of a clinical condition for each training subject. That is, in such implementations, the stage of clinical condition is not provided.

[0066] refer to Figure 2BIn some embodiments, each corresponding feature in the first module corresponds to a biomarker associated with a first independent phenotype that is statistically significantly more abundant in subjects exhibiting the first independent phenotype than in subjects that do not exhibit the independent phenotype in the subject group of the species. The subject group of the species need not be the subjects of the first dataset. The subject group of the species is any group of subjects that meets the selection criteria and includes both subjects with and without clinical conditions. The selection criteria for non-limiting instances of the group in cases of sepsis are: 1) the presence and type of infection as determined by a physician (e.g., a severe bacterial infection, a severe viral infection, or a non-infectious inflammation); 2) a characteristic value characteristic in said plurality of modules; 3) being over 18 years of age; 4) being seen in a hospital setting (e.g., the emergency department, intensive care unit); 5) a community- or hospital-acquired infection; and 6) a blood sample was collected within 24 hours of the initial suspicion of infection and / or sepsis. In some such implementations, the determination of whether a biomarker is “statistically significantly more abundant” is assessed by applying a standard t-test, Welch t-test, Wilcoxon test, or permutation test to the abundance of the biomarker, as measured in subjects exhibiting a first independent phenotype (Group 1) and in subjects not exhibiting the first independent phenotype to reach a p-value (Group 2). In some such implementations, a biomarker is statistically significantly more abundant when the p-value in such a test is 0.05 or less, 0.005 or less, or 0.001 or less. In some such implementations, a biomarker is statistically significantly more abundant when the p-value in such a test is 0.05 or less, 0.005 or less, or 0.001 or less, adjusted for multiple tests using a false discovery rate procedure such as Benjamini-Hochberg or Benjamini-Yekutieli. See, for example, Benjamini and Hochberg, Journal of the Royal Statistical Society, Series B 57, pp. 289–300; and Benjamini and Yekutieli, 2005, Journal of American Statistical Association 100(469), pp. 71–80, each of which is incorporated herein by reference. In some implementations, biomarkers are considered statistically significantly more abundant through fixed-effects or random-effects meta-analysis of multiple datasets (groups or training datasets). See, for example, Sianphoe et al . ,2019, BMC Bioinformatics 20:18, incorporated herein by reference.

[0067] In some implementations, each module 152 is uniquely associated with the absence, presence, or stage of an independent phenotype 157 related to a clinical condition. However, for each training subject in the first training set, the first training dataset only provides indications of the absence, presence, or stage of the clinical condition itself and the absence, presence, or stage of independent phenotypes for some, but not all, of the multiple modules. For example, in the case of Table 1a, in some implementations, the first training dataset includes indications of the absence, presence, or stage of the clinical condition / phenotype “sepsis” and the absence, presence, or stage of the phenotype “severity,” but does not indicate whether each training subject has a fever.

[0068] refer to Figure 2BIn some embodiments, each corresponding feature in the first module corresponds to a biomarker associated with a first independent phenotype 157 that is statistically significantly less abundant in subjects exhibiting the first independent phenotype compared to subjects that do not exhibit the independent phenotype in the subject group of the species. In some embodiments, the determination of whether a biomarker is “statistically significantly less abundant” is assessed by applying a standard t-test, Welch t-test, Wilcoxon test, or permutation test to the abundance of the biomarker, as measured in subjects in the group exhibiting the first independent phenotype (Group 1) and in subjects in the group not exhibiting the first independent phenotype to reach a p-value (Group 2). In some such embodiments, a biomarker is statistically significantly less abundant when the p-value in such a test is 0.05 or less, 0.005 or less, or 0.001 or less. In some implementations of this kind, a biomarker is considered statistically significantly less abundant when the p-value in such tests is 0.05 or less, 0.005 or less, or 0.001 or less, adjusted for multiple tests using false discovery rate procedures such as Benjamini-Hochberg or Benjamini-Yekutieli. See, for example, Benjamini and Hochberg, Journal of the Royal Statistical Society, Series B 57, pp. 289-300; and Benjamini and Yekutieli, 2005, Journal of American Statistical Association 100(469), pp. 71-80, each of which is incorporated herein by reference. In some implementations, a biomarker is considered statistically significantly less abundant through a fixed-effects or random-effects meta-analysis of multiple datasets (group or training datasets). See, for example, Sianphoe et al . , 2019, BMC Bioinformatics 20:18, incorporated herein by reference.

[0069] refer to Figure 2BIn some embodiments, each corresponding feature in the first module is associated with a first independent phenotype 157 because its feature value is statistically significantly larger in objects exhibiting the first independent phenotype compared to objects that do not exhibit the independent phenotype in the species' object group. In some embodiments, the determination of whether a feature is “statistically significantly richer” is assessed by applying a standard t-test, Welch t-test, Wilcoxon test, or permutation test to the abundance of the feature, as measured in objects in the group exhibiting the first independent phenotype (Group 1) and in objects in the group not exhibiting the first independent phenotype to reach a p-value (Group 2). In some such embodiments, the feature value is statistically significantly larger when the p-value in such tests is 0.05 or less, 0.005 or less, or 0.001 or less. In some implementations of this kind, a feature is considered statistically significant (richer) when the p-value in such tests is 0.05 or less, 0.005 or less, or 0.001 or less, and this is adjusted for multiple tests using false discovery rate procedures such as Benjamini-Hochberg or Benjamini-Yekutieli. See, for example, Benjamini and Hochberg, Journal of the Royal Statistical Society, Series B 57, pp. 289-300; and Benjamini and Yekutieli, 2005, Journal of American Statistical Association 100(469), pp. 71-80, each of which is incorporated herein by reference. In some implementations, a feature is considered statistically significant by a fixed-effects or random-effects meta-analysis of multiple datasets (group or training dataset). See, for example, Sianphoe et al . , 2019, BMC Bioinformatics 20:18, incorporated herein by reference.

[0070] refer to Figure 2BIn some embodiments, each corresponding feature in the first module is associated with a first independent phenotype 157 because its feature value is statistically significantly less in objects exhibiting the first independent phenotype compared to objects that do not exhibit the independent phenotype in the species' object group. In some embodiments, the determination of whether a feature is “statistically significantly less” is assessed by applying a standard t-test, Welch t-test, Wilcoxon test, or permutation test to the abundance of the feature, as measured in objects in the group exhibiting the first independent phenotype (Group 1) and in objects in the group not exhibiting the first independent phenotype to reach a p-value (Group 2). In some such embodiments, a feature is statistically significantly less when the p-value in such tests is 0.05 or less, 0.005 or less, or 0.001 or less. In some implementations of this kind, a feature is considered statistically significantly less when the p-value in such tests is 0.05 or less, 0.005 or less, or 0.001 or less. This is adjusted for multiple tests using false discovery rate procedures such as Benjamini-Hochberg or Benjamini-Yekutieli. See, for example, Benjamini and Hochberg, Journal of the Royal Statistical Society, Series B 57, pp. 289-300; and Benjamini and Yekutieli, 2005, Journal of American Statistical Association 100(469), pp. 71-80, each of which is incorporated herein by reference. In some implementations, a feature is considered statistically significantly less by a fixed-effects or random-effects meta-analysis of multiple datasets (group or training dataset). See, for example, Sianphoe et al . , 2019, BMC Bioinformatics 20:18, incorporated herein by reference.

[0071] refer to Figure 2C In some embodiments, as indicated by block 228, the characteristic value of the first feature in module 152 of the plurality of modules is determined by physical measurement of the corresponding component in a biological sample of the reference object. Referring to block 230, examples of components include, but are not limited to, components (e.g., nucleic acids, proteins, or metabolites).

[0072] refer to Figure 2CIn some embodiments, the feature value of the first feature in module 152 of the plurality of modules is a linear or nonlinear combination of the feature values ​​of each of the corresponding components in a set of components obtained by physically measuring each of the corresponding components (e.g., nucleic acids, proteins, or their metabolites) in a biological sample of a reference object.

[0073] Regarding box 216, note that for multiple independent features, the first training set is obtained using biological samples of the corresponding training objects in a first form of transcriptomics, proteomics, or metabolomics. Referring to box 234, in some embodiments, the first form is transcriptomics. Referring to box 236, in some embodiments, the first form is proteomics.

[0074] Regarding box 216, note that for each of the first plurality of training objects, the first training set includes the first plurality of feature values ​​obtained through the first technical background. Referring to box 238, in some embodiments, the first technical background is a DNA microarray, MMChip, protein microarray, peptide microarray, tissue microarray, cell microarray, compound microarray, antibody microarray, glycan array, or reversed-phase protein lysate microarray.

[0075] In some implementations, the biological samples collected from each subject are the subject's blood, whole blood, plasma, serum, urine, cerebrospinal fluid, feces, saliva, sweat, tears, pleural effusion, pericardial fluid, or peritoneal fluid. In some implementations, the biological samples are specific tissues of the subject. In some implementations, the biological samples are biopsies of specific tissues or organs of the subject (e.g., breast, lung, prostate, rectum, uterus, pancreas, esophagus, ovary, bladder, etc.).

[0076] In some implementations, the feature is a nucleic acid abundance value corresponding to the nucleic acid of a species gene obtained from the sequencing read, which in turn is derived from nucleic acids in the biological sample and represents the abundance of such nucleic acids and the genes they represent in the biological sample. Any form of sequencing can be used to obtain nucleic acid sequence reads from biological samples, including but not limited to high-throughput sequencing systems such as the Roche 454 platform, the Applied Biosystems SOLID platform, Helicos True Single Molecule DNA sequencing technology, the hybridization sequencing platform of Affymetrix Inc., the single-molecule real-time (SMRT) technology of Pacific Biosciences, the synthetic sequencing platforms of 454 Life Sciences, Illumina / Solexa, and Helicos Biosciences, and the ligation sequencing platform of Applied Biosystems. Life Technologies' ION TORRENT technology and nanopore sequencing can also be used to obtain cell-free nucleic acid sequence reads from biological samples.140

[0077] In some implementations, synthetic sequencing and reversible terminator-based sequencing (e.g., Illumina's Genome Analyzer; Genome Analyzer II; HISEQ 2000; HISEQ 2500 (Illumina, San Diego, California)) are used to obtain sequence reads of nucleic acids from biological samples. In some such implementations, millions of cell-free nucleic acid (e.g., DNA) fragments are sequenced in parallel. In one example of this type of sequencing technology, a flow cell is used, containing an optically clear slide with eight individual lanes on its surface to which oligonucleotide anchors (e.g., adapter primers) are bound. The flow cell is typically a solid support configured to retain and / or allow reagent solutions to pass through the bound analytes in an orderly manner. In some cases, the flow cell is planar, optically clear, typically millimeter or sub-millimeter in size, and typically has channels or lanes in which analyte / reagent interactions occur. In some implementations, the cell-free nucleic acid sample may include a signal or tag to facilitate detection. In some such implementations, obtaining nucleic acid sequence reads from biological samples includes obtaining quantitative information about signals or tags through a variety of techniques, such as flow cytometry, quantitative polymerase chain reaction (qPCR), gel electrophoresis, gene chip analysis, microarrays, mass spectrometry, cell fluorescence analysis, fluorescence microscopy, confocal laser scanning microscopy, laser scanning cell counting, affinity chromatography, manual batch separation, electric field suspension, sequencing, and combinations thereof.

[0078] Referring to box 240, in some implementations, the first independent phenotype of a module is the same as the clinical condition. This is illustrated for modules 152-3 and 152-4 in Table 1a, where the clinical condition is sepsis, the first independent phenotype of module 152-3 is "sepsis resolution," and the first independent phenotype of module 152-4 is sepsis resolution. Therefore, for modules 152-3 and 152-4, all that is needed in the training set (except for feature abundance) is that each training subject is labeled as having sepsis.

[0079] Referring to block 242, in some embodiments, a second training dataset is obtained. For each of the second plurality of training objects of the species, the second training dataset includes: (i) a second plurality of feature values ​​obtained, for independent plurality of features, in a second form identical to the first form of at least the first module, using biological samples of the respective training object from a second technical background other than the first technical background; and (ii) an indication of the absence, presence, or stage of a first independent phenotype in the respective training object.

[0080] Referring to box 244, in some embodiments, the first technical background (through which the first training set is obtained) is RNAseq, while the second technical background (through which the second training set is obtained) is DNA microarray.

[0081] In some implementations, the first technical background is a first form of microarray experiment selected from cDNA microarrays, oligonucleotide microarrays, BAC microarrays and single nucleotide polymorphism (SNP) microarrays, and the second technical background is a second form of microarray experiment selected from cDNA microarrays, oligonucleotide microarrays, BAC microarrays and SNP microarrays, which is different from the first form of microarray experiment.

[0082] In some implementations, the first technical context is nucleic acid sequencing using sequencing technology from a first manufacturer, and the second technical context is nucleic acid sequencing using sequencing technology from a second manufacturer (e.g., Illumina bead arrays versus Affymetrix or Agilent microarrays).

[0083] In some implementation schemes, the first technical background is to use a first sequencer to perform nucleic acid sequencing at a first sequencing depth, and the second technical background is to use a second sequencer to perform nucleic acid sequencing at a second sequencing depth, wherein the first sequencing depth is not the second sequencing depth and the first sequencing instrument and the second sequencing instrument are of the same brand and model, but the first and second instruments are different instruments.

[0084] In some implementations, the first technical background is a first type of nucleic acid sequencing (e.g., microarray-based sequencing) and the second technical background is a second type of nucleic acid sequencing other than the first type (e.g., next-generation sequencing).

[0085] In some implementations, the first technical background is paired-end nucleic acid sequencing, while the second technical background is single-read nucleic acid sequencing.

[0086] The above are non-limiting examples of different technical backgrounds. Generally speaking, two technical backgrounds are different when feature abundance data are captured under different technical conditions, such as different machines, different methods, or different technical conditions, such as different reagents, or different technical parameters (such as different coverage in the case of nucleic acid sequencing).

[0087] Referring to box 248, in some implementations, each corresponding biological sample in the first training dataset and the second training dataset belongs to a designated tissue or designated organ of the corresponding training object. For example, in some implementations, each biological sample is a blood sample. In another example, each biological sample is a breast biopsy, lung biopsy, prostate biopsy, rectal biopsy, uterine biopsy, pancreatic biopsy, esophageal biopsy, ovarian biopsy, or bladder biopsy.

[0088] refer to Figure 2D In some embodiments, box 252 indicates that a first normalization algorithm is performed on the first training dataset based on each corresponding distribution of the feature values ​​of the corresponding features in the first training dataset. Further, a second normalization algorithm is performed on the second training dataset based on each corresponding distribution of the feature values ​​of the corresponding features in the second training dataset. (See reference...) Figure 2D In some implementations, the first or second normalization algorithm is a robust multi-array averaging algorithm, a GeneChip RMA algorithm, or a normal exponential convolution algorithm for background correction, followed by a quantile normalization algorithm.

[0089] In some implementations, this normalization is not performed in the disclosed method. As a non-limiting example, in such an implementation, normalization of box 252 is not performed because the dataset has already been normalized. As another non-limiting example, in some implementations, normalization of box 252 is not performed because such normalization is determined to be unnecessary.

[0090] Referring to box 256, the feature values ​​of features present in at least the first and second training datasets are co-normalized across at least the first and second training datasets to remove batch effects between datasets, thereby computing co-normalized feature values ​​for at least a first module of a respective training object for each of the first plurality of training objects and for each of the second plurality of training objects. In some such embodiments, such normalization provides co-normalized feature values ​​for each of the plurality of modules for the respective training object.

[0091] Referring to box 258, in some embodiments, the first independent phenotype (of the first module) represents a disease status. Furthermore, a first subset of the first training dataset consists of objects without a disease status, and a first subset of the second training dataset consists of objects without a disease status. Additionally, the co-normalization of the feature values ​​present in at least the first and second training datasets involves estimating the inter-dataset batch effect between the first and second training datasets using only the first subsets of the respective first and second training datasets. Referring to box 260, in some such embodiments, the inter-dataset batch effect includes additive and multiplicative components, and co-normalization is performed by solving a common least squares model using feature values ​​across the first subsets of the respective first and second training datasets and reducing the resulting parameters representing the additive and multiplicative components using an empirical Bayesian estimator. See, for example, Sweeney. et al. , 2016, Sci Transl Med 8(346), pp. 346ra91, which is incorporated herein by reference.

[0092] Referring to box 264, in some embodiments, the co-normalization of eigenvalues ​​present in at least the first and second training datasets across at least the first and second training datasets includes estimating the inter-dataset batch effect between the first and second training datasets. Referring to box 266, in some embodiments, the inter-dataset batch effect includes additive and multiplicative components, and co-normalization is performed by solving an ordinary least squares model for eigenvalues ​​across a first subset of each of the first and second training datasets and reducing the resulting parameters representing the additive and multiplicative components using an empirical Bayesian estimator. See, for example, Sweeney. et al. , 2016, Sci Transl Med 8(346), pp. 346ra91, which is incorporated herein by reference.

[0093] refer to Figure 2E In some embodiments, according to box 266, the co-normalization of feature values ​​present in at least the first and second training datasets across at least the first and second training datasets includes utilizing invariant features, quantile normalization, or rank normalization. See Qiu et al., 2013, BMC Bioinformatics 14, p. 124; and Hendrik et al. , 2007, PLoS One 2(9), p. e898, each of which is incorporated herein by reference.

[0094] refer to Figure 2F Box 258, in some embodiments, each feature in the first and second datasets is a nucleic acid. The first technical background is a first form of microarray experiment selected from cDNA microarrays, oligonucleotide microarrays, BAC microarrays, and single nucleotide polymorphism (SNP) microarrays. The second technical background is a second form of microarray experiment selected from cDNA microarrays, oligonucleotide microarrays, BAC microarrays, and SNP microarrays, different from the first form of microarray experiment. See, for example, Bumgarner, 2013. Current protocols in molecular biology, Chapter 22, the references of which are incorporated herein by reference. In some such implementations, common normalization is robust multiarray average (RMA), GeneChip robust multiarray average (GC-RMA), MAS5, probe log intensity error (Plier), dChip, or chip calibration. See, for example, Irizarry, 2003, Biostatistics 4(2), pp. 249-264; Welsh et al References are included in this paper by way of citation: 2013, BMC Bioinformatics 14, p. 153; Therneau and Ballman, 2008, Cancer Inform 6, pp. 423-431; and Oberg, 2006, Bioinformatics 22, pp. 2381-2387.

[0095] refer to Figure 2F The method continues to train the master classifier on a composite training set to assess the clinical status of the test subjects. For each corresponding training subject in the first plurality of training subjects and each corresponding training subject in the second plurality of training subjects, the composite training set includes: (i) a summary of the co-normalized feature values ​​of the first module, and (ii) an indication of the absence, presence, or stage of the first independent phenotype in the corresponding training subject.

[0096] Referring to box 270, in some such embodiments, for each of the first and second plurality of training objects, the sum of the common normalized eigenvalues ​​of the first module is a measure of the central tendency (e.g., arithmetic mean, geometric mean, weighted mean, median range, midhinge, three-mean, Winsorized mean, median, or mode) of the common normalized eigenvalues ​​of the first module in the biological samples obtained from the respective training object. For example, in some such embodiments, in the biological samples obtained from the respective training objects, for each of the first and second plurality of training objects, the sum of the common normalized eigenvalues ​​of the first module is a measure of the central tendency (e.g., arithmetic mean, geometric mean, weighted mean, median range, midhinge, three-mean, Winsorized mean, median, or mode) of the common normalized eigenvalues ​​of each of the plurality of modules. This is in Figure 3 This is explained in the text, where each module f up f dn m up m dn s up and s dn Each of the training objects provides a measure of the central tendency of their respective common normalized eigenvalues.

[0097] Referring to box 274, in an alternative implementation, for each corresponding training object in a first plurality of training objects and for each corresponding training object in a second plurality of training objects, the summation of the common normalized feature values ​​of the first module is the output of the component classifier associated with the first module when the common normalized feature values ​​of the first module in the biological samples obtained from the corresponding training object are input. This is in Figure 4 The diagram illustrates that each module uses miniature network 'spokes'. Individual features are summarized by local networks (rather than by their geometric mean) and then fed into the main classification network (master classifier). Referring to box 276, in some implementations, the component classifiers are neural network algorithms, support vector machine algorithms, decision tree algorithms, unsupervised clustering algorithms, supervised clustering algorithms, logistic regression algorithms, mixture models, or hidden Markov models.

[0098] As used in this article, the master classifier refers to a model with fixed (locked) parameters (weights) and a threshold, which is prepared to be applied to previously unseen samples (e.g., test subjects). In this context, the model refers to a machine learning algorithm, such as logistic regression, neural networks, decision trees, etc. (similar to a model in statistics). Therefore, refer to... Figure 2GIn some embodiments, as indicated by box 278, the master classifier is a neural network. That is, in such embodiments, the master classifier is a neural network with fixed (locked) parameters (weights) and a threshold. In some such embodiments, referring to box 280, the first independent phenotype and clinical condition are identical.

[0099] Referring to box 282, in some implementations where the main classifier is a neural network, for each corresponding training object in a first plurality of training objects of a species, the first training dataset further includes: (iii) a plurality of feature values ​​obtained using biological samples of the corresponding training object from a second module of a plurality of modules through a first technical background, and (iv) an indication of the absence, presence, or stage of a second independent phenotype in the corresponding training object. For each corresponding training object in a second plurality of training objects of a species, the second training dataset further includes: (iii) a plurality of feature values ​​obtained using biological samples of the corresponding training object from a second module through a second technical background, and (iv) an indication of the absence, presence, or stage of a second independent phenotype in the corresponding training object. In other words, as Figure 3 and Figure 4 As shown, there can be more than one module. In the case of block 282, there are two modules. According to block 284, in some such embodiments, the first independent phenotype and the second independent phenotype are identical to the clinical condition (e.g., sepsis). Each corresponding feature in the first module is associated with the first independent phenotype by having a feature value that is statistically significantly larger in the species group than in the group of objects exhibiting the first independent phenotype compared to those not exhibiting the independent phenotype. This is in Figure 3 In the middle as module m upThe following description is provided. In some implementations, the determination of whether a trait is “statistically significantly greater” is assessed by applying a standard t-test, Welch t-test, Wilcoxon test, or permutation test to the abundance of the trait, as measured in subjects exhibiting the first independent phenotype (Group 1) and in subjects not exhibiting the first independent phenotype to reach a p-value (Group 2). In some such implementations, a trait is statistically significantly less abundant (less rich) when the p-value in such a test is 0.05 or less, 0.005 or less, or 0.001 or less. In some such implementations, a trait is statistically significantly less abundant when the p-value in such a test is 0.05 or less, 0.005 or less, or 0.001 or less, adjusted for multiple tests using a false discovery rate procedure such as Benjamini-Hochberg or Benjamini-Yekutieli. See, for example, Benjamini and Hochberg, Journal of the Royal Statistical Society, Series B 57, pp. 289-300; and Benjamini and Yekutieli, 2005, Journal of American Statistical Association 100(469), pp. 71-80, each of which is incorporated herein by reference. In some implementations, the feature is determined to be statistically significantly less by a meta-analysis of fixed or random effects across multiple datasets (groups or training datasets). See, for example, Sianphoe et al . , 2019, BMC Bioinformatics 20:18, incorporated herein by reference.

[0100] Each corresponding feature in the second module is associated with the first independent phenotype by having a statistically significantly fewer feature value in the species group compared to those exhibiting the first independent phenotype. This is in Figure 3 In the middle as module m dnThe following description is provided. In some embodiments, the determination of whether a trait is “statistically significantly less” is assessed by applying a standard t-test, Welch t-test, Wilcoxon test, or permutation test to the abundance of the trait, as measured in subjects exhibiting the first independent phenotype (Group 1) and in subjects not exhibiting the first independent phenotype to reach a p-value (Group 2). In some such embodiments, a trait is statistically significantly less (less abundant) when the p-value in such a test is 0.05 or less, 0.005 or less, or 0.001 or less. In some such embodiments, a trait is statistically significantly less when the p-value in such a test is 0.05 or less, 0.005 or less, or 0.001 or less, adjusted for multiple tests using a false discovery rate procedure such as Benjamini-Hochberg or Benjamini-Yekutieli. See, for example, Benjamini and Hochberg, Journal of the Royal Statistical Society, Series B 57, pp. 289–300; and Benjamini and Yekutieli, 2005, Journal of American Statistical Association 100(469), pp. 71–80, each of which is incorporated herein by reference. In some implementations, the feature is determined to be statistically significantly less by meta-analysis of fixed or random effects across multiple datasets (group or training datasets). See, for example, Sianphoe et al . , 2019, BMC Bioinformatics 20:18, incorporated herein by reference.

[0101] Referring to box 286, in some embodiments of the implementation of box 282, the first independent phenotype and the second independent phenotype are different (e.g., as in...). Figure 3 As shown in the figure, module f up Relative to module s up ).

[0102] Referring to block 288, in some embodiments, the neural network is a feedforward artificial neural network. For disclosures of feedforward artificial neural networks, see, for example, Swozil et al., 1997, Chemometrics and Intelligent Laboratory Systems 39(1), pp. 43-62, which are incorporated herein by reference.

[0103] refer to Figure 2H In some implementations of block 290, the main classifier includes a linear regression algorithm or a penalized linear regression algorithm. For disclosures regarding linear regression algorithms and penalized linear regression algorithms, see, for example, Hastie. et al ., 2001, The Elements of Statistical Learning Springer-Verlag, New York.

[0104] In some implementations, the main classifier is a neural network. See, for example, Hassoun, 1995. Fundamentals of Artificial Neural Networks MIT, which is incorporated herein by reference.

[0105] In some implementations, the main classifier is a support vector machine (SVM) algorithm. SVM is described in the following literature: Cristianini and Shaw-Taylor, 2000, “An Introduction to Support Vector Machines”, Cambridge University Press, Cambridge, Cambridge City; Boser et al ., 1992, “A training algorithm for optimal margin classifiers”, Proceedings of the 5th Annual ACM Symposium on Computational Learning Theory, ACM Press, Pittsburgh, PA, pp. 142-152; Vapnik, 1998, Statistical Learning Theory Wiley, New York; Mount, 2001. Bioinformatics: sequence and genome analysis Cold Spring Harbor Laboratory Press, Cold Spring Harbor, New York; Duda Pattern Classification , Second edition, 2001, John Wiley & Sons, Inc., pp. 259, 262-265; and Hastie, 2001, The Elements of Statistical Learning Springer, New York; and Furey et al ., 2000, Bioinformatics 16, 906-914, each of which is included in full here by reference.

[0106] In some implementations, the main classifier is a tree-based algorithm (e.g., a decision tree). [Reference] Figure 2H In some implementations, the master classifier, as indicated by box 292, is a tree-based algorithm selected from random forest and decision tree algorithms. Decision trees are generally developed by Duda (2001). Pattern Classification John Wiley & Sons, Inc., New York, pp. 395-396, described herein by reference.

[0107] See Figure 2HIn some implementations, the main classifier consists of an ensemble of classifiers subjected to an ensemble optimization algorithm (e.g., AdaBoost, XGBoost, or LightGBM). See Alafate and Freund, 2019, “FasterBoosting with Smaller Memory”, arXiv:1901.09047v1, which is incorporated herein by reference.

[0108] refer to Figure 2H In some implementations, the main classifier, as shown in box 295, is an ensemble of neural networks. See Zhou. et al. , 2002, Artificial Intelligence 137, pp. 239-263, incorporated herein by reference.

[0109] refer to Figure 2H In box 296, in some implementations, the clinical condition is a multi-class clinical condition and the master classifier outputs the probability of each class in the multi-class clinical condition. For example, refer to Figure 3 In some implementation schemes, the clinical condition is bacterial infection (I... bac ), viral infection (I vira ) or non-viral, non-bacterial infection (I non The classifier identifies three categories of situations, and provides the classifier with I... bac The probability of the object having I vira The probability and the object having I non The probability of (where the sum of probabilities is 100%).

[0110] Referring to box 297, in some embodiments, multiple supplementary training datasets are obtained (e.g., 3 or more, 4 or more, 5 or more, 6 or more, 10 or more, or 30 or more). For each of the species-independent corresponding training objects, each of the multiple supplementary datasets includes: (i) multiple feature values ​​obtained using biological samples of the corresponding training object with independent corresponding technical contexts, for a first form of multiple features of the corresponding module in the multiple modules, and (ii) an indication of the absence, presence, or stage of the corresponding phenotype in the corresponding training object corresponding to the corresponding module. In such an embodiment, the co-normalization of box 256 also includes co-normalizing the feature values ​​of features present in the corresponding two or more training datasets in the training group, spanning at least two or more corresponding training datasets including the first training dataset, the second training dataset, and the multiple supplementary training datasets, to remove batch effects between datasets, thereby computing co-normalized feature values ​​for each module in the multiple modules for each of the corresponding training objects in the two or more corresponding training datasets in the multiple training datasets. Furthermore, for each corresponding training object in each training dataset in the training group, the composite training set also includes: (i) a summary of the common normalized feature values ​​of one module in the corresponding training object among the plurality of modules and (ii) an indication of the absence, presence or stage of the corresponding independent phenotype in the corresponding training object.

[0111] Referring to box 298, in some implementations, a test dataset comprising multiple feature values ​​is obtained. For features in at least the first module, multiple feature values ​​are measured in a biological sample of the test subject in a first form (transcriptomics, proteomics, or metabolomics). The test dataset is input into a master classifier to assess the clinical condition of the test subject. That is, the master classifier provides a determination of the clinical condition of the test subject in response to input. In some implementations, the clinical condition is multi-classified, such as... Figure 3 As shown, the determination of the clinical condition of the test subject provided by the master classifier is the probability that the test subject has each constituent class in a multi-class clinical condition.

[0112] In some embodiments, this disclosure relates to a method 1300 for training a classifier for assessing the clinical condition of test subjects, as described below. Figure 13 Detailed description follows. In some embodiments, method 1300 is performed at a system as described herein, for example, system 100 as described above with respect to Figure 1. In some embodiments, method 1300 is performed at a system having a subset of modules and / or databases as described with respect to system 100.

[0113] Method 1300 includes obtaining (1302) characteristic values ​​and clinical status of a first group of training subjects. In some embodiments, the characteristic values ​​are collected from biological samples from training subjects in the first group, such as those described above with respect to method 200. Non-limiting examples of biological samples include solid tissue samples and liquid samples (e.g., whole blood or plasma samples). Further details regarding samples useful to method 1300 are described above with reference to method 200 and will not be repeated here for brevity. In some embodiments, the method described herein includes the step of measuring various characteristic values. In other embodiments, the method described herein acquires previously measured characteristic values, such as characteristic values ​​stored in one or more clinical databases, for example, electronically.

[0114] Two examples of measurement techniques include nucleic acid sequencing (e.g., qPCR or RNAseq) and microarray measurements (e.g., using DNA microarrays, MMChips, protein microarrays, peptide microarrays, tissue microarrays, cell microarrays, compound microarrays, antibody microarrays, glycan arrays, or reversed-phase protein lysate microarrays). However, those skilled in the art will be familiar with other measurement techniques used to measure characteristics from biological samples. Further details regarding feature measurement techniques (e.g., technical background) useful to Method 1300 are described above in reference to Method 200 and will not be repeated here for brevity.

[0115] In some implementations, the same measurement technique is used to collect feature values ​​for each training object in the first group. For example, in some implementations, each feature is of the same type, such as the abundance of proteins, nucleic acids, carbohydrates, or other metabolites, and the technique used to measure the feature value for each value is consistent throughout the first group. For example, in some implementations, the feature is the abundance of mRNA transcripts, and the measurement technique is RNAseq or nucleic acid microarray. In other implementations, for example, when feature values ​​are not co-normalized across different training object groups, different techniques are used to measure feature values ​​across the first training object group. However, in some implementations where feature values ​​are not co-normalized across different groups, for example, when a classifier is trained using a single training object group, the same technique is used to measure feature values ​​across the first group.

[0116] In some implementations, method 1300 includes obtaining (1304) feature values ​​and clinical status of additional training subject groups. In some implementations, feature values ​​are collected for at least two additional groups. In some implementations, feature values ​​are collected for at least three, four, five, six, seven, eight, nine, ten, or more additional groups. In some implementations, the feature values ​​obtained for each group are measured using the same technique. That is, all feature values ​​obtained for the first group are measured using a first technique, all feature values ​​obtained for the second group are measured using a second technique different from the first technique, all feature values ​​obtained for the third group are measured using a third technique different from the first and second techniques, and so on. Further details regarding the use of different feature measurement techniques (e.g., technical background) useful to method 1300 have been described above with reference to method 200 and will not be repeated here.

[0117] In some implementations, such as those in which feature values ​​of multiple training object groups are obtained, method 1300 includes co-normalizing (1306) the feature values ​​between a first group and any additional groups. In some implementations, the feature values ​​of features present in at least the first and second training datasets (e.g., for the first and second training object groups) are co-normalized across at least the first and second training datasets to remove batch effects between datasets, thereby computing co-normalized feature values ​​for multiple modules of the respective training object for each of the first plurality of training objects and each of the second plurality of training objects.

[0118] In some embodiments, the co-normalized eigenvalues ​​existing in at least the first and second training datasets (e.g., and any additional training datasets) across at least the first and second training datasets include estimating the inter-dataset batch effect between the first and second training datasets. In some embodiments, the inter-dataset batch effect includes additive and multiplicative components, and co-normalization is performed by solving an ordinary least squares model for eigenvalues ​​across a first subset of the respective first and second training datasets and reducing the resulting parameters representing the additive and multiplicative components using an empirical Bayesian estimator. In some embodiments, the co-normalization of eigenvalues ​​existing in at least the first and second training datasets across at least the first and second training datasets includes utilizing invariant features or quantile normalization.

[0119] In some implementations, a first phenotype of a corresponding module in a plurality of modules represents a disease state, a first subset of a first training dataset consists of objects without a disease state, and a first subset of a second training dataset (e.g., and any additional training datasets) consists of objects without a disease state. In some implementations, co-normalization of the feature values ​​present in at least the first and second training datasets includes estimating the inter-dataset batch effect between the first and second training datasets using only the first subsets of the respective first and second training datasets. In some implementations, the inter-dataset batch effect includes additive and multiplicative components, and co-normalization is performed by solving an ordinary least squares model using feature values ​​across the first subsets of the respective first and second training datasets and reducing the resulting parameters representing the additive and multiplicative components using an empirical Bayesian estimator.

[0120] The above reference method 200 describes more details about the common normalization techniques useful for method 1300 across various datasets corresponding to the various training groups, and will not be repeated here for the sake of brevity.

[0121] In some implementations, method 1300 includes summarizing (1308) feature values ​​associated with the phenotypes of clinical conditions across multiple modules. Specifically, in some implementations, sub-multiple acquired feature values ​​(e.g., sub-multiple mRNA transcript abundance values) each associated with a specific phenotype of one or more clinical conditions are grouped into modules, and these grouped feature values ​​are summed to form a corresponding summary of feature values ​​for the respective module for each training subject.

[0122] For example, Figure 3 and Figure 4 An example classifier is illustrated, trained to distinguish between three clinical conditions: bacterial infection, viral infection, and neither bacterial nor viral infection. Specifically, Figure 3 An example of a main classifier 300 as a feedforward neural network is illustrated. The input layer 308 is configured to receive a summary 358 of feature values ​​354 from multiple modules 352. For example, as... Figure 4 As shown on the right, module 352-1 includes eigenvalues ​​354-1, 354-2, and 354-3, which correspond to the mRNA abundance values ​​of genes IFI27, JUP, and LAX1, each of which is associated with one or more clinical condition phenotypes in a similar manner. In this case, IFI27, JUP, and LAX1 are all genes upregulated when the subject is infected with a virus. Figure 4As shown, the feature values ​​are summarized by feeding them into a feeder neural network through input layer 304, which includes hidden layer 306 and output summary 358-1, which is used as the input value of the main classifier 300. Each of the other modules 302-2 to 302-6 also includes sub-multiple features obtained for the object, which are different from the sub-multiple features in each of the other modules, and each module is similarly associated with different phenotypes related to one or more clinical conditions. For example, when the object is infected with a virus, the gene in module 302-2 is downregulated. Similarly, the genes in modules 302-3 and 302-4 are upregulated and downregulated, respectively, in patients with sepsis, in contrast to aseptic inflammation. Likewise, in patients who die within 30 days of admission for sepsis, the genes in modules 302-5 and 302-6 are both upregulated and downregulated, respectively.

[0123] In some implementations, method 1300 uses at least three modules, each module including features similarly associated with phenotypes of one or more clinical conditions assessed by the master classifier. In some implementations, method 1300 uses at least six modules, each module including features similarly associated with phenotypes of one or more clinical conditions assessed by the master classifier. In other implementations, method 1300 uses at least 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20 or more modules, each module including features similarly associated with phenotypes of one or more clinical conditions assessed by the master classifier. Further details regarding the modules, particularly regarding the grouping of features associated with specific phenotypes that are useful to method 1300, have been described above with reference to method 200 and will not be repeated here for the sake of brevity.

[0124] although Figure 4 The summarization method shown uses a feeder recurrent network, but other methods for summarizing features of corresponding modules are also considered. Example methods for summarizing module features include neural network algorithms, support vector machine algorithms, decision tree algorithms, unsupervised clustering algorithms, supervised clustering algorithms, logistic regression algorithms, mixture models, or hidden Markov models. In some implementations, the summarization is a measure of the central tendency of the feature values ​​of the corresponding module. Non-limiting examples of measures of central tendency include the arithmetic mean, geometric mean, weighted mean, median range, median pivot, three-mean, Winsorized mean, median, and mode of the feature values ​​of the corresponding module. Further details regarding methods for summarizing feature values ​​of modules that are useful to Method 1300 are described above in reference to Method 200 and will not be repeated here for brevity.

[0125] Method 1300 then includes training (1310) a master classifier for (i) derived values ​​of feature values ​​from one or more training groups of objects and (ii) the clinical state of objects in one or more training groups. In some embodiments, the master classifier is a neural network algorithm, a support vector machine algorithm, a decision tree algorithm, an unsupervised clustering algorithm, a supervised clustering algorithm, a logistic regression algorithm, a mixture model, or a hidden Markov model. In some embodiments, the master classifier is a neural network algorithm, a linear regression algorithm, a penalized linear regression algorithm, a support vector machine algorithm, or a tree-based algorithm. In some embodiments, the master classifier consists of an ensemble of classifiers subjected to an ensemble optimization algorithm. In some embodiments, the ensemble optimization algorithm includes AdaBoost, XGBoost, or LightGBM. Methods for training classifiers are well known in the art. Further details regarding the types of classifiers useful to method 1300 and the methods used to train those classifiers are described above with reference to method 200 and will not be repeated here for brevity.

[0126] In some implementations, the eigenvalues ​​derived are co-normalized eigenvalues ​​(1312). That is, in some implementations, method 1300 includes the step of co-normalizing eigenvalues ​​across two or more training datasets, for example, datasets formed from eigenvalues ​​obtained using different measurement techniques as described above with respect to methods 200 and 1300, but not the step of summing up groups of eigenvalues ​​subdivided into different modules.

[0127] In some implementations, the derived feature value is a summary of the feature values ​​(1314). That is, in some implementations, method 1300 does not include the step of co-normalizing feature values ​​across two or more training datasets, for example, in the case of using a single measurement technique to obtain all feature values, but does include the step of summarizing groups of feature values ​​subdivided into different modules, for example, as described above with respect to methods 200 and 1300.

[0128] In some implementations, the eigenvalue derived value is a summation of the common-normalized eigenvalues ​​(1316). That is, in some implementations, method 1300 includes the step of common-normalizing eigenvalues ​​across two or more training datasets, for example, datasets formed from eigenvalues ​​obtained using different measurement techniques, as described above with respect to methods 200 and 1300, and the step of summarizing the groups of common-normalized eigenvalues ​​subdivided into different modules, for example, as described above with respect to methods 200 and 1300.

[0129] In some implementations, the derived eigenvalues ​​are common-normalized summaries of the eigenvalues ​​(1318). That is, in some implementations, method 1300 includes a first step of summarizing groups of eigenvalues ​​subdivided into different modules, for example, as described above with respect to methods 200 and 1300, and a second step of common-normalizing the summaries from the modules across two or more training datasets using common-normalization techniques as described above with respect to methods 200 and 1300, for example, datasets formed from eigenvalues ​​obtained using different measurement techniques.

[0130] It should be understood that the description Figure 13 The specific order of operations described herein is merely an example and is not intended to indicate that the described order is the only possible order of operations. Those skilled in the art will recognize various ways in which the operations described herein can be reordered. For example, in some embodiments, the aggregation (1308) of feature values ​​for each module is performed prior to cross-group common normalization (1306), where feature data are collected using different measurement techniques. Additionally, it should be noted that other methods described herein (e.g., method 200 described above with respect to Figure 2 and method 200 described below with respect to Figure 2) are also relevant. Figure 14 The details of other processes described herein (method 1400) also apply in a similar manner to those described above regarding... Figure 13 Method 1300 is described above. For example, the feature values, modules, clinical conditions, clinical phenotypes, measurement techniques, etc., described above with reference to Method 1300 may optionally have one or more features of the feature values, modules, clinical conditions, clinical phenotypes, measurement techniques, etc., described herein with reference to other methods (e.g., Method 200 or 1400). Similarly, the methods used in the various steps (e.g., data collection, common normalization, summarization, classifier training, etc.) described above with reference to Method 1300 may optionally have one or more features of the data collection, common normalization, summarization, classifier training, etc., described herein with reference to other methods (e.g., Method 200 or 1400). For the sake of brevity, these details will not be repeated here.

[0131] In some implementations, this disclosure relates to method 1400 for assessing the clinical condition of test subjects, as described below. Figure 14 Detailed description. In some embodiments, method 1400 is performed at a system as described herein, for example, system 100 as described above with respect to Figure 1. In some embodiments, method 1400 is performed at a system having a subset of modules and / or databases as described with respect to system 100.

[0132] Method 1400 includes obtaining (1402) characteristic values ​​of the test subject. In some embodiments, the characteristic values ​​are collected from a biological sample from the test subject, such as those described above with respect to methods 200 and 1300. Non-limiting examples of biological samples include solid tissue samples and liquid samples (e.g., whole blood or plasma samples). Further details regarding samples useful for method 1400 are described above with reference to methods 200 and 1300 and will not be repeated here for brevity. In some embodiments, the methods described herein include the step of measuring various characteristic values. In other embodiments, the methods described herein acquire previously measured characteristic values, such as characteristic values ​​stored in one or more clinical databases, for example, electronically.

[0133] Two examples of measurement techniques include nucleic acid sequencing (e.g., qPCR or RNAseq) and microarray measurements (e.g., using DNA microarrays, MMChips, protein microarrays, peptide microarrays, tissue microarrays, cell microarrays, compound microarrays, antibody microarrays, glycan arrays, or reversed-phase protein lysate microarrays). However, those skilled in the art will be familiar with other measurement techniques used to measure characteristics from biological samples. Further details regarding feature measurement techniques (e.g., technical background) useful to Method 1400 are described above with reference to Methods 200 and 1300 and will not be repeated here for brevity.

[0134] In some implementations, for example, where the classifier is trained to evaluate feature values ​​obtained from various different measurement methods (e.g., technical background), method 1400 includes common normalization (1404) of feature values ​​for a predetermined pattern. In some implementations, the predetermined pattern is derived from common normalization of feature data across two or more training datasets, for example, using different measurement methods. Various methods for common normalization across different training datasets have been described in detail above with reference to methods 200 and 1300, and will not be repeated here for the sake of brevity. In some implementations, the feature values ​​obtained for the test object are not normalized, and the normalization describes the measurement technique used to obtain these values.

[0135] In some embodiments, method 1400 includes grouping (1406) feature values ​​or normalized feature values ​​of an object into a plurality of modules, wherein each feature value in a respective module is associated in a similar manner with a phenotype associated with one or more categories of the clinical condition being assessed. That is, in some embodiments, sub-multiple acquired feature values ​​(e.g., sub-multiple mRNA transcript abundance values) each associated with a specific phenotype of one or more clinical conditions are grouped into modules. In some embodiments, method 1400 uses at least three modules, each module including features similarly associated with a phenotype of one or more clinical conditions assessed by a master classifier. In some embodiments, method 1400 uses at least six modules, each module including features similarly associated with a phenotype of one or more clinical conditions assessed by a master classifier. In other embodiments, method 1400 uses at least 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20 or more modules, each module including features similarly associated with a phenotype of one or more clinical conditions assessed by a master classifier. Further details regarding the modules, particularly the grouping of features associated with specific phenotypes that are useful for Method 1400, have been described above with reference to Methods 200 and 1300, and will not be repeated here for the sake of brevity. In some implementations, feature values ​​are not grouped into modules but are directly input into the main classifier.

[0136] In some implementations, method 1400 includes summarizing (1408) the feature values ​​in each corresponding module to form a corresponding summary of the feature values ​​of the corresponding modules of the test object. For example, as described above for... Figure 3 and 4 As shown in module 352-1.

[0137] although Figure 4 The summarization method shown uses a feeder recurrent network, but other methods for summarizing features of corresponding modules are also considered. Example methods for summarizing module features include neural network algorithms, support vector machine algorithms, decision tree algorithms, unsupervised clustering algorithms, supervised clustering algorithms, logistic regression algorithms, mixture models, or hidden Markov models. In some implementations, the summarization is a measure of the central tendency of the feature values ​​of the corresponding module. Non-limiting examples of measures of central tendency include the arithmetic mean, geometric mean, weighted mean, median range, median pivot, three-mean, Winsorized mean, median, and mode of the feature values ​​of the corresponding module. Further details regarding the methods for summarizing feature values ​​of modules that are useful for method 1400 are described above in references to methods 200 and 1300, and will not be repeated here for brevity.

[0138] Method 1400 then includes inputting derived values ​​of the feature values ​​(1410) into a classifier trained to distinguish different categories of clinical conditions. In some embodiments, the classifier is trained to distinguish two clinical conditions. In some embodiments, the classifier is trained to distinguish at least three different categories of clinical conditions. In other embodiments, the classifier is trained to distinguish at least four, five, six, seven, eight, nine, ten, fifteen, twenty, or more different categories of clinical conditions.

[0139] Referring to methods 200 and 1300, the master classifier is trained as described above. In short, the master classifier is trained on (i) derived values ​​of feature values ​​from one or more training groups and (ii) the clinical state of the training subjects in one or more training groups. In some embodiments, the master classifier is a neural network algorithm, support vector machine algorithm, decision tree algorithm, unsupervised clustering algorithm, supervised clustering algorithm, logistic regression algorithm, mixture model, or hidden Markov model. In some embodiments, the master classifier is a neural network algorithm, linear regression algorithm, penalized linear regression algorithm, support vector machine algorithm, or tree-based algorithm. In some embodiments, the master classifier consists of an ensemble of classifiers subjected to an ensemble optimization algorithm. In some embodiments, the ensemble optimization algorithm includes AdaBoost, XGBoost, or LightGBM. Methods for training the classifier are well known in the art. Further details regarding the types of classifiers useful to method 1400 and the methods used to train those classifiers have been described above with reference to methods 200 and 1300 and will not be repeated here for brevity.

[0140] In some implementations, the derived eigenvalues ​​are normalized eigenvalues ​​that depend on the measurement platform (1412). That is, in some implementations, method 1400 includes a step of normalizing the eigenvalues ​​based on the method used to acquire the eigenvalue measurement, which is the opposite of other measurement methods used in the training group, as described above with respect to methods 200 and 1300, but not a step of summing up groups of eigenvalues ​​subdivided into different modules.

[0141] In some implementations, the eigenvalue derived value is a summary of the eigenvalues ​​(1414). That is, in some implementations, method 1400 does not include a step of normalizing the eigenvalues ​​based on the method used to obtain the eigenvalue measurement, which is contrary to other measurement methods used in the training group, but does include a step of summarizing the groups of eigenvalues ​​subdivided into different modules, as described above with respect to methods 200 and 1300.

[0142] In some implementations, the derived eigenvalues ​​are sums of normalized eigenvalues ​​(1416). That is, in some implementations, method 1400 includes the step of normalizing the eigenvalues ​​based on the method used to obtain the eigenvalue measurement, which is the opposite of other measurement methods used in the training group, as described above with respect to methods 200 and 1300, and the step of summarizing the groups of normalized eigenvalues ​​subdivided into different modules, for example, as described above with respect to methods 200 and 1300.

[0143] In some implementations, the derived eigenvalues ​​are a common-normalized sum of eigenvalues ​​(1418). That is, in some implementations, method 1400 includes a first step of summarizing groups of eigenvalues ​​subdivided into different modules, for example, as described above with respect to methods 200 and 1300, and a second step of normalizing the eigenvalues ​​based on the method used to obtain the eigenvalue measurement, which is contrary to other measurement methods used in the training group, as described above with respect to methods 200 and 1300.

[0144] In some embodiments, method 1400 further includes a step of treating the test subject based on the output of a classifier. In some embodiments, the classifier provides a probability that the subject has one of several categories of the clinical condition being assessed. Treatment decisions can be based on the output when the probability from the classifier output definitively identifies a category of clinical condition or definitively excludes a specific category of clinical condition. For example, if the classifier output indicates that the subject has a first category of clinical condition, the subject is treated by administering a first therapy tailored for the first category of clinical condition. Conversely, if the classifier output indicates that the subject has a second category of clinical condition, the subject is treated by administering a second therapy tailored for the second category of clinical condition.

[0145] For example, refer to Figure 4 The classifier shown is trained to assess whether an object has a bacterial infection, a viral infection, or inflammation unrelated to either bacterial or viral infection. After test data is fed into the classifier, when the classifier indicates that an object has a bacterial infection, an antibacterial agent, such as an antibiotic, is administered to the object. However, when the classifier indicates that an object has a viral infection, the object is not given antibiotics but may be given antiviral agents. Similarly, when the classifier indicates that an object has inflammation unrelated to either bacterial or viral infection, the object is not given antibiotics or antiviral agents but may be given anti-inflammatory agents.

[0146] It should be understood that the description Figure 14The specific order of operations described herein is merely an example and is not intended to indicate that the described order is the only possible order of operations. Those skilled in the art will recognize various ways in which the operations described herein can be reordered. For example, in some embodiments, the aggregation (1408) of feature values ​​for each module is performed prior to normalization (1404) across groups, where feature data are collected using different measurement techniques. Additionally, it should be noted that other methods described herein (e.g., method 200 described above with respect to Figure 2 and method 200 described above with respect to Figure 2) are also relevant. Figure 13 The details of other processes described herein (method 1300) also apply in a similar manner to the above description of... Figure 14 Method 1400 is described above. For example, the feature values, modules, clinical conditions, clinical phenotypes, measurement techniques, etc., described above with reference to Method 1400 may optionally have one or more features of the feature values, modules, clinical conditions, clinical phenotypes, measurement techniques, etc., described herein with reference to other methods (e.g., Method 200 or 1300). Similarly, the methods used in the various steps (e.g., data collection, common normalization, summarization, classifier training, etc.) described above with reference to Method 1400 may optionally have one or more features of the data collection, common normalization, summarization, classifier training, etc., described herein with reference to other methods (e.g., Method 200 or 1300). For the sake of brevity, these details will not be repeated here.

[0147] Example 1 Systematic search and inclusion criteria for clinical infection gene expression studies IMX training datasets meeting the defined inclusion criteria for clinical infection studies were obtained from the NCBI GEO (www.ncbi.nlm.nih.gov / geo / ) and EMBL-EBI ArrayExpress (www.ebi.ac.uk / arrayexpress) databases. Specifically, inclusion criteria included that patients in the study 1) must have been diagnosed by a physician as having an infection of a specific type (e.g., a strict bacterial infection, a strict viral infection, or a non-infectious inflammation), and 2) have gene expression measurements of 29 diagnostic biomarkers previously identified by Sweeney et al. (Sweeney et al.). et al ., 2015, Sci Transl Med 7(287),pp. 287ra71; Sweeney et al , 2016, Sci Transl Med 8(346), pp. 346ra91; and Sweeney et al, 2018, Nature Communications 9, p. 694), 3) over 18 years of age, 4) already seen in a hospital setting (e.g., emergency department, intensive care unit), 5) having a community- or hospital-acquired infection, and 6) having a blood sample collected within 24 hours of initial suspicion of infection and / or sepsis. Furthermore, the normalization / batch effect control method used required that each included study must measure at least a control sample (e.g., a sample not diagnosed with any of the three considered conditions). Studies were excluded if the patient had experienced trauma or had a condition that would not be encountered in a typical clinical setting (e.g., experimental LPS attack) or that was confounded by infection (e.g., anaphylactic shock).

[0148] Example 2 Normalization of expression data and co-normalization of COCONUT Then, normalization is performed for each study, using one of two methods depending on the platform. For Affymetrix arrays, Robust Multi-array Average (RMA) (Irizarry) is used. et al ., 2003,Biostatistics, 4(2):249-64) or gcRMA(Wu et al (Journal of the American Statistical Association, 2004, 99:909-17) Normalized expression data. Exponential convolution was used to normalize expression data from other platforms for background correction, followed by quantile normalization.

[0149] After normalizing the original expression data, the COCONUT algorithm (Sweeney) was used. et al. , 2016, Sci Transl Med 8(346), pp. 346ra91; and Abouelhoda et al (BMBB Bioinformatics 9, p. 476) These measurements are co-normalized to ensure their comparability in studies. COCONUT is based on ComBat (Johnson...). et al(2007, Biostatistics, 8, pp. 118-127) Based on the empirical Bayesian batch correction method, the expected expression value of each gene from healthy patients was calculated and adjusted for study-specific location modifications (mean) and scale (standard deviation) in gene expression. For this analysis, ComBat's parametric priors were used, where the gene expression distribution was assumed to be Gaussian, and the empirical prior distributions for study-specific location and variance modification parameters were Gaussian and inverse gamma distributions, respectively.

[0150] Example 3 Developing a sepsis classifier using machine learning. To develop a sepsis classifier, a machine learning approach was employed. This approach involves specifying candidate models, evaluating the performance of different classifiers using training data and specified performance statistics, and then selecting the best-performing model to evaluate independent data.

[0151] In this context, "model" refers to a machine learning algorithm, such as logistic regression, neural networks, decision trees, etc. (similar to the models used in statistics). Similarly, in this case, the master classifier refers to a model with fixed (locked) parameters (weights) and a threshold, ready to be applied to previously unseen samples. The classifier uses two types of parameters: weights learned by the core learning algorithm (e.g., XGBoost), and additional user-provided parameters that serve as input to the core learner. These additional parameters are called hyperparameters. Classifier development requires learning (fixed) weights and hyperparameters. Weights are learned by the core learning algorithm; hyperparameters are learned by the core learning algorithm. For this study, a random search method (Bergstra) is employed. et al ., 2012, Journal of Machine Learning Research 13, pp. 281-305).

[0152] The performance of four different types of predictive models was compared: 1) logistic regression with lasso (L1) penalty, 2) support vector machine (SVM) classifier with radial basis function kernel (RBF), 3) extreme gradient boosting tree (XGBoost), and 4) multilayer perceptron (MLP). The accuracy of each type of predictive model in classifying patient samples into one of the following: a) strict bacterial infection, b) strict viral infection, or c) non-infectious inflammation was evaluated.

[0153] To evaluate each predictive model on this three-class classification task, a metric called the average area under the paired ROC curve (APA) was developed. APA is defined as the average of the three area under the ROC curves for all (OVA) classes; that is, the average of bacteria against other AUCs, viruses against other AUCs, and non-infectious against other AUCs.

[0154] Various methods for evaluating the performance of a particular classifier (e.g., models with a fixed set of weights and hyperparameters) can be used in machine learning. Here, cross-validation (CV) is employed, a well-established method for small-sample scenarios such as sepsis research. Two CV variants are used, as described below.

[0155] Example 4 Model cross-validation method Two different types of cross-validation schemes were initially considered: traditional 5-fold cross-validation and leave-one-out-of-studies (LOSO) cross-validation. For 5-fold CV trials, all IMX samples were randomly partitioned into five non-overlapping subsets of roughly similar sample sizes using standard methods. For LOSO CV trials, each study was considered a CV partition. In this way, at each step (“fold”) of LOSO CV, the candidate model was trained on all studies except one, and then the trained model was used to generate predictions for the remaining studies.

[0156] The basic principle of using LOSO CV is as follows. In short, k-fold CV assumes that the training and validation samples for cross-validation come from the same distribution. However, due to the exceptional heterogeneity of sepsis research, this assumption is not even roughly satisfied. LOSO aims to support the model that is most robust to this heterogeneity empirically; in other words, the model most likely to generalize well to previously unseen studies. This is a key requirement for the clinical application of sepsis classifiers.

[0157] The LOSO method relates to previous work that proposed clustering the training data before cross-validation as a means of assessing heterogeneity (Tabe-Bordbar, 2018). et al ., Sci Rep 8(1), pp.6620). In this case, clustering is not necessary because it naturally follows the process from the partitioning of the training data to the study.

[0158] In both k-fold CV and LOSO, predictions are concentrated on the left fold of all folds to evaluate model performance. Alternatively, CV statistics can be calculated by estimating the statistics of interest for each fold and then averaging the results for each fold. In this study, LOSO requires pooling because most studies do not have samples from all three classes, so most of the statistics of interest cannot be calculated on a single LOSO fold. In this case, a pooling approach is uniformly applied to ensure a fair comparison with k-fold CV.

[0159] Hierarchical cross-validation (HCV) was used to determine the appropriate cross-validation scheme and feature set for the selection and prospective validation of the diagnostic classifier. HCV is technically equivalent to nested cross-validation (NCV). However, it is referred to as HCV here because it serves a different purpose than NCV. Specifically, in NCV, the goal is to estimate the performance of the selected model. In contrast, HCV is used here to evaluate and compare the components (steps) of the model selection process.

[0160] HCV divides the IMX dataset into three folds; each fold is constructed such that all samples from a given study appear in only one fold. These three HCV folds are manually constructed to have similar compositions of bacteria, viruses, and uninfected samples. To evaluate 5-fold and LOSO CV within this framework, each CV method is performed on samples from two HCV folds (inner folds). Models are then ranked according to their CV performance on the inner folds (based on APA), and the top 100 models from each CV method are evaluated on the remaining third HCV fold (outer fold). This procedure is performed three times, each time setting the outer fold as one HCV fold and the inner fold as the remaining two HCV folds.

[0161] Example 5 Predictive model evaluation and hyperparameter search Discovering promising candidate predictive models involves identifying the hyperparameter values ​​for each model that lead to robust generalization performance. The four predictive models evaluated here can be broadly categorized into models with small (low-dimensional) or large (high-dimensional) numbers of hyperparameters. More specifically, the predictive model with a low-dimensional hyperparameter space is logistic regression with a lasso-penalized SVM, while the predictive models with a high-dimensional hyperparameter space are XGBoost and MLP. For the predictive models with a low-dimensional hyperparameter space, 5000 model instances (with different values ​​of the corresponding hyperparameters) were sampled for cross-validation evaluation. For the predictive models with a high-dimensional hyperparameter space (e.g., XGBoost and MLP), 100,000 model instances were randomly sampled. In the case of logistic regression, only one hyperparameter is considered: the lasso penalty coefficient. For SVM, the values ​​of the C penalty term and the kernel coefficient γ are sampled. For XGBoost, the following hyperparameters were sampled: 1) pseudo-random number generator seed, 2) learning rate, 3) minimum loss reduction required to introduce splits in the classifier tree, 4) maximum tree depth, 5) minimum child node weights, 6) minimum sum of instance weights required per child node, 7) maximum increment step size, 8) L2 penalty coefficient for weight regularization, 9) tree method (exact or approximate), and 10) number of rounds. For MLP, the batch size was fixed at 128, and the optimization algorithm was ADAM. The following hyperparameters were then sampled: 1) number of hidden layers, 2) number of nodes per hidden layer, 3) activation function type per hidden layer (e.g., ReLU and variants, straight chain, sigmoid, hyperbolic tangent), 4) learning rate, 5) number of training iterations, 6) type of weight regularization (L1, L2, none), and 7) presence (enabled) and quantity (probability) of dropout in the input and hidden layers. The number of nodes per hidden layer was the same across all hidden layers. The β1, β2 and ε parameters of ADAM are fixed at 0.9, 0.999 and 1e-08, respectively.

[0162] In the case of XGBoost and MLP, some hyperparameters are sampled uniformly from the grid, while others are sampled from a continuous range according to the Bergstra & Bengio method described above.

[0163] Example 6 Fine-tuning of neural network hyperparameters In neural network analysis, significant variations in observed results are related to the seed value used to initialize network weights. To explain this variability, various approaches, including different ensemble models, were considered. Based on empirical evidence, an approach was adopted that incorporates the seed as an additional hyperparameter into the search. The “core” hyperparameters are searched randomly, while the seed is searched exhaustively using a fixed, predefined list of 1000 values.

[0164] Adding a random seed significantly increases the hyperparameter search space. To reduce computation, a large grid of hyperparameters (excluding the seed) is used as a starting point. For each random sample from the grid, more than 250 seed values ​​are searched. After the initial search is complete, a smaller grid is selected for the most promising hyperparameters. The hyperparameter values ​​are then refined by searching around the promising hyperparameter configurations. For each randomly sampled fine-tuning point, an additional, larger set of seed values ​​(e.g., 750) is searched. The configuration with the largest APA is selected as the final, locked set of hyperparameter values. This set includes the random number generator seed.

[0165] Example 7 Diagnostic markers and geometric mean feature groups Two sets of input features were considered in these analyses. The first set consisted of 29 genetic markers previously identified as highly discriminative of the presence, type, and severity of infection (Sweeney). et al ., 2015, Sci Transl Med 7(287), pp. 287ra71; Sweeney et al , 2016, Sci Transl Med 8(346), pp. 346ra91; and Sweeney et al (See Nature Communications, 2018, p. 694). The second set of input features is based on modules (subsets of relevant genes). These 29 genes are divided into 6 modules, such that each module consists of genes that share an expression pattern (trend) under a given infection or severity condition. For example, genes in the fever module are overexpressed (upregulated) in patients with fever. The composition of the modules is shown in Table 1b.

[0166] Table 1b. Definition and composition of sepsis-related modules (genomes). Fever / Fever Reduction: Genes with increased / decreased expression in severe viral infections. Sepsis Increase / Decrease: Genes with increased / decreased expression in patients with sepsis compared to patients with aseptic inflammation. Severity Increase / Decrease: Genes with increased / decreased expression in patients who died within 30 days of admission. The module-based features used in these analyses are geometric mean values ​​calculated based on the gene expression values ​​within each module, resulting in six geometric mean scores for each patient sample. This approach can be viewed as a form of "feature engineering," a method known to sometimes significantly improve the performance of machine learning classifiers.

[0167] Example 8 Iterative application of COCONUT to compare the IMX and ICU datasets To validate the predictive model trained on IMX using external validation of the clinical dataset, two datasets were generated using comparable first-stage production expression levels on different technology platforms (e.g., microarrays for IMX and NanoStrings for validating the clinical data). After normalizing the raw expression data, we used the COCONUT algorithm (Sweeney's algorithm) to perform external validation on the clinical dataset. et al. (2016, Sci Transl Med 8(346), pp. 346ra91) These measurements were co-normalized to ensure their comparability in the study. COCONUT is based on ComBat (Johnson) et al (Biostatistics, 2007, 8, pp. 118-127) Based on the empirical Bayesian batch correction method, the expected expression value of each gene from healthy patients is calculated and adjusted for study-specific location modifications (mean) and scale (standard deviation) in gene expression. For this analysis, we use the parametric prior of ComBat, where the gene expression distribution is assumed to be Gaussian, and the empirical prior distributions for study-specific location and variance modification parameters are Gaussian and inverse gamma distributions, respectively. Advantageously, the COCONUT algorithm is applied iteratively, applying co-normalization to the healthy samples of the IMX dataset while keeping the healthy samples of the validation clinical dataset unchanged at each step. In this setting, the NanoString healthy samples represent the target dataset because it remains constant throughout the process, while the IMX healthy samples represent the query dataset similar to the target dataset. The process terminates when the mean absolute deviation (MAD) between the mean expression vectors of the 29 diagnostic biomarkers in IMX and NanoString does not change by more than 0.001 in successive iterations. More detailed pseudocode for this process is shown in [link to pseudocode]. Figure 12 middle.

[0168] According to Figure 1 and Figure 12 This disclosure provides a computer system 100 for dataset common normalization, the computer system including at least one processor 102 and memory 111 / 112 storing at least one program (e.g., data common normalization module 124) executed by at least one processor.

[0169] The at least one procedure further includes instructions for (A) obtaining a first training dataset electronically. For each of the first plurality of training objects of a species, the first training dataset includes: (i) first plurality of feature values ​​obtained for a plurality of features using biological samples of the corresponding training object, and (ii) an indication of the absence, presence, or stage of a clinical condition in the corresponding training object, wherein a first subset of the first training dataset consists of objects that do not present a clinical condition (e.g., Figure 12(Q dataset).

[0170] The at least one procedure also includes instructions for (B) obtaining a second training dataset electronically. For each corresponding training object in the second plurality of training objects of a species, the second training dataset includes: (i) second plurality of feature values ​​obtained for a plurality of features using biological samples of the corresponding training object, and (ii) an indication of the absence, presence, or stage of a clinical condition in the corresponding training object, wherein a first subset of the second training dataset consists of objects that do not present a clinical condition (e.g., Figure 12 The T dataset).

[0171] The at least one procedure further includes instructions for (C) estimating an initial mean absolute deviation between (i) a vector of the average representation of a subset of features across a first plurality of objects and (ii) a vector of the average representation of a subset of features across a second plurality of objects (e.g., Figure 12 (Step 2). For example, such as Figure 12 As shown in step 2, in some implementations, (C) estimating the initial mean absolute deviation between (i) the vector of the average representation of a subset of features across a first plurality of objects and (ii) the vector of the average representation of a subset of features across a second plurality of objects includes setting the initial mean absolute deviation to zero.

[0172] The at least one procedure further includes instructions for (D) co-normalizing feature values ​​of a subset of multiple features across at least the first and second training datasets to remove inter-dataset batch effects, wherein a subset of features exists in at least the first and second training datasets, the co-normalization comprising estimating inter-dataset batch effects between the first and second training datasets using only a first subset of the respective first and second training datasets, and the inter-dataset batch effects comprising additive and multiplicative components, and the co-normalization solving an ordinary least squares model for feature values ​​across the first subset of the respective first and second training datasets, and narrowing the resulting parameters representing the additive and multiplicative components using an empirical Bayesian estimator, thereby using the resulting parameters to compute: for each respective training object in the first plurality of training objects, the co-normalized feature value of each feature value among the plurality of features (e.g., Figure 12 Step 3a and as in Sweeney et al (Published in Sci Transl Med 8(346), pp. 346ra91, 2016).

[0173] The at least one procedure further includes instructions for (F) estimating (i) a vector of the average representation of the common normalized eigenvalues ​​of multiple features across a first training dataset and (ii) a vector of the average representation of a subset of multiple features across a second training dataset (e.g., Figure 12 Steps 3b, 3c, 3d, and 3e.

[0174] The at least one procedure further includes instructions for (G) repeating the common normalization (E) and estimation (F) until the common normalized mean absolute deviation converges (e.g., Figure 12 Steps 3f and 3g, and the while condition τ > 0001 in step 3.

[0175] Example 9 Commercial health samples used for general comparison with NanoString representation data. Deploying the iterative COCONUT procedure described above in a clinical setting is not feasible because it requires acquiring healthy samples at the deployment site and re-comparing all healthy samples (both previously acquired and newly acquired). To establish a universal model of NanoString expression in healthy patients, a cohort of 40 commercially available healthy control samples was identified, including ten PAXGENE™ whole blood RNA samples, each acquired from four different locations in the continental United States. The donors providing these samples self-reported as healthy and negative for HIV and hepatitis C. In terms of gender, 12 healthy samples were from female donors, and the remaining 28 samples were from male donors.

[0176] Example 10 Validation of clinical study sample descriptions and NanoString expression profiles This study recruited patients admitted for suspected sepsis. To generate NanoString expression in ICU samples, RNA was isolated using the RNeasy Plus Micro Kit (Qiagen, part #74034) on a QIAcube (Qiagen), and PAXgene RNA was extracted from each sample using a custom script for QIAcube RNA isolation. Each expression profiling reaction consisted of 150 ng RNA per sample. Following the manufacturer's instructions, a custom probe code set for detecting the expression of our biomarker set and sample RNA were hybridized at 65°C for 16 hours. NanoString expression was then generated using the nCounter SPRINT standard protocol, producing raw RCC expression files. These raw expression values ​​were not normalized. After processing, a total of 104 data samples were available for analysis.

[0177] As stated above, 18 studies in the public domain that met the inclusion criteria and were used for classifier training were identified. These studies included 1069 different patient samples. The composition and key characteristics of the studies are shown in Table 2.

[0178] Table 2. Characteristics of the training study. ED = Emergency Department; ICU = Intensive Care Unit. ED / ICU is the percentage of samples collected from the ED (the remainder from the ICU). Platform = Gene Expression Platform. Numbers in parentheses represent percentages. 1 Platform: A = Agilent, I = Illumina Normalization Following the procedure described above, the training data for study normalization was iteratively adjusted using COCONUT, PROMPT data, and 40 commercial control samples processed on a NanoString instrument. The resulting batch-adjusted training data was then used for exploratory data analysis and machine learning. To illustrate the iterative process of COCONUT co-normalization, Figure 5 plots the distribution of selected genes in the training set before, during, and after normalization. As expected, with iteration, the distributions in the target and query datasets became visually more similar.

[0179] Exploratory data analysis The distribution of co-normalized expression values ​​for each of the 29 genes used in the algorithm for bacterial, viral, and uninfected samples was then visualized, as shown in Figure 6. This histogram demonstrates a moderate (bacteria vs. virus) to minimal (uninfected) class separation at the individual gene level, highlighting the need for advanced multigene modeling to achieve clinical utility for the sepsis classifier. Next, the projections of the three classes were visualized in 2D and 3D using t-distributed random neighborhood embeddings (t-SNE) (Figure 7) and principal component analysis (PCA) (Figure 8). Both analyses confirm the preliminary finding that a high-dimensional classifier needs to be developed to achieve clinical feasibility.

[0180] The samples were also plotted in two-dimensional PCA space through research, such as... Figure 9 As shown, this result indicates that residual research effects exist after normalization using COCONUT. This observation, along with existing research in the field, suggests that classifiers must be tested on diverse, previously unseen studies to avoid being confounded by the studies (e.g., to avoid learning batches rather than disease signals). This is particularly important given that some studies in the training set are for a single disease.

[0181] Leave-one-out research and cross-validation Disease heterogeneity and the residual batch effect suggest that conventional cross-validation for model selection can lead to significant overfitting. To test this hypothesis, a comparative analysis of two model selection methods was conducted: 5-fold cross-validation and leave-one-out cross-validation. This analysis used 3-fold hierarchical cross-validation (HCV), where each outer fold simulates independent validation of the best classifier selected in the inner loop. This exposes the potential overfitting of certain classifier selection methods without a separate (and unavailable) validation set. These studies were combined to make the class distributions in each partition as similar as possible.

[0182] In HCV, classifier tuning is performed in each inner loop using either standard CV or LOSO. To select the best model, we ranked candidates by the mean pairwise AUROC statistic (APA). The reasons for choosing APA are: (1) it showed the most consistent behavior across the training and test data of all relevant statistics in the preliminary analysis, (2) it is clinically highly relevant in diagnosing sepsis, and (3) the choice of statistics for model selection is not important, as previous evidence suggests a large gap in generalization ability between CV and LOSO. In other words, other statistics could be used, but APA is a straightforward choice.

[0183] The comparisons were made using SVMs with RBF kernels, deep learning MLPs, logistic regression (LR), and XGBoost classifiers. The underlying principles for using these classifiers are: (1) for SVMs, prior experience and use in existing clinical diagnostic tests; (2) for LRs, widespread acceptance in general medicine, particularly in the diagnosis of infectious diseases; (3) for XGBoosts, widespread acceptance in the machine learning community and a proven track record of best performance in major competitive challenges such as Kaggle; and (4) for deep neural networks, recent breakthroughs in multiple application areas (image analysis, speech recognition, natural language processing, reinforcement learning).

[0184] The analysis used 29 normalized expression profiles as input features and 6 GM scores as input features for the classifier. The reason for using 6 GM scores is that they showed very promising results in previous studies and preliminary analyses (internal data, not shown). The results are shown in Figures 10 and 11.

[0185] In all analyses, except for one of the GM logistic regression runs, the LOSO CV AUC estimate was closer to the test set value than the k-fold CV estimate. This is evidenced by the closeness of the blue (LOSO) points to the vertical dashed line compared to the red (k-fold) points. Based on this finding, the remaining analyses used LOSO.

[0186] Furthermore, the analysis showed that the test set using 6 GM scores performed better than the 29 gene expression features. Table 3 shows a comparison of the test set APA of the two sets of features and different classifiers. LOSO was used as the model selection criterion for this comparison because it was previously found to have much less bias.

[0187] Table 3. Comparison of test set performance using GM scores and gene expression as input features. This table contains the APA values ​​for the GM scores (GMS) and the 29 gene expression values ​​(GENEX). The APA column contains the average of the 10 models shown in Figure 11, for the three HCV test sets. The best model was found using the LOSO cross-validation method. For each GMS / GENEX pair, higher APAs are indicated by bold letters. As shown in Table 3, the GMS score yielded higher performance in almost all cases. Based on this finding, the remaining analyses used the GM score as an input feature for the classification algorithm. This use of the GM score is a combination of the above... Figure 1A and 1B Instantiation of module 152 / summary algorithm 156 under discussion.

[0188] Classifier development To develop the classifier, hyperparameter searches were performed on four different models. The LOSO cross-validation method was used for the search, with six GM scores as input features. For each configuration, LOSO learning was performed, and the predicted probabilities from the missing data set were pooled. For each configuration, the result was a set of predicted probabilities for all samples in the training set. The pooled probabilities were then used to calculate the APA, and the hyperparameter configurations were ranked using the APA values. The optimal configuration was the one with the highest APA. Table 4 presents the summarized LOSO results for the different algorithms.

[0189] Table 4. LOSO Training Results. The “APA LOSO” column contains LOSO cross-validation statistics for the best-performing hyperparameter configurations of the corresponding models. Among the four classifiers, MLP yielded the best LOSO cross-validation APA results. The winning configuration used the following hyperparameters: two hidden layers, four nodes each, 250 iterations, linear activation, no dropouts, learning rate = 1e-5, batch size = 128, batch normalization, regularization: L1 (penalty = 0.1), and input layer weight initialization using weight priors. Table 5 contains additional performance statistics for pooled LOSO probability estimation using the winning configuration.

[0190] Table 5. Detailed LOSO statistics for the winning neural network classifier. This analysis shows that network performance is highly sensitive to the pseudo-random initialization of network weights. To explore the space of these initial starting points, additional LOSO analysis was performed on the model with the winning hyperparameter configuration, and the network weights were randomly initialized 5000 times using weight priors specified by the chosen configuration. The network was trained and evaluated using the same methods as the initial run, e.g., by pooling the predicted probabilities of all folds in the LOSO run and calculating the APA of the pooled probabilities. The winning seed is the seed corresponding to the model with the highest APA.

[0191] The final locked model was applied to validate clinical data. Specifically, the locked classifier was applied to validate clinical NanoString expression data to compute validated clinical outcomes. This generated three probabilities for each sample: bacteria, virus, and uninfected. The classifier's utility was evaluated using multiple clinically relevant statistics by comparing the predictions to clinically determined diagnoses. Table 6 contains the results.

[0192] Table 6. Performance statistics of the BVN1 classifier applied to independently validated clinical samples (n = 104). In clinical use, key variables of interest when diagnosing patients are the likelihood of bacterial and viral infections. These values ​​are emitted by the top (softmax) layer of the neural network.

[0193] discuss As described above, a machine learning classifier was developed for diagnosing bacterial and viral sepsis in patients suspected of having the condition, and preliminary validation was performed on independent test data. This project faced several major challenges. First, regarding platform transition, the classifier was specifically developed using public domain data and analyzed on various microarray chips. In contrast, the test data was analyzed using NanoString, a platform never before encountered in training. Second, there was significant heterogeneity among the available training datasets. Third, the training sample size was relatively small, especially considering the heterogeneity of the training data. To address these challenges, multiple research directions were applied.

[0194] First, a method for selecting the optimal machine learning model for sepsis classification was investigated. Previous research has shown that standard randomized cross-validation introduces excessive optimism bias due to the highly significant technical and biological heterogeneity in sepsis data. Based on empirical findings and previous research on this subject, the Leave-One-Side Study (LOSO) method was chosen for classifier development.

[0195] Next, the impact of input feature engineering was analyzed. LOSO has consistently favored a custom-engineered input consisting of six geometric mean scores, and therefore used it as the input to the final locked classifier. This is a somewhat unexpected result and warrants further investigation, including the possibility of automatic learning and improvement of feature engineering transformations.

[0196] The probability distributions of the independent test data exhibit a clear trend in the expected direction, as bacterial samples tend to have a high bacterial probability, and so do viral samples. Furthermore, uninfected samples tend to have lower bacterial and viral probabilities. These trends are quantified by favorable pairwise AUROC estimation and class-conditional accuracy. However, significant residual overlap between the distributions has also been noted, which is a focus of ongoing research.

[0197] The platform migration attempt has been successful so far. Nevertheless, to improve clinical performance in testing, future enhancements to our sepsis classifier will include adding NanoString data to the training set.

[0198] This study demonstrates the feasibility of successfully learning a complex sepsis classifier using publicly available data, and then transferring it to previously unseen platforms to analyze previously unseen samples. To our knowledge, this has not been reported in previous sepsis literature, and perhaps not elsewhere in molecular diagnostics.

[0199] Conclusion Multiple instances can be provided for the components, operations, or structures described herein as a single instance. Finally, the boundaries between individual components, operations, and data stores are somewhat arbitrary and specific operations are described within the context of a particular illustrative configuration. Other functional assignments are envisioned and may fall within the scope of the stated implementation(s). In general, structures and functions presented as separate components in the example configuration can be implemented as composite structures or components. Similarly, structures and functions presented as single components can be implemented as separate components. These structures and functions, as well as other variations, modifications, additions, and improvements, fall within the scope of the stated implementation(s).

[0200] It will also be understood that while the terms “first,” “second,” etc., may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used only to distinguish one element from another. For example, without departing from the scope of this disclosure, a first subject may be referred to as a second subject, and similarly, a second subject may be referred to as a first subject. Although both the first and second objects are objects, they are not the same object.

[0201] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in the specification and appended claims, the singular forms “a,” “an,” and “described” are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and covers any and all possible combinations of one or more of the associated listed items. It should be further understood that, when used in this specification, the term “comprises and / or comprising” specifies the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0202] As used herein, depending on the context, the term “if” can be interpreted as “when” or “in response to determination” or “in response to detection”. Similarly, depending on the context, the phrase “if determination” or “if [the stated condition or event] is detected” can be interpreted as “when determination” or “in response to determination” or “when [the stated condition or event] is detected” or “in response to detection”.

[0203] The foregoing description includes example systems, methods, techniques, instruction sequences, and computer program products embodying illustrative embodiments. Numerous specific details have been set forth for purposes of explanation in order to provide an understanding of various embodiments of the subject matter of the invention. However, it will be apparent to those skilled in the art that the subject matter of the invention can be practiced without these specific details. Generally, well-known illustrative examples, protocols, structures, and techniques have not been shown in detail.

[0204] For illustrative purposes, the foregoing description has been given with reference to specific embodiments. However, the illustrative discussion above is not intended to be exhaustive or to limit the embodiments to the precise forms disclosed. Many modifications and variations are possible in light of the above teachings. These embodiments were chosen and described in order to best explain the principles and their practical application, thereby enabling others skilled in the art to best utilize these embodiments and various embodiments with a variety of modifications suited to the intended particular use.

Claims

1. A computer system for assessing the clinical condition of a test subject of a species using prior feature grouping, wherein the prior feature grouping comprises a plurality of modules, each of the plurality of modules comprising a plurality of independent features, the respective feature values ​​of which are associated with the absence, presence, or stage of an independent phenotype related to the clinical condition, the computer system comprising: At least one processor; and A memory storing at least one program executed by the at least one processor, the at least one program comprising instructions for: (A) Obtaining a first training dataset in electronic form, wherein for each of the first plurality of training objects of the species, the first training dataset comprises: (i) a first plurality of feature values ​​obtained in a first form for the independent plurality of features by means of a first assay using biological samples of the respective training object, the first form being one of transcriptomics, proteomics or metabolomics of at least a first module of the plurality of modules; and (ii) an indication of the absence, presence or stage of a first independent phenotype corresponding to the first module in the respective training object; (B) Obtain a second training dataset in electronic form, independent of the first training dataset, wherein for each of the second plurality of training objects of the species, the second training dataset includes: (i) second plurality of feature values ​​obtained, for the independent plurality of features, in a second form identical to the first form of at least the first module, using biological samples of the respective training object by a second determination method different from the first determination method; and (ii) an indication of the absence, presence, or stage of the first independent phenotype in the respective training object. (C) The feature values ​​of features present in at least the first and second training datasets are co-normalized across at least the first and second training datasets to remove batch effects between datasets, thereby calculating the co-normalized feature values ​​of at least a first module of the corresponding training object for each corresponding training object of the first plurality of training objects and for each corresponding training object of the second plurality of training objects, wherein, prior to the co-normalization, the first and second training datasets are mapped to a common variable set; and (D) Training a master classifier for a composite training set to assess the clinical condition of test subjects, wherein for each corresponding training subject in a first plurality of training subjects and for each corresponding training subject in a second plurality of training subjects, the composite training set includes: (i) a summary of the co-normalized feature values ​​of the first module, and (ii) an indication of the absence, presence, or stage of the first independent phenotype in the corresponding training subject.

2. The computer system of claim 1, wherein each corresponding feature in the first module corresponds to a biomarker associated with a first independent phenotype that is statistically significantly richer in objects exhibiting the first independent phenotype than objects that do not exhibit the independent phenotype in the object group of the species.

3. The computer system of claim 1, wherein each corresponding feature in the first module corresponds to a biomarker associated with a first independent phenotype that is statistically less abundant in objects exhibiting the first independent phenotype than objects that do not exhibit the independent phenotype in the object group of the species.

4. The computer system of claim 1, wherein each corresponding feature in the first module corresponds to a biomarker associated with the first independent phenotype by having a feature value that is statistically significantly larger in the group of objects exhibiting the first independent phenotype than in objects not exhibiting the independent phenotype.

5. The computer system of claim 1, wherein each corresponding feature in the first module corresponds to a biomarker associated with the first independent phenotype by having a statistically significantly fewer feature value in the group of objects exhibiting the first independent phenotype compared to objects not exhibiting the independent phenotype.

6. The computer system according to any one of claims 1-5, wherein the feature value of the first feature in one of the plurality of modules is determined by physical measurement of the corresponding component in a biological sample of a reference object.

7. The computer system according to claim 6, wherein the component is a composition.

8. The computer system according to claim 7, wherein the composition is a nucleic acid, protein, or metabolite.

9. The computer system according to any one of claims 1-5, wherein the feature value of the first feature in the plurality of modules is a linear or nonlinear combination of the feature values ​​of each corresponding component in the component group obtained by physical measurement of each corresponding component in a biological sample of a reference object.

10. The computer system of claim 9, wherein each corresponding component in the group of components is a nucleic acid, a protein, or a metabolite.

11. The computer system of claim 1, wherein the species is a human.

12. The computer system of claim 1, wherein the first form is the transcriptome.

13. The computer system of claim 12, wherein the first assay is RNAseq and the second assay is a DNA microarray.

14. The computer system of claim 1, wherein the first form is proteomics.

15. The computer system of claim 1, wherein each corresponding biological sample in the first training dataset and the second training dataset is whole blood of the corresponding training object.

16. The computer system of claim 1, wherein each corresponding biological sample in the first training dataset and the second training dataset belongs to a specified tissue or specified organ of the corresponding training object.

17. The computer system according to claim 1, wherein: The first independent phenotype represents the disease status. The first subset of the first training dataset consists of objects without disease conditions. The first subset of the second training dataset consists of objects without disease conditions. The co-normalization of feature values ​​present in at least the first and second training datasets includes estimating the inter-dataset batch effect between the first and second training datasets using only a first subset of the respective first and second training datasets.

18. The computer system of claim 17, wherein the batch effect between the datasets comprises additive and multiplicative components, and the co-normalization is to solve an ordinary least squares model using eigenvalues ​​across a first subset of each of the first and second training datasets and to reduce the resulting parameters representing the additive and multiplicative components using an empirical Bayesian estimator.

19. The computer system of claim 1, wherein the common normalized eigenvalues ​​present in at least the first and second training datasets, spanning at least the first and second training datasets, include estimating the inter-dataset batch effect between the first and second training datasets.

20. The computer system of claim 19, wherein the batch effect between the datasets comprises additive and multiplicative components, and the co-normalization is performed by solving an ordinary least squares model using eigenvalues ​​across respective first and second training datasets and reducing the resulting parameters representing the additive and multiplicative components using an empirical Bayesian estimator.

21. The computer system of claim 1, wherein the co-normalization of feature values ​​present in at least the first and second training datasets across at least the first and second training datasets comprises normalization using invariant features or quantiles.

22. The computer system according to claim 1, wherein: Each feature in the first and second datasets is a nucleic acid. The first assay is a first-form microarray assay selected from the group consisting of cDNA microarrays, oligonucleotide microarrays, BAC microarrays, and single nucleotide polymorphism (SNP) microarrays. The second assay is a different form of microarray assay selected from the group consisting of cDNA microarrays, oligonucleotide microarrays, BAC microarrays, and SNP microarrays, and is distinct from the first form of microarray assay. The common normalization is either robust multiarray average (RMA) or GeneChip robust multiarray average (GC-RMA).

23. The computer system according to claim 1, wherein the first assay is a DNA microarray, MMChip, protein microarray, peptide microarray, tissue microarray, cell microarray, chemical compound microarray, antibody microarray, glycan array, or reverse-phase protein lysate microarray.

24. The computer system of claim 1, wherein for each corresponding training object in the first plurality of training objects and for each corresponding training object in the second plurality of training objects, the sum of the common normalized eigenvalues ​​of the first module is a measure of the central tendency of the common normalized eigenvalues ​​of the first module in biological samples obtained from the corresponding training object.

25. The computer system of claim 24, wherein for each of the first and second plurality of training objects, the measure of central tendency of the common normalized eigenvalues ​​of the first module is the arithmetic mean, geometric mean, weighted mean, median range, median pivot, three-mean, Winsorized mean, median, or mode of the common normalized eigenvalues ​​of the first module of the corresponding training object.

26. The computer system of claim 1, wherein for each corresponding training object in the first plurality of training objects and for each corresponding training object in the second plurality of training objects, the summation of the common normalized feature values ​​of the first module is the output of the component classifier associated with the first module when the common normalized feature values ​​of the first module in the biological samples obtained from the corresponding training objects are input.

27. The computer system according to claim 26, wherein the component classifier is a neural network algorithm, a support vector machine algorithm, a decision tree algorithm, an unsupervised clustering algorithm, a supervised clustering algorithm, a logistic regression algorithm, a hybrid model, or a hidden Markov model.

28. The computer system according to claim 1, wherein: The at least one procedure further includes instructions for obtaining, in electronic form, a plurality of supplementary training datasets in addition to the first and second training datasets, wherein for each of the plurality of independent training subjects of the species, each of the plurality of supplementary training datasets comprises: (i) a plurality of feature values ​​obtained by independent corresponding assays using biological samples of the corresponding training subject for a first form of a plurality of features of a plurality of modules, and (ii) an indication of the absence, presence, or stage of a corresponding phenotype in the corresponding training subject corresponding to the corresponding module, and The co-normalization (C) further includes co-normalizing the feature values ​​of features present in the corresponding two or more training datasets across a training group comprising a first training dataset, a second training dataset, and multiple additional training datasets, to remove batch effects between datasets, thereby computing the co-normalized feature values ​​for each module across multiple modules for each corresponding training object in each of the multiple training datasets. For each corresponding training object in each training dataset in the training group, the composite training set further includes: (i) a summary of the common normalized feature values ​​of one module in the corresponding training object among the plurality of modules and (ii) an indication of the absence, presence or stage of the corresponding independent phenotype in the corresponding training object.

29. The computer system of claim 28, wherein the plurality of additional training datasets comprises three or more additional training datasets.

30. The computer system of claim 1, wherein the master classifier is a neural network.

31. The computer system of claim 30, wherein the first independent phenotype is identical to the clinical condition.

32. The computer system according to claim 30, wherein For each corresponding training object in the first plurality of training objects of the said species, the first training dataset further includes: (iii) multiple feature values ​​obtained by a first assay using biological samples of the corresponding training subjects from the second module of the multiple modules, and (iv) indications of the absence, presence, or stage of a second independent phenotype in the corresponding training subjects, and For each of the second plurality of training objects of the species, the second training dataset further includes: (iii) a plurality of feature values ​​obtained by a second assay using biological samples of the corresponding training object using the second module, and (iv) an indication of the absence, presence, or stage of a second independent phenotype in the corresponding training object.

33. The computer system according to claim 32, wherein The first and second independent phenotypes were identical to the clinical condition described. Each corresponding feature in the first module is associated with the first independent phenotype by having a feature value that is statistically significantly larger in the group of said species than in the group of objects exhibiting the first independent phenotype compared to objects not exhibiting the independent phenotype. Each corresponding feature in the second module is associated with the first independent phenotype by having a statistically significantly fewer feature value in the group of said species compared to those of objects exhibiting the first independent phenotype.

34. The computer system of claim 32, wherein the first independent phenotype and the second independent phenotype are different.

35. The computer system according to any one of claims 30-34, wherein the neural network is a feedforward artificial neural network.

36. The computer system according to claim 1, wherein the master classifier includes a linear regression algorithm, a penalized linear regression algorithm, a support vector machine algorithm, or a tree-based algorithm.

37. The computer system of claim 36, wherein the master classifier is a tree-based algorithm selected from the group consisting of random forest algorithm and decision tree algorithm.

38. The computer system according to any one of claims 1, 36 or 37, wherein the master classifier is composed of an ensemble of classifiers subjected to an ensemble optimization algorithm.

39. The computer system of claim 38, wherein the integrated optimization algorithm includes AdaBoost, XGBoost, or LightGBM.

40. The computer system according to any one of claims 1, 36 or 37, wherein the master classifier is composed of an ensemble of neural networks.

41. The computer system of claim 1, wherein the at least one program further comprises instructions for: A test dataset is obtained in electronic form, wherein the test dataset includes multiple feature values ​​measured in a first form in a biological sample of the test subject for at least the features in the first module, and The test dataset is input into the main classifier to assess the clinical condition of the test subjects.

42. The computer system of claim 1, wherein the clinical condition is a binary clinical condition.

43. The computer system of claim 1, wherein the clinical condition is a multi-category clinical condition.

44. The computer system of claim 43, wherein the clinical condition comprises three categories of clinical conditions: (i) severe bacterial infection, (ii) severe viral infection, and (iii) non-infectious inflammation.

45. The computer system of claim 43, wherein the master classifier is configured to output the probability of each category in the multi-category clinical condition.

46. ​​The computer system of claim 1, wherein the plurality of modules comprises at least three modules.

47. The computer system of claim 1, wherein the plurality of modules comprises at least six modules.

48. The computer system of claim 1, wherein each of the plurality of modules, each of the plurality of features, comprises at least three features.

49. The computer system of claim 1, wherein each of the plurality of modules, each of the plurality of features, comprises at least five features.

50. The computer system of claim 1, wherein the at least one program further comprises instructions for: Before common normalization (C), a first normalization algorithm is performed on the first training dataset based on each corresponding distribution of the feature values ​​of the corresponding features in the first training dataset, and Before common normalization (C), a second normalization algorithm is performed on the second training dataset based on the corresponding distribution of the feature values ​​of the corresponding features in the second training dataset.

51. The computer system of claim 50, wherein the first normalization algorithm or the second normalization algorithm is a robust multi-array averaging algorithm, a GeneChip RMA algorithm, or a normal exponential convolution algorithm for background correction, followed by a quantile normalization algorithm.

52. A computer system for assessing the clinical condition of a test subject of a species, the computer system comprising: At least one processor; and A memory storing at least one program executed by the at least one processor, the at least one program comprising instructions for: (A) Obtaining a first training dataset in electronic form, wherein for each of the first plurality of training objects of the species, the first training dataset comprises: (i) a first plurality of feature values ​​obtained for a plurality of features using biological samples of the corresponding training object by a first assay, and (ii) an indication of the absence, presence or stage of a first independent phenotype in the corresponding training object, wherein the first independent phenotype represents a disease condition, and wherein a first subset of the first training dataset consists of objects without a disease condition; (B) Obtain a second training dataset in electronic form, independent of the first training dataset, wherein for each of the second plurality of training objects of the species, the second training dataset comprises: (i) second plurality of feature values ​​obtained for a plurality of features using biological samples of the corresponding training object by a second assay different from the first assay, and (ii) an indication of the absence, presence or stage of a first independent phenotype in the corresponding training object, and wherein a first subset of the second training dataset consists of disease-free objects; (C) Commonly normalize the feature values ​​of subsets of at least the first and second training datasets to remove batch effects between datasets, where A subset of the features exists in at least the first and second training datasets. Prior to the common normalization, the first and second training datasets were mapped to a common variable set. The common normalization includes estimating the inter-dataset batch effect between the first and second training datasets using only a first subset of the corresponding first and second training datasets, and The batch effect between the datasets includes additive and multiplicative components, and the co-normalization is achieved by solving an ordinary least squares model using eigenvalues ​​across a first subset of each of the first and second training datasets and reducing the resulting parameters representing the additive and multiplicative components using an empirical Bayesian estimator, thereby using the resulting parameters to compute: the co-normalized eigenvalues ​​of a subset of the plurality of features for each corresponding training object in the first plurality of training objects and for each corresponding training object in the second plurality of training objects; and (D) Training a master classifier for a composite training set to assess the clinical condition of the test subjects, wherein for each corresponding training subject in a first plurality of training subjects and for each corresponding training subject in a second plurality of training subjects, the composite training set includes: (i) common normalized feature values ​​of a subset of the plurality of features, and (ii) an indication of the absence, presence, or stage of a first independent phenotype in the corresponding training subject.

53. The computer system of claim 52, wherein the feature value of the first feature in the subset of the plurality of features is a predetermined physical measurement of a component in the biological sample of the training object.

54. The computer system of claim 53, wherein the component is a composition.

55. The computer system of claim 54, wherein the composition is a nucleic acid, protein, or metabolite.

56. The computer system of claim 52, wherein the first feature in the subset of the plurality of features is a linear or nonlinear combination of the feature values ​​of each corresponding component in the component group obtained by physically measuring each corresponding component in the biological sample of the training object.

57. The computer system of claim 56, wherein each corresponding component in the group of components is a nucleic acid, a protein, or a metabolite.

58. The computer system of claim 52, wherein the species is a human.

59. The computer system of claim 52, wherein the first assay is RNAseq and the second assay is a DNA microarray.

60. The computer system of claim 52, wherein each corresponding biological sample in the first training dataset and the second training dataset is whole blood of the corresponding training object.

61. The computer system of claim 52, wherein each corresponding biological sample in the first training dataset and the second training dataset is a designated tissue or designated organ of the corresponding training object.

62. The computer system according to claim 52, wherein Each feature in the first dataset is a nucleic acid, and The first assay is a first form of microarray assay selected from the group consisting of cDNA microarrays, oligonucleotide microarrays, BAC microarrays, and single nucleotide polymorphism (SNP) microarrays.

63. The computer system according to claim 52, wherein, The first assay is selected from the group consisting of DNA microarrays, MMChips, protein microarrays, peptide microarrays, tissue microarrays, cell microarrays, chemical compound microarrays, antibody microarrays, glycan arrays, or reverse-phase protein lysate microarrays.

64. The computer system according to any one of claims 52-63, wherein The at least one procedure further includes instructions for obtaining, in electronic form, a plurality of supplementary training datasets in addition to the first and second training datasets, wherein for each of the plurality of independent training objects of the species, each of the plurality of supplementary training datasets comprises: (i) using biological samples of the corresponding training object, multiple feature values ​​obtained for multiple independent features, and (ii) indications of the absence, presence, or stage of the corresponding independent phenotype in the corresponding training object of the corresponding module, and The co-normalization (C) further includes co-normalizing the feature values ​​of features present in the corresponding two or more training datasets across a training group comprising a first training dataset, a second training dataset, and multiple additional training datasets, to remove batch effects between datasets, thereby computing the co-normalized feature values ​​for each module across multiple modules for each corresponding training object in each of the multiple training datasets. For each corresponding training object in each of the additional training datasets, the composite training set further includes: (i) common-normalized feature values ​​of features derived from common-normalization (C) and (ii) indications of the absence, presence, or stage of the corresponding independent phenotype in the corresponding training object.

65. The computer system of claim 64, wherein the plurality of additional training datasets comprises three or more additional training datasets.

66. The computer system of claim 52, wherein the master classifier is a neural network.

67. The computer system according to claim 52, wherein The first independent phenotype is the same as the described clinical condition.

68. The computer system of claim 52, wherein the first independent phenotype and the clinical condition are different.

69. The computer system of claim 66, wherein the neural network is a feedforward artificial neural network.

70. The computer system of claim 52, wherein the master classifier comprises a linear regression algorithm, a penalized linear regression algorithm, a support vector machine algorithm, an unsupervised clustering algorithm, a supervised clustering algorithm, or a tree-based algorithm.

71. The computer system of claim 70, wherein the master classifier is a tree-based algorithm selected from the group consisting of random forest algorithm and decision tree algorithm.

72. The computer system of claim 52, wherein the main classifier is composed of an ensemble of classifiers subjected to an ensemble optimization algorithm.

73. The computer system of claim 72, wherein the integrated optimization algorithm includes AdaBoost, XGBoost, or LightGBM.

74. The computer system of claim 52, wherein the at least one program further comprises instructions for: A test dataset is obtained in electronic form, wherein the test dataset includes multiple feature values ​​measured in a first form in a biological sample of the test subject for at least the features in the first module, and The test dataset is input into the main classifier to assess the clinical condition of the test subjects.

75. The computer system of claim 52, wherein the at least one program further comprises instructions for: Before common normalization (C), a first normalization algorithm is performed on the first training dataset based on the corresponding distribution of the feature values ​​of the corresponding features in the first training dataset, and Before common normalization (C), a second normalization algorithm is performed on the second training dataset based on each corresponding distribution of the feature values ​​of the corresponding features in the second training dataset.

76. The computer system of claim 75, wherein the first normalization algorithm or the second normalization algorithm is a robust multi-array averaging algorithm, a GeneChip RMA algorithm, or a normal exponential convolution algorithm for background correction, followed by a quantile normalization algorithm.

77. A computer system for common normalization of datasets, the computer system comprising: At least one processor; and A memory storing at least one program executed by the at least one processor, the at least one program comprising instructions for: (A) Obtain a first training dataset in electronic form, wherein for each of the first plurality of training objects of a species, the first training dataset includes: (i) a first plurality of feature values ​​obtained for a plurality of features using biological samples of the respective training object by a first assay, and (ii) an indication of the absence, presence or stage of a clinical condition in the respective training object, and wherein a first subset of the first training dataset consists of objects that do not present a clinical condition. (B) Obtaining a second training dataset in electronic form, independent of the first training dataset, wherein for each of the second plurality of training objects of the species, the second training dataset includes: (i) a second plurality of feature values ​​obtained for a plurality of features using biological samples of the corresponding training object by a second assay different from the first assay, and (ii) an indication of the absence, presence or stage of the clinical condition in the corresponding training object, and wherein a first subset of the second training dataset consists of objects that do not present the clinical condition; (C) Estimate the initial mean absolute deviation between (i) the vector of the average representation of a subset of features across a first plurality of objects and (ii) the vector of the average representation of a subset of features across a second plurality of objects; (D) Commonly normalize the feature values ​​of subsets of at least the first and second training datasets to remove batch effects between datasets, where A subset of the features exists in at least the first and second training datasets. Prior to the common normalization, the first and second training datasets were mapped to a common variable set. The common normalization includes estimating the inter-dataset batch effect between the first and second training datasets using only a first subset of the corresponding first and second training datasets, and The batch effect between the datasets includes additive and multiplicative components, and the conormalization is to solve the ordinary least squares model for the feature values ​​across the first subset of the respective first and second training datasets and reduce the resulting parameters representing the additive and multiplicative components using an empirical Bayesian estimator, thereby using the resulting parameters to compute: the conormalized feature value of each feature value among the plurality of features for each corresponding training object in the first subset of the training objects; (E) Estimate the common-normalized mean absolute bias between (i) the vector of the average representation of the common-normalized eigenvalues ​​of multiple features across the first training dataset and (ii) the vector of the average representation of a subset of multiple features across the second training dataset; and (F) Repeat the common normalization (D) and estimation (E) until the common normalized mean absolute deviation converges.

78. The computer system of claim 77, wherein the at least one program further comprises instructions for: (G) Train the master classifier on a composite training set to assess the clinical condition of the test subjects, the composite training set comprising: For each of the first plurality of training objects: (i) the common normalized eigenvalues ​​of the plurality of features and (ii) the indication of the absence, presence or stage of the clinical condition in the corresponding training object; and For each of the second plurality of training objects: (i) the second plurality of feature values ​​of the plurality of features and (ii) the indication of the absence, presence or stage of the clinical condition in the corresponding training object.

79. The computer system of claim 77, wherein the common normalized mean absolute deviation converges when the change in the common normalization (D) does not exceed 0.001 in successive iterations.

80. The computer system of claim 77, wherein the common normalization mean absolute deviation converges when the change in the common normalization (D) does not exceed 0.0001 in successive iterations.

81. The computer system of claim 77, wherein the common normalization mean absolute deviation converges when the common normalization (D) has been performed a threshold number of times.

82. The computer system according to claim 81, wherein the threshold number is five times.

83. The computer system according to claim 81, wherein the threshold number is ten times.

84. The computer system of claim 77, wherein the common normalization mean absolute deviation converges when its variation in successive iterations of the common normalization (D) does not exceed a first threshold amount.

85. The computer system of claim 77, wherein the common normalization mean absolute deviation converges when the change in the common normalization (D) during successive iterations does not exceed a first threshold amount or a threshold number of times the common normalization (D) has been performed.

86. The computer system according to claim 85, wherein the first threshold is 0.001, and the threshold count is twenty times.

87. The computer system of claim 85, wherein (C) estimating the initial mean absolute deviation between (i) a vector of average representations of a subset of features across a first plurality of objects and (ii) a vector of average representations of a subset of features across a second plurality of objects comprises setting the initial mean absolute deviation to zero.

88. A computer system for assessing the clinical condition of a test subject of a species using prior feature grouping, wherein the prior feature grouping comprises a plurality of modules, each of the plurality of modules comprising independent plurality of features, the respective feature values ​​of which are associated with the absence, presence, or stage of the clinical condition, the computer system comprising: At least one processor; and A memory storing at least one program executed by the at least one processor, the at least one program comprising instructions for: (A) Obtaining a first training dataset in electronic form, wherein for each of the first plurality of training subjects of the species, the first training dataset comprises: (i) a first plurality of feature values ​​obtained in a first form for the independent plurality of features by means of a first assay using biological samples of the respective training subject, the first form being one of transcriptomics, proteomics or metabolomics of at least a first module of the plurality of modules; and (ii) an indication of the absence, presence or stage of the clinical condition in the respective training subject. (B) Obtain a second training dataset in electronic form, independent of the first training dataset, wherein for each of the second plurality of training subjects of the species, the second training dataset includes: (i) second plurality of feature values ​​obtained, for the independent plurality of features, in a second form identical to the first form of at least the first module, using biological samples of the respective training subject by a second assay different from the first assay; and (ii) an indication of the absence, presence, or stage of the clinical condition in the respective training subject. (C) The feature values ​​of features present in at least the first and second training datasets are co-normalized across at least the first and second training datasets to remove batch effects between datasets, thereby calculating the co-normalized feature values ​​of at least a first module of the corresponding training object for each corresponding training object of the first plurality of training objects and for each corresponding training object of the second plurality of training objects, wherein, prior to the co-normalization, the first and second training datasets are mapped to a common variable set; and (D) Training a master classifier for a composite training set to assess the clinical condition of test subjects, wherein for each corresponding training subject in a first plurality of training subjects and for each corresponding training subject in a second plurality of training subjects, the composite training set includes: (i) a summary of the common normalized feature values ​​of the first module, and (ii) an indication of the absence, presence, or stage of the clinical condition in the corresponding training subject.

89. A computer system for assessing the clinical condition of test subjects of a species using feature grouping, wherein the feature grouping comprises a plurality of modules, each of the plurality of modules comprising independent plurality of features, the respective feature values ​​of which are associated with the absence, presence, or stage of a phenotype related to the clinical condition, the computer system comprising: At least one processor; and A memory storing at least one program executed by the at least one processor, the at least one program comprising instructions for: (A) Obtain a first training dataset in electronic form, wherein for each of the first plurality of training objects of the species, the first training dataset includes: (i) for each of the plurality of modules, multiple feature values ​​of the independent plurality of features obtained from biological samples from the corresponding training object, and (ii) an indication of the absence, presence or stage of the clinical condition in the corresponding training object. (B) For each corresponding training object in the first plurality of training objects, and for each corresponding module in the plurality of modules, the plurality of feature values ​​are summarized to form a corresponding summary of the feature values ​​of the corresponding module of each corresponding training object. (C) Train a master classifier for a composite training set to evaluate the clinical condition of the test subject. For each of the first plurality of training subjects, the composite training set includes: (i) a corresponding summary of the feature values ​​of the corresponding module for each of the plurality of modules, and (ii) an indication of the absence, presence, or stage of the clinical condition in the corresponding training subject. A second training dataset, independent of the first training dataset, is obtained electronically, wherein for each of the second plurality of training objects of the species, the second training dataset includes: (i) for each of the plurality of modules, multiple feature values ​​of the independent plurality of features obtained from biological samples from the corresponding training object; and (ii) an indication of the absence, presence, or stage of the clinical condition in the corresponding training object; and For each corresponding training object in the second plurality of training objects, and for each corresponding module in the plurality of modules, the plurality of feature values ​​are summarized to form a corresponding summary of the feature values ​​of the corresponding module of the corresponding training object, wherein: For each corresponding training object in the first plurality of training objects, and for each corresponding module in the plurality of modules, each corresponding feature value in the plurality of feature values ​​is obtained using a first measurement technique. For each corresponding training object in the second plurality of training objects, and for each corresponding module in the plurality of modules, a second measurement technique different from the first measurement technique is used to obtain each corresponding feature value among the plurality of feature values, and The composite training set further includes, for each corresponding training object in the second plurality of training objects: (i) for each corresponding module in the plurality of modules, a corresponding summary of the feature values ​​of the corresponding module, and (ii) an indication of the absence, presence, or stage of the clinical condition in the corresponding training object; and The feature values ​​of features present in at least the first and second training datasets are co-normalized across at least the first and second training datasets to remove batch effects between datasets, thereby calculating co-normalized feature values ​​of multiple modules of the corresponding training objects for each corresponding training object of the first plurality of training objects and for each corresponding training object of the second plurality of training objects, wherein the first and second training datasets are mapped to a common variable set prior to the co-normalization.

90. The computer system according to claim 89, wherein, For each of the plurality of modules, each of the corresponding independent features is the abundance level of mRNA transcripts measured from biological samples of the corresponding training object.

91. The computer system of claim 89 or 90, wherein for each of the first plurality of training objects, and for each of the plurality of modules, the sum of the feature values ​​of the respective module is a measure of the central tendency of the plurality of feature values ​​of the respective module obtained from biological samples of the respective training object.

92. The computer system of claim 89 or 90, wherein for each of the first plurality of training objects, and for each of the plurality of modules, the summation of the feature values ​​of the corresponding module is the output of a component classifier associated with the first module after inputting a plurality of feature values ​​of the corresponding module obtained from biological samples of the corresponding training object.

93. The computer system according to claim 92, wherein the component classifier is a neural network algorithm, a support vector machine algorithm, a decision tree algorithm, an unsupervised clustering algorithm, a supervised clustering algorithm, a logistic regression algorithm, a hybrid model, or a hidden Markov model.

94. The computer system of claim 89, wherein the first measurement technique is RNAseq and the second measurement technique is a nucleic acid microarray.

95. The computer system according to claim 89, wherein The first phenotype of the corresponding module in the plurality of modules represents the disease status. The first subset of the first training dataset consists of objects without disease conditions. The first subset of the second training dataset consists of objects without disease conditions. The co-normalization of feature values ​​present in at least the first and second training datasets includes estimating the inter-dataset batch effect between the first and second training datasets using only a first subset of the respective first and second training datasets.

96. The computer system of claim 89, wherein the batch effect between the datasets comprises additive and multiplicative components, and the co-normalization is to solve an ordinary least squares model using eigenvalues ​​across a first subset of each of the first and second training datasets and to reduce the resulting parameters representing the additive and multiplicative components using an empirical Bayesian estimator.

97. The computer system of claim 89, wherein the co-normalized eigenvalues ​​present in at least the first and second training datasets, spanning at least the first and second training datasets, include estimating the inter-dataset batch effect between the first and second training datasets.

98. The computer system of claim 89, wherein the batch effect between the datasets comprises additive and multiplicative components, and the co-normalization is performed by solving an ordinary least squares model using eigenvalues ​​across respective first and second training datasets and reducing the resulting parameters representing the additive and multiplicative components using an empirical Bayesian estimator.

99. The computer system of claim 89, wherein the co-normalization of feature values ​​present in at least the first and second training datasets across at least the first and second training datasets comprises normalization using invariant features or quantiles.

100. The computer system according to claim 89, wherein Each feature in the first and second training datasets is a nucleic acid. The first measurement technique is a first-form microarray experiment selected from the group consisting of cDNA microarrays, oligonucleotide microarrays, BAC microarrays, and single nucleotide polymorphism (SNP) microarrays. The second measurement technique is a second form of microarray experiment, selected from the group consisting of cDNA microarrays, oligonucleotide microarrays, BAC microarrays, and SNP microarrays, which differs from the first form of microarray experiment. The common normalization is either robust multiarray average (RMA) or GeneChip robust multiarray average (GC-RMA).

101. The computer system according to any one of claims 89 and 95-99, wherein the first measurement technique and the second measurement technique are independently selected from the group consisting of DNA microarrays, MMChips, protein microarrays, peptide microarrays, tissue microarrays, cell microarrays, chemical compound microarrays, antibody microarrays, glycan arrays, or reverse-phase protein lysate microarrays.

102. The computer system of claim 89, wherein the at least one program further comprises instructions for: A third training dataset is obtained electronically, wherein for each corresponding training object among a third plurality of training objects of the said species, the third training dataset comprises: (i) for each of the plurality of modules, multiple feature values ​​of the independent plurality of features obtained from biological samples from the corresponding training subjects, and (ii) an indication of the absence, presence or stage of the clinical condition in the corresponding training subjects; and For each corresponding training object in the plurality of training objects, and for each corresponding module in the plurality of modules, the plurality of feature values ​​are summarized to form a corresponding summary of the feature values ​​of the corresponding module of the corresponding training object, wherein: For each corresponding training object in the third plurality of training objects, and for each corresponding module in the plurality of modules, a third measurement technique is used to obtain each corresponding feature value in the plurality of feature values, and The composite training set further includes, for each of the third plurality of training objects: (i) for each of the plurality of modules, a corresponding summary of the feature values ​​of the corresponding module, and (ii) an indication of the absence, presence, or stage of the clinical condition in the corresponding training object.

103. The computer system of claim 89, wherein the master classifier is a neural network.

104. The computer system of claim 89, wherein the clinical condition is a binary clinical condition.

105. The computer system of claim 89, wherein the clinical condition is a multi-category clinical condition.

106. The computer system according to claim 105, wherein, For the corresponding module among the plurality of modules, the phenotype associated with the clinical condition is the same as one of the categories of the multi-category clinical condition.

107. The computer system according to claim 105, wherein, For the corresponding module among the plurality of modules, the phenotype associated with the clinical condition is not the same as any category of the multi-category clinical condition.

108. The computer system of claim 105, wherein the clinical condition comprises three categories of clinical conditions: (i) severe bacterial infection, (ii) severe viral infection, and (iii) non-infectious inflammation.

109. The computer system of claim 105, wherein the master classifier is configured to output the probability of each category in the multi-category clinical condition.

110. The computer system of claim 89, wherein the plurality of modules comprises at least six modules.

111. The computer system of claim 89, wherein each of the plurality of modules, each of the plurality of features, comprises at least three features.

112. A computer system for assessing the clinical condition of test subjects of a species using feature grouping, wherein the feature grouping comprises a plurality of modules, each of the plurality of modules comprising independent plurality of features, the respective feature values ​​of which are associated with the absence, presence, or stage of a phenotype related to the clinical condition, the computer system comprising: At least one processor; and A memory storing at least one program executed by the at least one processor, the at least one program comprising instructions for: (A) Obtain a test dataset in electronic form, wherein for each of the multiple modules, the test dataset includes multiple feature values ​​of multiple independent features obtained from biological samples of the test object; (B) For each corresponding module in the multiple modules, the multiple feature values ​​are summarized to form a corresponding summary of the feature values ​​of the corresponding module of the test object; and (C) For each corresponding module among multiple modules, the corresponding summary of feature values ​​of the corresponding module is input into a classifier trained to distinguish two or more categories of clinical conditions, thereby providing a classification of the clinical condition of the test subject, wherein the classifier is trained using a computer system comprising: At least one processor; and A memory storing at least one program executed by the at least one processor, the at least one program comprising instructions for: (A) Obtaining a first training dataset in electronic form, wherein for each of the first plurality of training objects of the species, the first training dataset comprises: (i) a first plurality of feature values ​​obtained in a first form for the independent plurality of features by means of a first assay using biological samples of the respective training object, the first form being one of transcriptomics, proteomics or metabolomics of at least a first module of the plurality of modules; and (ii) an indication of the absence, presence or stage of a first independent phenotype corresponding to the first module in the respective training object; (B) Obtaining a second training dataset in electronic form, independent of the first training dataset, wherein for each of the second plurality of training objects of the species, the second training dataset includes: (i) second plurality of feature values ​​obtained, for the independent plurality of features, in a second form identical to the first form of at least the first module, using biological samples of the respective training object by a second assay different from the first assay; and (ii) an indication of the absence, presence, or stage of the first independent phenotype in the respective training object. (C) The feature values ​​of features present in at least the first and second training datasets are co-normalized across at least the first and second training datasets to remove batch effects between datasets, thereby calculating the co-normalized feature values ​​of at least a first module of the corresponding training object for each corresponding training object of the first plurality of training objects and for each corresponding training object of the second plurality of training objects, wherein, prior to the co-normalization, the first and second training datasets are mapped to a common variable set; and (D) Training a master classifier for a composite training set to assess the clinical condition of test subjects, wherein for each corresponding training subject in a first plurality of training subjects and for each corresponding training subject in a second plurality of training subjects, the composite training set includes: (i) a summary of the co-normalized feature values ​​of the first module, and (ii) an indication of the absence, presence, or stage of the first independent phenotype in the corresponding training subject.

Images