Methods for predicting and generating mixtures of microbiota samples

Abstract

Prediction of mixtures of complex microbial communities involves, for example, matrix-based linear prediction, modified using an interaction model, for example a matrix, learned from a reference true mixture profile and a corresponding reference linear prediction profile. Inverse prediction allows for the determination of the mixture of samples that will be generated given a target mixture profile.

Description

[Technical field]

[0001] The present invention relates to the mixing or "pooling" of complex communities of microorganisms, i.e., microbiota, and more particularly to methods and apparatus that use learning models, e.g., matrix-based predictive models, that link the individual profiles or compositions of initial microbiota samples with the resulting mixture profile. [Background technology]

[0002] Complex communities of microorganisms, also known as the microbiome, play important roles in health and disease. In particular, it has been discovered that administering or transplanting complex communities of microorganisms, for example through fecal microbiota transplantation (FMT), could potentially treat infections and diseases.

[0003] When administering or transplanting complex communities of microorganisms, it is important that the administered or transplanted sample has an appropriate profile in terms of microbial viability and diversity, including bacteria, archaea, viruses, phages, protozoa and / or fungi.

[0004] Many administration and implantation methods are empirical and do not take special precautions to ensure the diversity of microorganisms present in the samples used or to optimize microbial viability.

[0005] Furthermore, samples taken from donors may not provide a satisfactory profile of the complex microbial community for efficient treatment.

[0006] Therefore, mixtures of complex microbial community samples from multiple donors have been considered to increase the diversity of samples that can be used as inocula for administration and transplantation.

[0007] To test various mixtures, samples are actually mixed randomly and the resulting products are sequenced to obtain a final mixture profile from which hardening and processing properties can be inferred. This test-based approach has several drawbacks, notably the difficulty in obtaining samples from donors, the consumption of scarce materials, and the time required for sequence analysis, which can take weeks to complete.

[0008] Prediction of mixture composition, i.e., prediction of the mixture product profile, has been considered in this way.

[0009] A simple way to predict the mixture composition from the individual profiles of a complex community of microorganisms used as starting material is to apply linear prediction for each profiling feature, for example by summing the relative abundance of said profiling features in the individual profiles after weighting with the proportion of the corresponding complex community in the mixture.

[0010] However, some shift or deviation was observed between such linearly predicted profiles and the true profiles (obtained by profiling the resulting mixture products), and therefore linear prediction of the profiles is considered to be an erroneous approach.

[0011] Our hypothesis is that the shift may result from some important and rapid adjustments due to microbial interactions in a shared environment, e.g., adaptation to new conditions in the shared environment or competition between microorganisms.

[0012] Thus, accurate predictions need to be made, for example, to perform pooling in a manner that ensures delivery of the correct conjugate product composition with the expected therapeutic effect. Summary of the Invention [Problem to be solved by the invention]

[0013] The present invention seeks to overcome some of the aforementioned concerns by computer-aided design of these shifts in predicting mixture compositions, particularly with the goal of driving, controlling or directing the actual mixing of microbial samples in administration or implantation procedures.

[0014] In this regard, the present invention proposes a computer-aided method for predicting the mixture composition resulting from the mixing of complex microbial community samples belonging to an initial sample collection, the method comprising: Predicting intermediate mixture profiles for selected complex microbial community sample mixtures using linear methods; and modifying the intermediate mixture profile to a predictive mixture profile using an interaction model learned from a reference linear predictive mixture profile and a corresponding reference true mixture profile.

[0015] In particular, the predicted mixing profile can be used to control the actual removal and mixing of a complex microbial community sample from an initial sample collection to obtain a mixed result product. Sample removal may simply mean taking or obtaining an appropriate and sufficient amount of sample from the initial collection. The removal may be manual by an operator or automated by a controlled robot.

[0016] It has also been found that the present invention provides a method for producing a complex microbial community product, the method comprising: selecting a complex microbial community sample from the initial sample collection; predicting a mixture profile resulting from a mixture of the selected samples using the prediction method described above; Comparing the predicted mixture profile to a selection criterion, such as sufficient presence of a taxon of interest or any target mixture profile; and actually taking and mixing the selected samples depending on the results of the comparison to obtain a mixed result product. The selected samples are preferably mixed using the relative abundances used in the prediction.

[0017] Of course, if the comparison is inaccurate, no actual mixing can be performed and other sample selections can be made to predict other mixing profiles using the teachings of the present invention. In this manner, multiple sets of selected samples can be considered in succession. For each set, a using and comparing step is then performed, and the actual collection and mixing is performed depending on the results of the comparison.

[0018] Conversely, the present invention also proposes an inverse approach by a computer-assisted method to determine a target set of complex microbial community samples in an initial sample collection given a target mixture profile representing a target mixture result product, the method comprising: selecting a candidate set of complex microbial community samples from the initial sample collection; For each selected candidate set, predicting a mixture profile resulting from a mixture of samples of the selected candidate set using the prediction method described above; comparing the predicted mixture profile to the target mixture profile and selecting one candidate set as the target set.

[0019] A target mixture profile may be general, i.e. quantifying each of the profile features considered, or it may be specific to one or several profile features, e.g. defining several feature specifications such as the presence or absence of one (or several) profile features and / or their relative abundance or amount or amount range, or it may define a minimum level of diversity, e.g. in terms of the number of profile features having a minimum relative abundance. Thus, a target mixture profile may be a set of profiles examining various possible values ​​for a given feature specification.

[0020] The target sample set can then be used to control the actual collection and mixing of complex microbial community samples from the initial sample collection, and obtain a function of the mixture result product of the target mixture profile (which may have the target mixture profile or be close to a given approximation).

[0021] It has also been found that the present invention provides a method for producing a complex microbial community product having a target mixture profile representative of a target mixture result product, the method comprising: selecting a target set of complex microbial community samples belonging to an initial sample collection given a target mixture profile using the determination method; and actually taking and mixing samples of the selected target set to obtain a mixed result product.

[0022] The present invention makes it possible to instantly simulate various mixture compositions at low cost, in particular without consuming real material (samples of the initial sample collection).

[0023] Furthermore, an efficient set of complex microbial community samples can be found in terms of obtaining a mixture result product that meets the mixture criteria, for example, a target group profile or composition adapted to treat a disease.

[0024] Thus, depending on the needs of the intended use (therapeutic, preventive, environmental, etc.), pooling policies can be established prior to production routines.

[0025] The resulting mixture product can be administered or implanted in humans or animals, in plants as fertilizer, or in environmental media such as water, soil, subsurface materials, etc. for the treatment of contamination by bioremediation.

[0026] Preferably, a Microbiome Ecosystem Therapy product can be produced using the above method.

[0027] In this regard, the present invention also provides a computer device including at least one microprocessor configured to perform the steps of any of the above methods, and the computer device may thus be configured to emit signals to control a mixing device that actually collects and mixes the complex microbial community sample from the initial sample collection to obtain a mixed result product.

[0028] Optional features of embodiments of the invention are defined in the accompanying claims, some of which are described below with reference to a method but which may also be replaced by device features.

[0029] In some embodiments, predicting the intermediate mixture profile comprises calculating a matrix product between a first matrix that defines the mixture in terms of the proportions of the complex microbial community samples of the initial sample collection and a second matrix that defines the individual profiles of the complex microbial community samples, the second matrix being hereinafter denoted A and defined by the available initial sample collection.

[0030] In some embodiments, modifying the intermediate mixture profile includes computing a matrix product between a matrix representing the intermediate mixture profile and a square interaction matrix of a learned interaction model, where the interaction model may be a square interaction matrix learned from a baseline linear predictive mixture profile and a corresponding baseline true mixture profile.

[0031] Using a matrix to make a prediction of a sample mixture advantageously takes into account a large number of profile features and allows for rapid calculations to obtain one or more predicted mixture profiles for the mixture result product or products.

[0032] In one embodiment, the prediction method further comprises clipping each negative value in the predicted mixture profile, i.e., setting negative values ​​to 0, in order to correct theoretical predictions (e.g., negative relative abundances) to natural realities.

[0033] In some embodiments, the prediction method further comprises normalizing the sum of the relative abundances of the profiling features that define the predicted mixture profile to 1. Again, the goal is to normalize the theoretical prediction to natural reality, which is to have true relative abundances, whose sum represents the entire composition.

[0034] Also, features that are not present in the selected sample (which is a mixture) should not be present in the predicted mixture profile, so for features that are not present in the selected sample, their nonzero abundance in the predicted mixture profile is set to zero.

[0035] In some embodiments of the inverse approach, determining the set of samples may include determining the relative abundance of each sample in the set. In other words, the inverse prediction aims to obtain the relative proportions of complex microbial community samples that are mixed together.

[0036] In another embodiment of the inverse approach, comparing the predicted mixture profile to the target mixture profile includes calculating a distance between each predicted mixture profile and the target mixture profile and selecting the candidate set having the shortest distance as the target set.

[0037] In some embodiments, a profile of a complex community of microorganisms (sample or mixture) comprises the relative abundance of a profiling feature in the complex community of microorganisms.

[0038] In a specific embodiment, the relative abundance is representative of the mass or volume fraction of a profile feature in a complex community of microorganisms.

[0039] In some embodiments, the profile features forming the profile of a complex community of microorganisms include one or more features from taxonomic group, gene, antibiotic resistance gene, function, metabolite trait, metabolite and protein production, preferably taxonomic group.

[0040] In some embodiments, individual profiles of complex microbial community samples are obtained using profiling techniques such as 16S rRNA gene amplicon sequencing, NGS shotgun sequencing, non-16S rRNA gene-based amplicon sequencing, NGS amplicon-based targeted sequencing, phylochip-based profiling, whole metagenomic sequencing (WMS), polymerase chain reaction (PCR) identification, mass spectrometry (such as LC / MS type or GC / MS type), near-infrared (NIR) spectroscopy, nuclear magnetic resonance (NMR) spectroscopy, etc., preferably using 16S rRNA gene amplicon sequencing or NGS.

[0041] In some embodiments, the profile of a complex community of microorganisms defines profiling characteristics for one or more microorganisms present in the complex community of microorganisms from bacteria, archaea, viruses, phages, protozoa and fungi, preferably bacteria and / or archaea.

[0042] In some embodiments, the profile of a complex community of microorganisms defines profile features that define the relative abundance of the microorganisms considered at one or more taxonomic levels from strain, species, genus, family, order, preferably at one or more taxonomic levels from genus, family, order.

[0043] In some embodiments, the profile of a complex microbial community comprises the relative abundance in the complex microbial community of bacterial and / or archaeal taxa considered at the taxonomic level of genus, family and order.

[0044] In some embodiments, a complex microbial community profile comprises the relative abundance of bacterial and / or archaeal taxa in a complex microbial community defined by the presence or absence or expression of specific genes and / or functions (e.g., production of butyrate, production of antibiotic resistance genes, production of enzymes such as organophosphate hydrolases, phosphodiesterases, superoxide dismutases, production of antimicrobial peptides, production of organophosphate hydrolases or other enzymes useful in bioremediation processes, ...).

[0045] In some embodiments, the initial sample collection comprises a sample selected from the group consisting of a raw complex microbial community sample, an artificial / engineered complex microbial community sample, an artificial complex microbial community sample (e.g., a bacterial consortium obtained by mixing isolated strains), and a virtual complex microbial community sample.

[0046] In some embodiments, the initial sample collection includes one or more of fecal, skin, buccal, vaginal, nasal, tumor, human, animal, plant, water, soil samples, etc. For example, it may include one or more fecal samples obtained from at least one donor, and preferably from at least two donors.

[0047] In some embodiments, the interaction model (e.g., a square interaction matrix) is a reference linear predictive mixture profile and a reference predictive mixture profile obtained from the interaction model (preferably by performing a matrix multiplication with a square interaction matrix); the corresponding baseline true mixture profile, The method is obtained using machine learning to minimize the functional expression of the difference between

[0048] The reference data (here, profiles) are known as the training data for the machine learning process. They are searched for to minimize the error between the matrix-based predicted profile and the corresponding true profile, sometimes subject to a regularization term.

[0049] At this point, the formula can add a regularization term, preferably a ridge-based regularization term, to the difference.

[0050] In certain embodiments, the regularization term comprises the difference between a square interaction matrix of the interaction model and an identity matrix.

[0051] Regularization tends to penalize model solutions that deviate too far from the identity. Indeed, model solutions that deviate too far from the identity will be far from biological reality, since interactions between microorganisms in a mixture are not expected to be significant. The regularization term therefore avoids obtaining such accidental solutions (theoretical solutions resulting from a particular training dataset).

[0052] In some embodiments, negative values ​​of the relative abundance of the profiling feature in the baseline predicted mixture profile are removed prior to minimizing the equation.

[0053] In some embodiments, the method further comprises, prior to minimizing the equation, normalizing the sum of the relative abundances of the profiling features that define one of the baseline predictive mixture profiles to 1. Preferably, some or all of the baseline predictive mixture profiles are individually normalized to 1, as desired.

[0054] In some embodiments, a reference linear predicted mixture profile is predicted using linear techniques from the individual profiles of complex microbial community samples that are mixed together to generate a reference mixture product, and a corresponding reference true mixture profile is obtained from profiling (e.g., sequencing or 16S rRNA gene amplicon sequencing) of the reference mixture product.

[0055] In some embodiments related to the generation methods, the selection criteria include one or more of a diversity criterion representative of increasing diversity of profiling features, a minimum or maximum relative abundance of one or more profiling features, a non-zero relative abundance of one or more specific profiling features or a minimum number of profiling features, a relative ratio between at least two profiling features, a closeness (or similarity, such as a minimum distance) to a target mixture profile.

[0056] In some embodiments, the selected complex microbial community sample is a virtual sample, and the method further comprises actually generating a complex microbial community sample corresponding to the selected virtual sample from the isolated strains and / or complex microbial community samples. This advantageously allows the pooling method to be defined in advance without consuming or having samples yet. The bacterial consortium identified as useful for generating the desired mixed result product can be generated by simply mixing the isolated strains. Similarly, the sample identified as useful for generating the desired mixed result product can be generated by mixing one or more of the isolated strains with one or more samples, resulting in an artificial sample enriched in the desired strains.

[0057] Another aspect of the invention relates to a non-transitory computer readable medium storing a program which, when executed by a microprocessor or computer system within an apparatus, causes the apparatus to carry out any of the methods defined above.

[0058] At least some of the methods according to the invention may be computer-implemented. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, microcode, etc.), or an embodiment combining software and hardware aspects, all of which may be referred to generally herein as a "circuit," "module," or "system." Furthermore, the present invention may take the form of a computer program product embodied in any tangible medium of expression having computer usable program code embodied in the medium.

[0059] Since the present invention can be implemented in software, the present invention can be embodied as computer readable code for provision to a programmable device on any suitable carrier medium. Tangible carrier media can include storage media such as hard disk drives, magnetic tape devices, or solid-state memory devices. Transient carrier media can include signals such as electrical, electronic, optical, acoustic, magnetic, or electromagnetic signals, e.g., microwave and RF signals. [Brief description of the drawings]

[0060] [Figure 1] FIG. 1 is a diagram illustrating a complex microbial community mixing platform for implementing an embodiment of the present invention. [Figure 1a] Figure 1a shows the behavior of the error measure depending on the hyperparameter of the regularization term we model. [Diagram 2] FIG. 2 illustrates, by means of a flow chart, the general steps of generating a blending result product, including prediction of a blending profile, according to an embodiment of the invention. [Diagram 3] FIG. 3 illustrates, by way of a flow chart, the general steps of determining and then generating a blending result product given a target blending profile, according to an embodiment of the present invention. [Figure 4] FIG. 4 is a schematic diagram of a computing device according to an embodiment of the present invention. [Figure 5a] FIG. 5a shows the results of a first experiment of the invention, based on a mixture of natural complex community samples of microorganisms. [Figure 5b] FIG. 5b shows the results of a first experiment of the invention, based on a mixture of natural complex community samples of microorganisms. [Figure 5c] FIG. 5c shows the results of a first experiment of the invention, based on a mixture of natural complex community samples of microorganisms. [Figure 6a] FIG. 6a shows the results of another experiment of the present invention, based on a mixture of fermentation complex community samples of microorganisms. [Figure 6b] FIG. 6b shows the results of another experiment of the present invention, based on a mixture of fermentation complex community samples of microorganisms. [Figure 6c] FIG. 6c shows the results of another experiment of the present invention, based on a mixture of fermentation complex community samples of microorganisms. [Figure 7a] FIG. 7a shows the results of yet another experiment of the present invention in which natural and fermented samples were mixed. [Figure 7b] FIG. 7b shows the results of yet another experiment of the present invention in which natural and fermented samples were mixed. [Figure 8] FIG. 8 shows a collection of sample profiles used in a second experiment of the present invention. [Figure 9a] FIG. 9a shows the results of a second experiment that attempts to find a mixture composition to obtain a mixture product sufficiently close to the target mixture product. [Figure 9b] FIG. 9b shows the results of a second experiment that attempts to find a mixture composition to obtain a mixture product sufficiently close to the target mixture product. [Figure 10] FIG. 10 shows the similarity of the actual mixtures or the actual and predicted mixtures when comparing the target mixtures and the best predicted mixtures shown in FIGS. 9a and 9b. [Figure 11] FIG. 11 shows PCA based on genomic relative abundances obtained from NGS shotgun sequencing of samples from experiment 3. [Figure 12] FIG. 12 shows the PCA-based approach used in Experiment 3. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0061] The present invention relates to mixtures or "pools" of complex communities of microorganisms, i.e. "microbiota" or "microbiota samples", and more particularly to methods and apparatus that use learned predictive models to correlate the individual profiles or compositions of initial complex microbial community samples with the resulting mixture profiles.

[0062] As used herein, the terms "microbiota", "microbiota composition" and "complex community of microorganisms" can be used interchangeably to refer to a microbial population that includes a large number of different species of microorganisms that may coexist and interact. Microorganisms that may be present in a complex community of microorganisms include yeast, bacteria, archaea, viruses, fungi, algae, phages, and any protozoa of different origins, such as soil-derived, water-derived, plant-derived, animal-derived, and human-derived.

[0063] Microbial communities according to the present invention include naturally occurring complex communities of microorganisms (e.g., gut microbiota, i.e., the population of microorganisms that live in the intestines of animals), as well as "artificial complex communities of microorganisms", i.e., complex communities obtained from transformation processes such as the addition of isolated beneficial strains, treatment to remove potentially harmful microorganisms (e.g., using rare-cutting endonucleases that target genes specific to pathogens), and growth in culture under specific conditions (e.g., fermentation in a suitable medium). As used herein, "isolated beneficial strains" refers to naturally occurring strains known to have beneficial effects under specific conditions (e.g., Akkermansia muciniphila), as well as genetically modified strains, including strains in which potentially harmful genes have been knocked out (e.g., using rare-cutting endonucleases such as Cas9) and strains in which transgenes have been introduced (e.g., using bacteriophages or CRISPR systems).

[0064] The microbial and microbiota complexes according to the present invention include "raw" or "native" complex flora or microbiota, i.e., complex flora or microbiota obtained directly from a source, a donor or donors, without any downstream processing, and "processed complex flora of microorganisms", including engineered complex flora or microbiota, and one or more naturally occurring raw complex microbiota (e.g., complex communities or microbiota that have been filtered, frozen, thawed, and / or lyophilized, and / or extracted, isolated, or separated from an initial matrix by techniques well known to those skilled in the art, such as those described in WO2016 / 170285 and WO2017 / 103550).

[0065] Here, the expressions "sample", "complex microbial community sample" and "microbiota sample" can be used interchangeably and refer to an initial complex community or microbiota of microorganisms within the meaning of the present invention, i.e. available as a mixture.

[0066] As used herein, the term "microbial community ecosystem therapy product" refers to any composition that includes a complex of microbial communities (naturally occurring or artificially created, natural or processed), provided that it is in a form suitable for administration to an individual in need thereof. Microbial community ecosystem therapy aims to modify an individual's microbiome to obtain a health benefit (e.g., prevent or alleviate symptoms of disease, increase the likelihood that an individual will respond to treatment, etc.). Typically, microbial community ecosystem therapy is performed by replacing at least a portion of a dysfunctional and / or damaged ecosystem in a subject in need thereof with a different complex community of microorganisms. Microbial community ecosystem therapy includes fecal microbiota transplantation (FMT). As used herein, unless otherwise noted, the term "FMT" is used broadly to refer to any type of microbiota ecosystem therapy.

[0067] As shown in Figure 1, which illustrates a complex microbial community mixing platform 1 implementing an embodiment of the present invention, samples 100 are available through an initial sample bank or collection 10. Although a single collection or bank is shown, samples may be stored in multiple sub-banks that collectively form the collection or bank 10.

[0068] A sample of the present invention may comprise or consist of microorganisms derived from one or more sources and / or one or more donors 101 .

[0069] The sample of the present invention is Single source, At least two sources Single donor, At least two donors, Single source and single donor, Single source and at least two donors, At least two sources and a single donor, or At least two sources and at least two donors; can be obtained from

[0070] As used herein, the term "source" refers to any environment from which a sample is derived, such as soil, water, a part of a plant, a part of an animal's body or body fluid, a part of a human body or body fluid, etc. In humans and animals, the source may be any part of the body (e.g., skin, nasal mucosa, etc.) or a body fluid such as the contents of the intestine (e.g., a stool sample).

[0071] As used herein, the term "donor" refers to a plant, a physical location (a source such as soil or water), an animal, or a human, preferably a human.

[0072] Donors can be pre-selected according to methods and criteria as described in the prior art, for example, WO2019 / 171012 A1.

[0073] In this example, the samples designated 100d, 100e, 100f, and 100g are raw microbial complexes or microbial communities obtained directly from a donor or donors without any post-processing.

[0074] The other samples, labeled 100a, 100b, 100c, are "processed samples", i.e., artificial complex microbial communities resulting from processing or subsequent processing or transformation of one or more natural raw complex communities. As mentioned above, processing includes filtration, centrifugation, fermentation, freezing, lyophilization of initial complex communities, and even mixing of initial complex communities, but also treatments aimed at isolating spores and spore-forming bacteria, such as the use of ethanol, chloroform, heat, etc.

[0075] As shown, the initial composite assemblage may be one of samples 100d, 100e, 100f, 100g from the initial sample collection 10 or may be an external sample 99.

[0076] The initial sample collection 10 may include one or more samples from any source (fecal, skin, nasal, buccal, vaginal, tumor...) of any origin (human, animal, plant, soil,...), and preferably includes one or more fecal samples from at least one donor, more preferably from at least two donors.

[0077] According to certain embodiments, the sample of collection 10 comprises a fecal sample.

[0078] The fecal sample collected from the donor can be controlled according to the methods and qualitative criteria described in the prior art, for example, in WO2019 / 171012 A1. For example, the qualitative criteria for the sample can include the sample being consistent between 1 and 6 on the Bristol scale, the absence of blood and urine in the sample, and / or the absence of certain bacteria, parasites and / or viruses, as described in WO2019 / 171012 A1.

[0079] The fecal sample can be collected according to any method described in the prior art, for example, WO2016 / 170285 A1, WO2017 / 103550 A1 and / or WO2019 / 171012 A1. Preferably, the sample is collected and then placed under anaerobic conditions. For example, the sample may be placed in an oxygen-tight collection device within 5 minutes of collection, as described in WO2016 / 170285 A1, WO2017 / 103550 A1 and / or WO2019 / 171012 A1.

[0080] The sample can be prepared according to the methods described in the prior art, for example, in WO2016 / 170285 A1, WO2017 / 103550 A1, and / or WO2019 / 171012 A1.

[0081] All of the samples 100a-100g shown in the figure are actual samples stored in at least one bank.

[0082] Samples 100y-100z, represented by dotted lines, are theoretical samples that are not actually taken from or produced by donors, and therefore are not actually stored in a preservation bank or bank 10. As explained below, these "virtual" samples 100y-100z are drawn to illustrate a theoretical composite community profile 110z imagined by an entity, such as a computer, an operator, a researcher, or the like.

[0083] The initial sample collection 10 may include only natural samples 100d-100g, may include only processed samples 100a-100c, may include only virtual samples 100y-100z, or may include any combination thereof.

[0084] The first objective of the present invention is to predict the mixture composition resulting from the mixing of samples 100a to 100z belonging to the initial sample collection 10. The prediction includes: Predicting intermediate mixing profiles for the mixture of selected complex microbial community samples using linear methods; and and correcting the intermediate mixture profile to a predictive mixture profile using an interaction model learned from a reference linear predictive mixture profile and a corresponding reference true mixture profile. The interaction model is preferably a squared interaction matrix learned from a reference linear predictive mixture profile and a corresponding reference true mixture profile.

[0085] The inventors have surprisingly discovered that methods based on learned interaction models, and more particularly matrices, provide accurate prediction results once the interaction models or matrices are trained, thus giving relevant hints to the final product without consuming material from the initial sample collection.

[0086] Because the predictions can be performed computationally, predicted mixture profiles can be obtained quickly despite the large number of mixtures to predict, the large number of samples available in the initial sample collection 10, and the large number of features to profile the complex community of microorganisms (samples and mixtures).

[0087] A second object of the present invention relates to the reverse operation, where a target set of samples is determined from an initial sample collection 10 given a target mixture profile representative of a target mixture result product. The target mixture result product may represent, for example, a desired complex community of microorganisms having curative properties with respect to a disease or infection.

[0088] For the reverse operation, sequentially selecting a candidate set of complex microbial community samples from the initial sample collection; for each successively selected candidate set, predicting, using the prediction method described above, a mixture profile resulting from a mixture of samples of the selected candidate set; comparing the predicted mixture profile to a target mixture profile and selecting one candidate set as a target set; Includes:

[0089] Both the predictive and inverse operations can be used to actually generate the blended result product.

[0090] For the predictive operation, the predicted mixture profile is compared to a selection criterion, e.g., sufficient presence of the taxon of interest. Then, depending on the result of the comparison, selected samples are taken and actually mixed to obtain a mixed result product. In this way, the predicted mixture profile can be used to control the actual collection and mixing of samples from the initial sample collection to obtain a mixed result product.

[0091] The selection criteria can be set as a function of desired properties for the resulting blended product.

[0092] Such approaches, including interaction model-based prediction and matrix-based prediction, are illustrated in Figure 2 and described in further detail below.

[0093] With regard to the inverse operation, for example, a target set of samples belonging to an initial sample collection is selected using the inverse operation given a target mixing profile corresponding to a mixing result product having desired hardening properties. Samples of the target set are then selected and actually mixed to obtain the desired mixing result product. In this way, the target set of samples determined by the inverse prediction method can be used to control the actual collection and mixing of samples from the initial sample collection to obtain a function of the mixing result product of the target mixing profile.

[0094] An embodiment of such a reverse approach using a target mixture profile is described below with reference to FIG.

[0095] "Mixing" refers to the actual mixing of samples, which results in a new complex community of microorganisms or a new microbiota composition. The result can be used for administration or transplantation as described above, and is therefore also called the mixed result product. The mixed result product can be used, for example, as a medium for inoculation of FMT.

[0096] "Profile" refers to a description of a microbial complex community composition or microbiome composition (be it a sample or a mixture). For example, a profile defines the relative abundance of a profiling feature in a complex community or microbiome composition. "Relative" means that the sum of the abundances is equal to 1. The relative abundance can be expressed as the mass (or weight) or volume fraction of the profiling feature in a complex community of microorganisms.

[0097] Depending on the application involved (e.g. in the therapeutic field depending on the disease of interest, in the bioremediation field depending on the pollutant to be removed), the profiling features may be of different types. They are usually selected from the group comprising taxa, genes, antibiotic resistance genes, functions, metabolite traits, metabolites and protein production. A profile can mix different types of profiling features, for example taxa and antibiotic resistance genes. In a particular embodiment, only taxa are considered to profile a complex community of microorganisms.

[0098] Functions can describe the known action of a protein or protein family (defined phylogenetically, e.g., KEGG KO, NCBI COGs, or Enzyme Commission Number databases), or they can define a metabolic context (e.g., BiGG model databases at the reaction level or KEGG pathways at the metabolic pathway level), with specialized databases, e.g., the CaZy database, a catalogue of carbohydrate-active enzymes. These functional categories (or a combination of them) can be used as features in matrix models.

[0099] KEGG stands for "Kyoto Encyclopedia of Genes and Genomes," KO stands for "KEGG Orthology," NCBI stands for "National Center for Biotechnology Information," COG stands for "Cluster of Orthologous Groups," and BiGG stands for "Biochemical Genetic and Genomic."

[0100] To obtain complex community profiles, various profiling techniques are known, such as 16S rRNA gene amplicon (i.e., metagenomic) sequencing, NGS shotgun sequencing, non-16S rRNA gene amplicon sequencing, NGS amplicon-based targeted sequencing, 18S / ITS gene sequencing, metagenomic sequencing, phylochip-based profiling, identification by polymerase chain reaction (PCR), mass spectrometry (e.g., LC / MS and GC / MS types), near-infrared (NIR) spectroscopy, and nuclear magnetic resonance (NMR) spectroscopy.

[0101] As shown in Figure 1, a profiler (or sequencer) 12 is preferably used to provide profiles, e.g. 16S sequences, of actual samples 100a-100g. The corresponding individual profiles thus obtained are referred to as 110a-110g and form an initial profile collection or bank 11. Of course, 16S rRNA sequencing is not mandatory and other methods, as defined above, can be used alone or in combination to provide the profiles 110.

[0102] The individual profiles are converted in the same format, whatever the sequencing technology used, into a matrix or vector a x The coefficients a of each profile "x" are stored in the computer memory (not shown).x (j) indicates the relative abundance of profiling feature "j" in the sample considered.

[0103] As described above, some individual profiles 110z may be represented by, for example, a coefficient a that represents the relative abundance of a profiling feature "j" in a theoretical sample. x It may be constructed artificially by an operator by defining (i).

[0104] Therefore, the initial profile collection 11 may include only individual profiles 110d-110g corresponding to the natural samples 100d-100g, or only individual profiles 110a-110c corresponding to the processed samples 100a-100c, or only virtual profiles 110y-110z corresponding to the virtual samples 100y-100z, or any combination thereof.

[0105] Other profiles that are subsequently treated (eg so-called intermediate or mixed profiles) follow the same profile format, eg a vector of the same profiling features "j" in the same order.

[0106] Preferably, the abundance profile of bacteria is obtained, which means that the profile defines the relative abundance of the profiling feature for bacteria.More generally, the profile of a complex biota of microorganisms can define the profiling feature for one or more microorganisms (bacteria, archaea, viruses, phages, protozoa and fungi) present in the complex biota, preferably for bacteria and / or archaea.Of course, the profiling features in the same profile can be for different microorganisms, as listed above.

[0107] Preferably, a genus-based bacterial abundance profile is obtained, which means that the profiling feature describes the relative abundance of bacteria at the genus level in a complex biota of microorganisms.More generally, the profile of a complex biota of microorganisms can define profiling features that define the relative abundance of the microorganisms considered at one or more taxonomic levels from strain, species, genus, family and order, preferably at one or more taxonomic levels from genus, family and order.

[0108] The prediction and inverse operations are performed by the pool prediction module 13 under the control of module 14. Module 14 is referred to as the "test and determination module" or "determination module" and drives platform 1 for the purpose of predicting a mixture profile and / or determining a sample set given a target mixture profile and / or generating at least one mixture result product.

[0109] The modules 13 and 14 are preferably implemented by a computer having input / output interfaces (keyboard, mouse, screen, etc.) to allow an operator to interact with the platform 1 .

[0110] As shown, the pooled predictor 13 is matrix-based and includes two steps to predict the resulting mixture profile from the initial profiles of the mixed samples.

[0111] The matrix A defines the individual profiles of all samples available in the collection 10. It is formed by the individual profiles obtained by, or at least from, the profiler or sequencer 12. Furthermore, any virtual individual profiles are also added to the matrix.

[0112] Preferably,

number

[0113] The square matrix W is the interaction matrix defined above that models the interactions between microorganisms. A detailed description of the modeling matrix W (including the learning method) is given below. The interaction matrix aims to represent the nonlinear interactions between the various profiling features of the samples when the latter are mixed together.

[0114] The prediction operation includes a first matrix-based step of predicting, using matrix A, an intermediate mixture profile formed by matrix I for at least one mixture of selected samples, where I=P*A, where P is a matrix representing at least one mixture of selected samples from the collection 10.

[0115] The matrix P can define each mixture in terms of mass or volume ratios of samples from the initial sample collection.

[0116] for example,

number

number

[0117] Matrix-based approaches advantageously allow for a varying number of mixtures to be predicted jointly, with each row of P defining a mixture to predict (so in the above example, "t" mixtures are defined), where the number "t" can vary from one prediction to another.

[0118] The mixtures {p} can be predefined in a list 140 of the determination module 14. When starting a new test procedure, the module 14 reads the list 140 and uses one, some or all of the mixture definitions from the list to form the matrix P defined above. Of course, as a variant or in combination, the operator can define at least one new mixture on the fly, by selecting samples 100 from the collection 10 and specifying their relative proportions, on a screen connected to the module 14. Thus, the mixtures {p x (k)} k can be created on the spot.

[0119] The prediction calculation I=P*A is implemented, for example, in a computer.

[0120] matrix

number

[0121] These mixing profiles are "naive" predictions because they do not take into account the interactions between microorganisms when actual mixing occurs.

[0122] Therefore, according to the invention, the prediction operation uses the interaction model, in particular the interaction matrix W, to convert the intermediate mixture profile, i.e., matrix I, into the matrix

number

[0123] In this way, predictive blend profiles can be obtained quickly for a range of blends without consuming material from the collection 10 .

[0124] Relative abundance r x The (j) are expected to be non-negative and together form the total composition (i.e., their sum equals 1 for a given mixture "x"). However, this may not be the case in matrix multiplication. Therefore, embodiments of the present invention include post-processing the result matrix R into R' to satisfy biological constraints.

[0125] For example, each negative value of R is clipped, i.e., negative abundances are set to 0. Then, the relative abundances r x (j) is normalized (e.g., by linear interpolation) to r´ x (j) so that their sum is equal to 1.

number

number

[0126] The efficiency of the method is based on modeling the actual positive and negative interactions between microorganisms in a mixed sample into a matrix, the so-called interaction matrix W. Then, a two-step process based on the matrix allows for efficient prediction of the actual mixture profile.

[0127] The interaction matrix W is learned for a given set of m profiling features. If the profiling features are reordered within a profile, the coefficients of the interaction matrix W are reordered accordingly.

[0128] The m profiling features may also evolve over time, for example, as new features are discovered, some features are removed because they are no longer as meaningful, and / or some features are split into more features to make them more accurate. Evolution of profiling features may also result from enhancements to profiling / sequencing methods and profilers / sequencers12 that provide new profiling data, as well as improvements in bioinformatics methods that combine algorithms and reference databases of features.

[0129] For example, different sets of profiling features may be considered depending on the target disease or treatment.

[0130] The number of features in a set, as well as the profiling features themselves, may evolve and change.

[0131] Thus, each time a new set of profiling features is considered, the interaction matrix W can be calculated anew, similar to the matrix A describing the initial profile collection 11. The calculated interaction matrix W can be stored in the memory of the pooling predictor 13 so that it can be reused when the corresponding set of profiling features is newly used.

[0132] The interaction matrix is ​​preferably obtained using machine learning, which is performed using a set of training data, which is a set of mixtures of samples k {p ref (k)}.

[0133] The actual reference mixture of samples is homogenized for between 10 minutes and 3 hours, preferably between 30 minutes and 1.5 hours. Homogenization is carried out at a temperature between 0° C. and 10° C., preferably between 2° C. and 6° C., more preferably at about 4° C.

[0134] The mixture is then believed to be stable for several hours, at least up to 16 hours after mixing, and preferably up to 24 hours.

[0135] Thus, this interaction matrix represents the interactions that should occur between microorganisms for a mixture stabilized at 4°C.

[0136] Other interaction matrices may be created to represent other mixing conditions.

[0137] The individual profiles of sample x {a x (j) j (j=1,…,m) are either known or obtained from a sequencer profiling the sample x. Thus, the reference linear predictive mixture profile {i ref (j) j can also be found by using the above linear equation I=P*A.

[0138] The mixing profile of the reference mixture product “ref” is the true mixing profile {r true (j) j , and is known or obtained from a sequencer profiling a reference mixing product "ref."

[0139] The baseline predictive mixture profile {r pred (j) j is the reference linear predictive mixture profile {i ref (j) j and the square interaction matrix W (during training), and for a single reference mixture product 'ref', R pred I ref *W or {r pred (j) j ={i ref(j) j *W.

[0140] Machine learning seeks to minimize the prediction error of a baseline mixture profile. In other words, the baseline true mixture profile

number

number

[0141] Machine learning training data is I ref and R true It is.

[0142] In some embodiments, the equation to be minimized is the residual vector {r true-i (j) j -{r pred-i (k)} k ={r true-i (j) j -{i ref (k)} k *W, Or, the residual matrix R true -R pred =R true -I ref *W.

[0143] Any norm can be used, such as L1, L2, Lp, etc. Preferably, the sum of squared differences (SSD) or its derivative, the mean squared error (MSE), can be used. Also, the least chi-square method can be used.

[0144] Machine learning attempts to solve the following convex optimization problem:

number

number

[0145] In an embodiment that avoids overfitting W, the formula adds a regularization term, preferably a ridge (L2) based regularization term, to the above difference. As a variant, a lasso (L1) based regularization term can also be used. The ridge approach advantageously lends itself to having more non-zero coefficients in W, and therefore more accurately models the interactions between profiling features.

[0146] Thus, machine learning attempts to solve the following convex optimization problem:

number

number

[0147] In addition, R pred Constraints can be imposed during machine learning such that I have no negative relative abundances and the sum of the relative abundances of each baseline predicted mixture profile is 1. In other words, I ref * Clip the negative relative abundances of W and calculate the predicted mixture profile for each standard, i.e., I ref *The correction matrix R' corresponds to normalizing the sum of the relative abundances of each row of W to 1. pred It is preferable to use the modified I ref *W is

number

number

[0148] The training data set (say N baseline mixtures) is split into two subsets, one for optimizing the hyperparameter λ and one for optimizing W.

[0149] Various methods are known for optimizing λ, in particular information criterion minimization methods (e.g., Akaike Information Criterion Minimization or Bayesian Information Criterion Minimization) and cross-validation residual minimization methods, which use an initial subset of the training data. In this optimization, W may be set by default to a value different from ID.

[0150] For example, if we set λ to 10 -5 and 10 3 We calculate the MSE of the above formula for the training and test datasets (which we split into subsets to optimize the hyperparameter λ) while varying between . The resulting MSE is shown in Figure 1a.

[0151] As shown, when λ is small, the MSE of the training data set is close to 0, but the MSE of the test data set is very high. In this situation, the model is overfitted.

[0152] On the other hand, when λ is high, the model is underfitted.

[0153] Therefore, λ can be chosen to minimize the MSE of the test data set.

[0154] Once λ is known, a second subset of the training data is used to learn W by minimizing the cross-validated residual, and a k-fold procedure is performed.

[0155] A subset of the training data (i.e., {r true-i (j) j and {i ref-i (j) j ) is divided into k subsets, where k is preferably selected from an integer between 3 and 20, preferably between 4 and 10, and more preferably is equal to 5.

[0156] Each of the k subsets is successively selected in a round robin fashion (cyclic order) to define a test subset, while the remaining k-1 subsets define training subsets.

[0157] Every k rounds, we train the model using the training subset. That is, to find W,

number

[0158] The learned interaction matrix W is then checked on a test subset, which is then used to model the matrix-based model R. true =I ref *Applied to W. Any norm, e.g., MSE

number

[0159] Repeating this process for each of the k test subsets gives us k scores.

[0160] The learned interaction matrix W corresponding to the best score (i.e., the lowest score) can then be selected to construct the pooled predictor 13.

[0161] Of course, other machine learning methodologies can be used once the learned interaction matrix W is obtained.

[0162] In some embodiments, the profiling features of the sample 100 (i.e., used to form matrix A) are the same as the profiling features of the final mixture result (i.e., used to form matrix R). As mentioned above, they may be taxa, genes, antibiotic resistance genes, functions, and metabolite traits, as well as metabolite and protein production (as in Experiments 1 and 2 below).

[0163] In other embodiments, the profiling features of the samples 100 (i.e., used to form matrix A) are different (partially or entirely) from the profiling features of the final mixture result (i.e., used to form matrix R). Any of the profiling features described above (taxon, gene, function, etc.) may be used.

[0164] As an example, when a profiling technique such as NGS shotgun sequencing is used, many more profiling features are obtained per sample 100 compared to 16S sequencing. Thus, the sample 100 can be profiled using NGS shotgun sequencing (hence, matrix A is formed with the NGS shotgun profiling features), while the final mixture result can retain a reduced number of profiling features obtained, for example, using 16S sequencing (hence, matrix R is formed with the 16S profiling features). In that case, matrix I is formed with the NGS-shotgun profiling features, and the interaction matrix W is not a square matrix and still models interactions between microorganisms, but in this example, as relationships between the NGS-shotgun profiling features and the 16S profiling features.

[0165] In a specific embodiment seeking to reduce a number of NGS shotgun profiling features, a principal component analysis (PCA) is performed to project the number of features into k principal components (k PCs). In one embodiment, PCA is performed in constructing the features profiling the sample, i.e., matrix A. In another embodiment, matrix I is generated using the number of profiling features, and PCA is performed on matrix I.

[0166] As described above, the pooled predictor 13

number

number

[0167] The mixture result matrix R′, i.e., the predicted mixture profile {r′ obtained by the test and decision module 14 x (j) j can then be used to control the process of actually producing the blended result product 19. For example, through signals S1 and optionally S2, can be used to control the actual removal and blending of samples from the initial sample collection 10 to obtain the blended result product.

[0168] One of the predicted blending profiles in R′ may be selected by the decision module 14 to trigger the generation of the blending result product 19 .

[0169] One or more selection criteria can be used to select one of the predictive mixture profiles.

[0170] The selection criteria may be stored in a file in memory 141. The criteria are entered into the system (list 141) by an operator and may reflect, for example, requirements for a mixed result product having hardening or processing properties.

[0171] The criteria relate to the profiling features of the profile, and are therefore synonymous with a target mixture profile, with more or less relaxed constraints on the profile features depending on the embodiment.

[0172] The criteria can include diversity criteria, for example bacterial diversity criteria.

[0173] "Diversity" or "bacterial diversity" refers to the variety or variability of a complex community (mixture or sample) of microorganisms, measured for example at the genus, species, gene, function or metabolite level. Diversity can be expressed as richness (number of observed species, genera, genes), alpha diversity parameters for describing complex communities, such as the Shannon index, Simpson index, and inverse Simpson index, and beta diversity parameters for comparing complex communities, such as the Bray-Curtis index, UniFrac index, and Jaccard index.

[0174] Thus, the diversity criterion can express a requirement in terms of a minimum number of profiling features (e.g. bacterial genera) or the presence (i.e., the corresponding relative abundance is non-zero) of one or more predefined profiling features. The minimum number of profiling features may be considered for all m profiling features or for a predefined sublist of m profiling features. This allows the selection process to focus on specific features for the desired mixture result product 19.

[0175] The predicted mixture profiles of R' are selected, for example, to satisfy the presence of a minimum number of profiling features.

[0176] The diversity criterion may represent a minimum or maximum relative abundance of one or more specific profiling features. For example, a given bacterial genus may be desired to be present in the mixture result product at a ratio (by mass) of at least 5% relative to other bacteria (identified by other profiling features). The diversity criterion may also specify a range within which the relative abundance of one or more specific profiling features should fall. Of course, various diversity criteria can be mixed, such as a minimum or maximum relative abundance of one profiling feature and a range of another feature, and / or a maximum relative abundance of a third feature, etc.

[0177] Similarly, a relative ratio between at least two profiling features (possibly a minimum and / or a maximum ratio) can also be used as a diversity measure.

[0178] For example, one can select a predicted mixture profile of R' that meets a minimum or maximum value for the relative abundance of a particular profiling feature.

[0179] The diversity criterion may also represent an increase in the diversity of the profiling feature.

[0180] The diversity criterion may specify a closeness or similarity to a particular target mixture profile. For example, a target mixture profile may be specified if a mixture result product that exactly matches the target mixture profile is desired. Typically, the target mixture profile is provided with a maximum value corresponding to the distance (measurement) to be evaluated between the profiles. If the distance (measurement given) between both profiles is less than or equal to the maximum value, the mixture profile is said to be close to the target mixture profile. The measurement may be a norm, L1, L2, ..., Lp, SSD, MSE, beta diversity index, or other known distance measurement between profiling features (e.g., Bray-Curtis distance, Jaccard distance, Unifrac distance or similarity measurement).

[0181] For example, the predicted mixture profile of R' that has the smallest distance to the target mixture profile is selected.

[0182] All or some of the criteria set out above may be combined.

[0183] One or more selection criteria used to select one of the predicted mixture profiles (or more generally the target mixture profile) are obtained by the decision module 14 and applied to R'.

[0184] The predicted mixture profiles in R' can be considered in sequence in order.

[0185] The first predicted blend profile that meets the selection criteria or criteria may be selected for generating the blend result product 19 .

[0186] In a variant, all predicted mixture profiles in R' are evaluated with respect to a selection criterion or criteria, and those with the best scores (eg, satisfying and / or closest to some criteria) are selected.

[0187] More generally, inverse prediction may be considered, where for example a selection criterion / target mixture profile is defined that corresponds to a target mixture having healing properties. The file 140 may define a candidate set of samples to be tested, from which a "best" set is searched for, taking into account the target mixture profile. This process may first test, through the file 140, a first group of heterogeneous candidate sets (i.e., having a heterogeneous mixture in terms of which samples are mixed together and in what proportions respectively) to find the "best", then another group of candidate sets in the vicinity of the "best" set (e.g., the proportion of samples in the "best" set p xThe iterations may be iterative, meaning that one may vary (k) and / or add or remove only one or at most two samples to / from the set). Of course, one or more additional iterations may be considered to incrementally refine the "best" performing sample set given the target mixture profile.

[0188] At each iteration, for example, the predicted mixture profile of R' with the smallest distance (Bray-Curtis distance, Jaccard distance, Unifrac distance, etc.) to the target mixture profile may be selected as the "best". Alternatively, multiple (e.g., matrix-based) predictions may be made and a "best" candidate set may be selected from the multiple (e.g., different matrix R') predicted mixture profiles obtained. Thus, a comparison is made between the predicted mixture profiles and the target mixture profiles and one candidate set is selected as the "best", i.e., the target set.

[0189] Once a predicted mixing profile has been selected and the corresponding target set of samples from the initial collection is known, the process of generating a mixing result product 19 begins.

[0190] The determination module 14 first determines the mixture composition {p x (k)} k That is, the ratio p of samples in the initial collection 10 to each sample k in the target set is obtained. x (k). Then, using S1, selector and mixer 15 receives these ratios {p x (k)} k In a variant, the signal S1 may be an indication to the operator, for example the ratio {p x (k)} k is displayed on a screen to the operator so that he or she manually performs the actual removal and mixing of the samples.

[0191] The selector and mixer 15 has mechanical access to the sample collection 10 (eg, via a controlled articulated arm) and may be a machine including a bioreactor that performs mixing of the sample.

[0192] In response to signal S1, selector and mixer 15 selects a non-zero proportion p from bank 10. x Pick, i.e., take or remove, a sample having (k) and a corresponding proportion p x (k) removes the amount of each sample given and the total amount or mass of the target mixed result product 19. All sample amounts removed are poured into the bioreactor where they are actually mixed.

[0193] Preferably, homogenization is performed for between 10 minutes and 3 hours, preferably between 30 minutes and 1.5 hours. Homogenization is performed at a temperature between 0° C. and 10° C., preferably between 2° C. and 8° C., more preferably at about 4° C. The resulting mixed product is believed to be stable for several hours thereafter, at least within 16 hours, preferably within 24 hours after mixing.

[0194] The true mixing profile of the resulting mixture product 19 is true (j) j can be obtained using the profiler / sequencer 12. This is done by using the selected predictive blend profile 192{r' x (j) j Since the corresponding intermediate mixture profiles {i x (j) j A new round as described above can be performed with this only new item (or items, if any) of training data, keeping λ unchanged and initializing W to its current value. Iteratively learning W while using platform 1 in this way advantageously improves W, allowing for better future mixture profile predictions.

[0195] As mentioned above, some samples 100y-100z may be hypothetical. When such a hypothetical sample is selected by the decision module 14 (i.e., its corresponding relative abundance p x If (k) is not zero, then there is a need to actually generate the samples from their virtual definitions (ie, the corresponding individual profiles).

[0196] When the decision module 14 detects such a non-zero relative abundance for a hypothetical sample 100y-100z corresponding to a bacterial consortium, it uses S2 to notify the sample generator 16 of the need to generate said artificial sample. S2 identifies the sample and determines the amount of material required (i.e., the corresponding fraction p x (k) multiplied by the total target amount or mass of the mixed result product 19.

[0197] The sample generator 16 may be a machine with mechanical access (e.g., via a controlled articulated arm) to a bank of isolated strains 160 and a memory device with access to a file 161 that defines the composition of the sample in terms of a mixture of individual strains. The sample generator 16 also includes a bioreactor in which the mixture of strains is performed.

[0198] In response to signal S2, the sample generator 16 obtains the definition of the artificial sample (bacterial consortium) in terms of strains and takes the appropriate amount of each strain required from the strain bank 16 given the required amount indicated by the signal. The taken amount of all required strains is injected into the bioreactor and mixed in practice, for example for 30 minutes at 4°C.

[0199] In an embodiment, the sample generator 16 has access to the bank 10 and / or the bank 99 of external samples. When the decision module 14 detects a non-zero relative abundance for a virtual sample corresponding to a manipulated or processed composite community (i.e., a mixture including the sample), it uses S2 to inform the sample generator 16 of the need to generate said manipulated or processed sample. S2 identifies each strain and / or each sample in the bank 10, and / or each external sample related to the mixture, and determines the amount of material required (i.e., the corresponding fraction p x (k) multiplied by the target total amount or mass of the mixture resultant product 19).

[0200] In response to signal S2, the sample generator 16 acquires or collects material and pours it into the bioreactor where it is actually mixed.

[0201] Once mixing is complete and stabilized, the sample has been generated and is stored in the initial sample collection or bank 10, and the selector and mixer 15 can actually generate the resulting mixed product 19.

[0202] Although signals S1 and S2 have been described above as control signals for driving the selector and mixer 15 and the sample generator 16, either or both of these may simply be signals displayed to the operator so that the operator actually performs the mixing manually.

[0203] 2 illustrates, by means of a flow chart, the general steps for generating such a blending result product 19, including the prediction of the blending profile. These steps are performed by the platform 1.

[0204] In step 200, the test and decision module 14 selects a set of samples from those available in the initial sample collection 11. This step simply selects one mixture rule {p x (k)} kThis may consist of providing or selecting

[0205] The definition of the mixture "x" is provided to the pool predictor 13 for prediction.

[0206] "x" is the number of mixed rules in list 140 from 1 to N mix When the index is up to , the initial value is set to 1.

[0207] In step 205, the Poole predictor 13 first performs a linear prediction of the mixture profile to obtain an intermediate mixture profile, {i x (j) j ={p x (k)} k *A. This is a matrix multiplication.

[0208] In step 210, the pooled predictor 13 then performs a correction of the intermediate mixture profile using the interaction matrix W to obtain {r x (j) j ={i x (j) j *W. This is a matrix multiplication.

[0209] Predicted mixture profile of mixture “x”, 215, {r’ x (j) j To obtain {r x (j) j Clipping and normalization of the

[0210] In step 220, {r' x (j) j is checked whether it satisfies a condition based on the selection criteria. This condition may be defined by a single selection criterion to be satisfied or by multiple selection criteria. This condition may reflect criteria that define a complex community of microorganisms or a microbial composition that has desired healing or therapeutic properties.

[0211] {r' x (j) jdoes not satisfy the condition, another mix is ​​considered via step 225 (incrementing the mix index) before looping back to step 200. Of course, if all mixes have not yet been considered (test 221), the process ends with no mixes satisfying the condition, in which case a warning message may be issued to the operator.

[0212] {r' x (j) j If satisfies the condition, it is selected and processing proceeds to step 230 .

[0213] In step 230, the corresponding mixed rule {p select (k)} k is retrieved from the list 140 by the decision module 14 and sent to the selector and mixer 15 or displayed to the operator (signal S1).

[0214] Optionally, a non-zero proportion p select If so, then the decision module 14 will trigger actual generation by the sample generator 16 through the transmission of a signal S2 or display the information to the operator.

[0215] In step 235, the corresponding non-zero p select A sample having (k) is removed from the bank 10 by the selector and mixer 15 and mixed in the bioreactor.

[0216] The result is the desired mixed result product 19.

[0217] Thanks to the above prediction operations, an accurate profile of the resulting mixture product can be quickly obtained (or at least estimated) without consuming any material (sample).

[0218] However, samples may disappear over time (either to actually produce some products or to deteriorate over time), while new samples may be collected from new donors. From this, it can be seen that after a mixture definition is determined to produce a target mixture result product, the collection 10 may evolve over time (hence A evolves). By implementing the present invention, the pool predictor 13 is newly configured with the evolved collection (A is redefined and W is learned), and another mixture definition corresponding to the evolved collection is also determined (using the prediction of the present invention) to generate a similar mixture result product.

[0219] The above sequence of steps 200-235 selects the first mixture in the list 140 that satisfies the condition.

[0220] In a variant, a predictive mixture profile can be estimated by the pool predictor 13 for all mixtures defined in the list 140 before checking the conditions in step 220 to find the “best” mixture, i.e. set of samples.

[0221] This method derives the mixture prescription {p x (k)} k (i.e., a set of samples of the collection 10). The determination of such a target mix rule (i.e., a target set of samples) may be performed in a production process, as just described. In fact, the test and determination module 14 determines the target mix rule {p x (k)} k Once obtained, it can be used to control the process of actually producing the mixed result product 19, as described above, and the decision module 14 can send a signal S1, and optionally a signal S2, to the control modules 15 and 16 in the production process or to inform an operator of the action to be taken.

[0222] 3 illustrates, by means of a flow chart, the general steps for generating such a mixing result product 19, given a target mixing profile. These steps are performed by the platform 1.

[0223] In step 300, a target mixing profile {r' x (j) j is set in file 141.

[0224] A target mixture profile may be a profile with clearly defined values ​​for the profiling features, but may also be a profile that defines more lenient values, such as, for example, a minimum relative abundance of one or more profiling features, a maximum relative abundance of one or more profiling features, a range of relative abundance of one or more profiling features, a minimum number of non-zero relative abundances in a profile or a predefined subset of profiling features, predefined profiling features with non-zero relative abundances, a defined ratio between the relative abundances of two or more profiling features, etc.

[0225] The target mixture profile is determined by the test and decision module 14 by dividing {r'(j)} j This target mixture profile can be specified in file 141 so that it can be fed to the pooled predictor 13 as (the output of the model).

[0226] In step 305, a group of candidate sets of samples are obtained from the collection 10. These may be predefined.

[0227] Each mixture ratio p x Similar to randomly selecting (k), a random selection of samples from the collection can be performed.

[0228] The number of samples to be mixed can be selected within the range of acceptable numbers, for example, 2 to 1000 samples, preferably 3 to 100 samples. In practice, 3 to 10 can be easily handled. Of course, the computer-implemented method according to the invention allows predictions for larger sample mixtures to be performed at low cost.

[0229] The blending ratio can be selected from a group of predefined ratios (the sum of the ratios must be 100%, given the number of blended samples).

[0230] The initial group of candidate sets may be formed randomly, and will typically be a highly distributed set of samples. Other groups of candidate sets may be formed given one or more known sets (e.g., a set determined as the "best" set in a previous iteration of the process of FIG. 3). Other groups of candidate sets may be formed, for example, given different mixing ratios p x Other candidate sets can include those that differ from the known set or sets only by (k) and / or by a limited number of different samples (e.g., only one or two different samples).

[0231] The candidate set is then passed to the test and decision module 14, which then passes these candidate mixtures ({p x (k)} k The list of the semaphores is provided in file 140 so that the semaphores can be provided.

[0232] Then, we calculate one or more mixture profiles {r x (j) j The above steps 205, 210, 215 are performed to predict P. The matrix-based approach allows for the prediction of mixture profiles for multiple (possibly all) candidate sets simultaneously, where P is a set of multiple or all {p x (k)} k Includes.

[0233] Step 310 checks whether all candidate sets / mixtures have been processed (test 216). If not, the next mixture rule is considered via step 225.

[0234] Next, in step 315, the predicted mixture profile is compared to the target mixture profile (selection criteria) with a view to selecting one candidate set as the target set.

[0235] For example, a distance, e.g., the Bray-Curtis distance, the Jaccard distance, the Unifrac distance, or a combination thereof, is computed for each predictive mixture profile (and thus for each candidate set).

[0236] Step 320 determines the predicted mixture profile or profiles that are closest to the target mixture profile at the considered distance. Preferably, the closest is determined.

[0237] A distance margin may be applied to ensure that the closest predictive blend profile is close enough to the target blend profile. In this case, the closest predictive blend profile must satisfy the margin, which means that the calculated distance must be smaller than this margin. If the predictive blend profile does not satisfy the test, the process is terminated and a warning message may be issued to the operator.

[0238] Otherwise, proceeding to step 230 above, the predicted mixture profile {r' select (j) j 325 has been decided.

[0239] In some embodiments (not shown), this target set / mixture composition can be used to define a new group of candidate sets in order to perform another (or more) rounds of the process as described above to refine the target set / mixture composition used in step 230.

[0240] In step 230, the mix rule "selection" is sent to the selector and mixer 15 or operator (signal S1).

[0241] Optionally, a non-zero proportion p select If so, the decision module 14 triggers its actual production by signaling the sample generator 16 through the transmission of a signal S2 or by notifying an operator.

[0242] In step 235, the corresponding p select Samples with (k) not equal to zero are removed or selected from the bank 10, for example by the selector and mixer 15, and mixed in the bioreactor.

[0243] The result is the desired mixed result product 19.

[0244] Fecal microbiota transplant (FMT) products, or more generally, microbiome ecosystem therapy products, can be constructed from multiple samples. The mixing method defined in the present invention can improve the diversity of the final FMT product more efficiently and without wasting materials compared to mono-sample methods.

[0245] The platform 1 described above with reference to Figure 1 includes several modules under the control of a central computer, for example the pool predictor 13 and the test and decision module 14 are implemented in the central computer, while the sequencer 12, the selector and mixer 15, the sample generator 16 and the bank 10 are separate machines connected to the central computer.

[0246] The above discussion mainly uses matrix-based predictive models, in particular square interaction matrices. An alternative to the latter are deep learning models such as neural networks consisting of multiple layers of parameterized differentiable nonlinear modules trained or learned by backpropagation.

[0247] 4 shows diagrammatically a computer arrangement 400 managing the production platform 1. The computer arrangement 400 can for example implement the pool predictor 13 and the test and decision module 14 and can control the sequencer 12, the selector and mixer 15 and the sample generator 16 via adapted signalling (S1 and S2).

[0248] A computing device 400 is configured to implement at least one embodiment of the present invention. The computing device 400 can be preferably a device such as a microcomputer, a workstation, or a lightweight portable device. The computing device 400 includes a communication bus 401, which preferably includes: - a central processing unit 402, such as a microprocessor, denoted as CPU; a read-only memory 403, denoted ROM, for storing a computer program for implementing the invention; a random access memory 404, denoted RAM, for storing the executable code of the method according to an embodiment of the invention, as well as registers adapted to record variables and parameters necessary for the implementation of the method according to an embodiment of the invention; a communication interface 405 connected to a network 499 for communicating with user or operator devices and / or with other devices of the platform 1, such as the sequencer 12, the selector and mixer 15, the sample generator 16; - a data storage means 406, such as a hard disk or a flash memory, for storing computer programs for implementing the methods according to one or more embodiments of the invention, as well as data required for the embodiments of the invention, including in particular the individual sample profiles (i.e. collections 11), the lists 140 and 141, is connected.

[0249] Optionally, the computing device 400 can be used to configure the platform, for example by means of a keyboard 408 or any other pointing means (for example, to define the lists 140, 141 and the collections 11 and the virtual samples 110y-z) and / or to display the results of the prediction process or the results of the inverse operation, for example by means of a target mixture definition {p select (j) j It may also include a screen 407 that serves as a graphical interface with an operator for displaying the

[0250] The computer system 400 can be optionally connected to various peripheral devices other than the sequencer 12 of the present invention, each of which is connected to an input / output card (not shown).

[0251] Preferably, a communications bus provides for communication and interoperation between the various elements included in or connected to computing device 400. The representation of a bus is not intended to be limiting, and in particular a central processing unit is operable to communicate instructions to any element of computing device 400 directly or by other elements of computing device 400.

[0252] The executable code may optionally be stored either in the read-only memory 403, in the hard disk 406 or on a removable digital medium (not shown). According to an optional variant, the executable code of the program may be received by the communication network 499 via the interface 405 so as to be stored in one of the storage means of the computer device 400, such as the hard disk 406, before being executed.

[0253] The central processing unit 402 is preferably adapted to control and direct the execution of instructions or parts of the software code of the program or programs according to the invention, these instructions being stored in one of the storage means mentioned above. On power-up, the program or programs stored in a non-volatile memory, for example the hard disk 406 or the read-only memory 403, are transferred to the random access memory 404, which contains the executable code of the program or programs and registers for storing variables and parameters necessary to implement the invention.

[0254] Experimental Results Scope of the experiment The objective of the experiments was to investigate the efficiency of the proposed machine learning procedure with interaction matrix W to predict the mixture profile of microbiota samples (experiment 1) and to determine the mixture composition given a target mixture profile (experiment 2).

[0255] Experiment 1 - Protocol We considered an initial sample collection of 10. Using 16S-based microbial community taxonomic profiling, we sequenced each microbial community sample to obtain a corresponding initial profile collection of 11. As a result, 131 taxa (genus level) were evaluated as profiling features.

[0256] The samples were then mixed. Each mixture was a combination of 3 to 6 samples in different ratios. Mixing was performed at 4°C and homogenized between 30 minutes and 1 hour and 30 minutes after mixing. Mixtures were sequenced using the same 16S-based microbial community taxonomic profiling during the steady-state period (i.e., several hours after homogenization and within 16 hours of mixing).

[0257] We employed k-fold cross-validation with k = 5 to construct a pooled predictor13, i.e., to learn λ and the interaction matrix W. The k-fold method ensured that no observation was used as both training data and test set in the same evaluation period.

[0258] The modeling approach described in the Materials and Methods chapter was tested and applied at three different taxonomic ranks (species, genus, family, order). However, the species-level dataset was excluded from the testing procedure because it was very sparse. Starting from the genus level, the taxonomic table was rich enough to allow the analysis, so the lower levels of precision (families, orders) were inferred from the genus table and used for visualization purposes only when necessary, but not as is in the modeling procedure. The main reason is that it is not possible to infer the composition of the higher levels of precision from the taxon level used in the training, and having the genus information is important from the point of view of our application.

[0259] Models were trained separately for only the native samples (Figure 5), only the fermented samples (Figure 6), and for both together (Figure 7). MSE was used to quantify the quality of the modeling when applied to the data. MSE was systematically compared between machine learning models and linear models (which provided naive predictions).

[0260] Experiment 1 - Results Figure 5a shows the initial profile collection 11 corresponding to the initial sample collection 10 containing only natural fecal microbiota samples. 27 microbiota samples were studied, whose individual profiles are shown.

[0261] Figure 5b shows the mixing profiles of 24 mixture products, each of which is made by mixing 3 to 6 microbial samples from the 27 microbial samples in Figure 5a, in terms of their respective ratios or percentages. x (k)} k has been preserved.

[0262] Figure 5c shows the mixed rule {p x (k)} k and individual sample profiles {a x (j) j denotes the error resulting from linear prediction of the given mixture profile. The linear prediction corresponds to the sole step 205, where I=A*P.

[0263] The figure also shows, on the right, the error resulting from the prediction according to the invention (steps 205 and 210), i.e. the error including the interaction matrix W. W was machine-learned using k-fold cross-validation using only the samples and mixture profiles (natural samples) of Figures 5a and 5b.

[0264] The model-based methods of the present invention provide better performance than linear methods on natural datasets.

[0265] Figure 6a shows the initial profile collection 11 corresponding to the initial sample collection 10 containing only fermented fecal microbiota samples. Thirty-six microbiota samples were studied, whose individual profiles are shown.

[0266] Figure 6b shows the mixing profiles of 48 mixture products, which are made by mixing 3 to 6 microbial samples from the 36 microbial samples in Figure 6a, in their respective ratios or percentages. x (k)} k has been preserved.

[0267] Figure 6c shows the mixed rule {p x (k)} kand individual sample profiles {a x (j) j denotes the error resulting from linear prediction of the given mixture profile. The linear prediction corresponds to the sole step 205, where I=A*P.

[0268] The figure also shows, on the right, the error resulting from the prediction according to the invention (steps 205 and 210), i.e. the error including the interaction matrix W. W was machine-learned by k-fold cross-validation using only the sample and mixture profiles of Figures 6a and 6b (fermentation samples).

[0269] The model-based method of the present invention performs significantly better than linear methods on fermentation datasets (median MSE is 5-fold lower for ML model predictions).

[0270] For Figures 7a and 7b, both the sample and mixture profiles from Figures 5a, 5b, 6a and 6b (i.e., native and fermented samples) were used as training data to machine-learn the interaction matrix W. Again, a k-fold cross-validation method was used.

[0271] FIG. 7a shows the results when the data sets of FIGS. 5a, 5b (i.e. the natural samples and their mixtures) are applied to a pooled predictor 13 constructed in this way.

[0272] The left side of the figure shows the mixed rule {p x (k)} k and individual sample profiles {a x (j) j 4 shows the error resulting from linear prediction of the mixture profile given

[0273] On the right hand side, the error resulting from the prediction according to the invention (steps 205 and 210) is shown, ie including the learned interaction matrix W.

[0274] Similar to the single dataset model, the combined dataset model results in slightly improved estimates when applied to the natural dataset.

[0275] FIG. 7b shows the results when the data set of FIGS. 6a, 6b (i.e. the fermenter samples and their mixtures) is applied to a pool predictor 13 configured in this way.

[0276] The left side of the figure shows the mixed rule {p x (k)} k and individual sample profiles {a x (j) j The figure shows the error of the linear prediction of the mixture profile given by

[0277] On the right hand side, the error resulting from the prediction according to the invention (steps 205 and 210) is shown, ie including the learned interaction matrix W.

[0278] Similar to the single dataset models, the combined dataset models significantly improved the estimations when applied to the fermentation dataset (median MSE was 4 times lower than the predicted values ​​of the ML model).

[0279] Experiment 1 - Observations and Conclusions In all cases, the model-based predictions improved the estimates of the naive (linear) methods. This was especially important in the fermentation dataset, where the naive methods perform poorly, especially for some taxonomic groups. The model-based correction methods were more efficient, probably because there was room for improvement. Adding more data to train the models would likely improve their overall performance and robustness. The learning methods that are also part of the present invention allow for such model refinement.

[0280] Experiment 2 - Protocol In this experiment, we use the interaction matrix W (i.e., W in Figures 7a and 7b) trained using the natural and fermented samples.

[0281] In this experiment, another collection of samples was considered, which consisted of 23 samples. Using 16S-based microbial community taxonomic profiling, in which 131 taxa (genus level) were considered as profiling features, the same sequencing was performed for each of the 23 microbial community samples, resulting in a corresponding collection of profiles. Figure 8 shows the collection of profiles (class level).

[0282] 160 mixtures were generated under different mixing conditions with different input microbial samples (mixing 2 to 4 samples out of 23) using the pooled predictor 13. Prediction was performed in four rounds (exp_1 to exp_4) in which eight different sample sets (chunk_1 to chunk_8) were considered with five different ratio sets (Mix1 to Mix5).

[0283] Each generated predictive mixture profile is identified by a triplet (i,j,k) where i=1,...,4 (exp), j=1,...,8 (chunk), k=1,...,5 (Mix) and has the corresponding name "exp_i-chunk_j-Mixk".

[0284] A different set of percentages was predefined (depending on the number of samples in the mixture) as follows: [Table 1] [Table 2] [Table 3]

[0285] The different candidate sets of samples were defined as follows: [Table 4]

[0286] In this experiment, the mixture “exp_1-chunk_7-Mix5”, i.e., a mixture of 33% sample-8, 33% sample-13, and 34% sample-15, was considered as the target mixture, and its predicted mixture profile was used as the target mixture profile.

[0287] The similarity between all proposed mixtures “exp_i-chunk_j-Mixk” and the target mixture “exp_1-chunk_7-Mix5” was assessed using the Bray-Curtis index at the genus level (Fig. 9).

[0288] The Bray-Curtis index is forced to 0 if at least two samples of a mixture come from the same donor, to avoid identification of mixtures with the same samples.

[0289] The mixture “exp_1-chunk_7-Mix5” and the most similar mixtures on both metrics were actually mixed, and the resulting products were sequenced using the same sequencing techniques to compare composition at the phylum and family level.

[0290] Experiment 2 - Results Table 5 below shows the 15 highest Bray-Curtis similarities (based on genus abundance) calculated from the predictive mixture profiles. Bray-Curtis similarity is equal to 1-Bray-Curtis dissimilarity. [Table 5]

[0291] Figure 9a shows the true profiles of samples containing exp_1-chunk_7-Mix5 and its closest mixture, exp_4-chunk_4-Mix4 (20% of sample-1, 30% of sample-10, 20% of sample-14, and 30% of sample-19), as well as their true mixture profiles at the phylum level.

[0292] Figure 9b shows the same comparison at the family level.

[0293] Although the initial samples have very different profiles in exp_1-chunk_7-Mix5 and exp_4-chunk_4-Mix4, the final products have very similar mixture profiles at both the phylum and family levels.

[0294] Table 6 and Figure 10 show the Bray-Curtis similarity (at genus level) results of comparing exp_1-chunk_7-Mix5 with its closest mixture, exp_4-chunk_4-Mix4, and their true mixture profiles. [Table 6]

[0295] Although the initial samples of the two mixtures have very different profiles, the final products have very similar mixture profiles at the genus level according to the Bray-Curtis distance index. Table 6 and Figure 10 show the similarity gap between the real mixtures and their predictions. However, this does not substantially affect the similarity between the real mixtures, which remains at an acceptable level.

[0296] Experiment 2 - Observations and Conclusions Experiment 2 shows that by iteratively using the prediction tool, we can predict mixtures that are very close to the target samples (Bray-Curtis similarity at the genus level ≥ 0.86) and select the corresponding samples and mixture ratios for the in vitro experimental stage.

[0297] We also show that the two mixtures generated according to the predicted recipe are very close to their real counterparts (Bray-Curtis similarity at genus level is above 0.75), demonstrating the performance of the prediction tool and its applicability in real-world use.

[0298] Experiment 3 - Protocol In this experiment, 100 samples were profiled using NGS shotgun sequencing. Metagenomic sequence data were obtained for 76 pools and 69 samples from donors or individual fermenters.

[0299] Due to the large number of NGS shotgun profiling features (compared to 16S sequencing, especially when viewed at the species level rather than the genus level, or when viewed at specific features), PCA was used to reduce the dimensionality of each sample profile to k PCs.

[0300] Figure 11 shows PCA based on genus relative abundances obtained from NGS shotgun sequencing of natural samples (natural, inoculum, mixture) and fermentation samples (fermentation, inoculum, mixture). The fermentation samples tend to cluster similarly to the natural samples.

[0301] This PCA-based method is summarized in Figure 12, which shows that instead of learning a taxon × taxon interaction matrix W (as in Experiments 1 and 2), in Experiment 3 we learn a top k principal components × taxon interaction matrix W.

[0302] The methodology for learning this interaction matrix W is the same as in the 16S analyses of Experiments 1 and 2.

[0303] Experiment 3 - Results [Table 7]

[0304] The interaction matrix W was trained using MSE, and the comparison of the predicted mixtures (using W) to the true mixtures was based on MSE or Bray-Curtis distance.

[0305] Both modeling approaches (with and without PCA) improve taxonomic profile predictions (by MSE or BC metrics) at the genus and species levels. The modifications according to the invention (based on matrix W) have a larger impact on the prediction of mixtures from fermented samples compared to natural samples.

[0306] Reduction of profiling features using PCA appears to significantly improve prediction accuracy from fermented samples, but only slightly for predictions from natural samples.

[0307] Although the present invention has been described with reference to specific embodiments, it is understood that the present invention is not limited to those embodiments, and modifications which fall within the scope of the present invention will be apparent to those skilled in the art.

[0308] Many further modifications and variations will be suggested to those skilled in the art upon reference to the exemplary embodiments described above, however, these embodiments are given by way of example only and are not intended to limit the scope of the invention, which is determined solely by the appended claims. In particular, different features from different embodiments can be interchanged where appropriate.

[0309] In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite articles "a" or "an" do not exclude a plurality. The mere fact that different features are recited in mutually different dependent claims does not indicate that a combination of these features cannot be used to advantage.

Claims

1. 1. A computer-assisted method for predicting a mixture composition resulting from a mixture of complex microbial community samples belonging to an initial sample collection, comprising: predicting intermediate mixture profiles for mixtures of selected complex microbial community samples using linear methods; modifying the intermediate mixture profile to a predictive mixture profile using an interaction model learned from a reference linear predictive mixture profile and a corresponding reference true mixture profile; The method includes:

2. 2. The method of claim 1, wherein the predicted mixing profile is used to control the actual removal and mixing of a complex microbial community sample from the initial sample collection to obtain a mixed result product.

3. 3. The method of claim 1 or 2, wherein predicting the intermediate mixture profile comprises calculating a matrix product between a first matrix defining the mixture in proportions of the complex microbial community samples of the initial sample collection and a second matrix defining individual profiles of the complex microbial community samples.

4. The method of claim 1 or 2, wherein modifying the intermediate mixture profile comprises computing a matrix product between a matrix representing the intermediate mixture profile and a square interaction matrix of the learned interaction model.

5. 1. A computer-assisted method for determining a set of complex microbial community samples in an initial sample collection given a target mixture profile representative of a target mixture result product, comprising: selecting a candidate set of complex microbial community samples from the initial sample collection; - predicting, for each selected candidate set, a mixture profile resulting from said mixture of samples of said selected candidate set using the prediction method of claim 1; comparing the predicted mixture profile to the target mixture profile to select one candidate set as a target set; The method includes:

6. 6. The method of claim 5, wherein the target set of samples is used to control the actual removal and mixing of complex microbial community samples from the initial sample collection to obtain a function of a mixture result product of the target mixture profile.

7. 7. The method of claim 5 or 6, wherein comparing the predictive mixture profile to the target mixture profile comprises calculating a distance between each predictive mixture profile and the target mixture profile and selecting the candidate set having the shortest distance as a target set.

8. The method of claim 1 or 2, wherein the profile of a complex community of microorganisms comprises the relative abundance of a profiling feature in the complex community of microorganisms.

9. The method of claim 1 or 2, wherein the profiling features forming the profile of the complex community of microorganisms comprise one or more features selected from the group consisting of taxon, gene, antibiotic resistance gene, function, metabolite trait, and metabolite and protein production, preferably comprising taxon.

10. The profile of the complex community of microorganisms defines profiling characteristics for one or more microorganisms present in said complex community of microorganisms from bacteria, archaea, viruses, phages, protozoa and fungi, preferably for bacteria and / or archaea, and / or Define profiling features that identify the relative abundance of the microorganisms considered at one or more taxonomic levels from strain, species, genera, families and orders, preferably at one or more taxonomic levels from genera, families and orders, and / or In the complex community of microorganisms, the relative abundance of bacterial and / or archaeal taxa considered at the taxonomic level of genus, family and order; The method according to claim 1 or 2.

11. 3. The method of claim 1 or 2, wherein the interaction model is obtained using machine learning to minimize a functional expression of the difference between the reference linear predictive mixture profile and a reference predictive mixture profile obtained from the interaction model, and the corresponding reference true mixture profile.

12. 1. A method for producing a complex microbial community product, comprising: selecting a complex microbial community sample from the initial sample collection; predicting a mixture profile resulting from said mixture of said selected samples using the prediction method of claim 1; comparing said predicted mixture profile to a selection criterion; Depending on the result of said comparison, actually taking and mixing said selected samples to obtain a mixed result product; The method includes:

13. The method of claim 12, wherein the selection criteria comprises one or more selected from a diversity criterion representative of an increased diversity of profiling features, a minimum or maximum relative abundance of one or more profiling features, a non-zero relative abundance of one or more specific profiling features or a minimum number of profiling features, a relative ratio between at least two profiling features, and a proximity to a target mixture profile.

14. 1. A method for generating a complex microbial community product having a target mixture profile representative of a target mixture outcome product, comprising: Selecting a target set of complex microbial community samples belonging to an initial sample collection given the target mixture profile using the determination method of claim 5; actually collecting and mixing the complex microbial community samples of the selected target set to obtain a mixing result product; The method includes:

15. 15. The method of claim 14, wherein one selected complex microbial community sample is a virtual sample, and the method further comprises actually generating an artificial complex microbial community sample from isolated strains that corresponds to the selected virtual sample.

16. 3. A computing device including at least one microprocessor configured to carry out the method according to claim 1 or 2.

17. A non-transitory computer readable medium storing a program which, when executed by a microprocessor or computer system within an apparatus, causes said apparatus to carry out the method according to claim 1 or 2.

Images