Computer-implemented method for the functional profiling of the microbiome
The method addresses inefficiencies in existing microbiome profiling by integrating de novo assembly, complete reference databases, and machine learning to provide accurate and comprehensive metabolic profiles, facilitating robust biomarker identification.
Patent Information
- Application Number
- PCT/IT2026/050123
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-13
- Filing Date
- 2026-05-12
- Publication Date
- 2026-11-19
AI Technical Summary
Existing methods for functional profiling of the microbiome, such as HUMAnN2, face limitations in accuracy and completeness due to reliance on two-step approaches and incomplete reference databases, leading to inefficiencies in identifying microbial pathways and metabolic potential.
A computer-implemented method involving de novo assembly of metagenomic sequencing data, followed by gene prediction using a statistical model, and functional annotation with a complete reference database enriched with UniProt and SwissProt/RefSeq sequences, combined with machine learning strategies to identify robust functional markers.
Enhances the accuracy and completeness of microbial function profiling by capturing rare functions, reducing false positives, and providing a comprehensive metabolic potential estimation, enabling reliable biomarker identification for diagnostic and therapeutic applications.
Smart Images

Figure IT2026050123_19112026_PF_FP_ABST
Abstract
Description
[0001] “COMPUTER-IMPLEMENTED METHOD FOR THE FUNCTIONAL PROFILING OF THE MICROBIOME”
[0002]
[0003] FIELD OF THE INVENTION
[0004] Some embodiments described here concern a computer-implemented method for the functional profiling of the microbiome.
[0005] In particular, the embodiments described here are used to perform a sequence of steps in order to analyze metagenomic data and derive from it information on microbial metabolic potential.
[0006] Some embodiments of the method find application in clinical diagnostics, for example, for the identification of microbial biomarkers correlated to specific health conditions or diseases, allowing to develop personalized therapeutic strategies based on the modulation of the human microbiota.
[0007] Other possible applications include, for example, pharmaceutical and nutraceutical research to assess the ability of microbial communities to produce bioactive metabolites (e.g., short-chain fatty acids, neurotransmitters, vitamins), useful for screening microbial strains or probiotic formulations for therapeutic or nutritional purposes.
[0008] BACKGROUND OF THE INVENTION
[0009] It is known that the functional analysis of microbial metagenomes is increasingly used to understand the properties and functions of microbial communities in different environments, thanks to the integration of assembly methods, gene prediction and functional annotation directly on DNA shotgun datasets.
[0010] There are software pipelines for efficient and accurate profiling of the presence / absence and abundance of microbial pathways in a community starting from metagenomic or metatranscriptomic sequencing data (typically millions of short DNA / RNA sequences). This process, called functional profiling, aims to describe the metabolic potential of a microbial community and its members.
[0011] Some known solutions, such as HUMAnN2, take a two-step approach: they initially map reads to databases of reference nucleotide sequences, so as to reconstruct genes, and then translate unassigned regions to align them to protein databases, for example UniRef which provides a set of sequences grouped by theUniProt Knowledge Base, allowing gene families and metabolic pathways to be quantified with high accuracy and speed.
[0012] In this context, there is the need to provide a new computer-implemented method for the functional profiling of the microbiome.
[0013] The Applicant has devised, tested and embodied the present invention to overcome the shortcomings of the state of the art and to obtain these and other purposes and advantages.
[0014] SUMMARY OF THE INVENTION
[0015] The present invention is set forth and characterized in the independent claims, while the dependent claims describe other characteristics of the present invention or variants to the main inventive idea.
[0016] DESCRIPTION OF SOME EMBODIMENTS
[0017] Some embodiments described here using fig. 1 concern a computer-implemented method for the functional profiling of the microbiome. In some embodiments, the computer- implemented method comprises:
[0018] assembly 120 and filtering 130 of metagenomic sequencing data of a biological sample;
[0019] gene prediction 140 and generation of ORF sequences using a statistical model; functional annotation 150, using an aligner software, of the predicted sequences, on the basis of a complete reference database 160;
[0020] aggregation 170 of EC number values, possibly normalizing 180 the EC number values, to estimate 190 the potential production of molecules of interest and therefore the metabolic potential.
[0021] Therefore, the method described here directly performs an assembly 120 of the metagenomic sequencing data, followed by gene prediction 140 using a statistical model. The assembly 120 reconstructs entire gene sequences, rather than directly analyzing individual reads.
[0022] The subsequent functional annotation 150 is done with respect to a complete reference database 160 that integrates bacterial EC number values available from a database (e.g. from UniProt) enriched with selected sequences (e.g. from SwissProt / RefSeq), achieving a more complete and reliable coverage of microbial functions. This increases the accuracy in recognizing the true functions present in the sample.
[0023] Furthermore, the method, via upstream aggregation 170 to aggregate biological molecules of interest, directly calculates the production potential of specific metabolites by summing normalized values of EC number values.
[0024] In some embodiments, the computer-implemented method can optionally also comprise selection 200 of variables, that is, enzymatic classes of relevant EC numbers, implemented with a hybrid strategy by means of several machine learning models to identify the most significant functional markers.
[0025] The possible hybrid selection 200 of functional markers by means of several machine learning models means that the extracted markers are statistically more robust and biologically relevant than those obtainable with standard methods.
[0026] The method described here combines an effective assembly 120 and a complete reference database 160 for the functional annotation 150 aimed at quantifying metabolic potential and possible machine learning strategies to provide a detailed functional profile of the microbiome adapted for the study of bioactive metabolites. The method is therefore particularly directed to analysis downstream of the assembly 120, and possibly directed to the extraction of specific functional markers, with a more direct metabolic vision.
[0027] Assembly and Filtering
[0028] In some embodiments, the method comprises assembly 120 and filtering 130 of sequencing data supplied at input 110.
[0029] The method therefore starts from metagenomic sequencing data, also called reads, and a software assembles the reads by overlapping them, obtaining longer and more continuous sequences called contigs. Filtering 130 discards smaller contigs, for example smaller than 300 bases, which do not usually contain useful information, keeping only those long enough to be reliable. In this way, work is only carried out with sequences of good quality and sufficient length, reducing errors in the subsequent steps.
[0030] The term “reads” refers to each of the base sequences (A, C, G, T) obtained from a single DNA fragment during a next-generation sequencing run (e.g. shotgun sequencing).
[0031] The term “contig” refers to a continuous fragment of DNA reconstructed from the assembly 120 of multiple overlapping reads; it represents a portion of an incomplete genome but without internal gaps.
[0032] For example, the sequencing data is metagenomic shotgun sequencing data. This data is generally provided at input to the assembly 120 as paired-end metagenomic sequencing files in FASTQ format (see block 110 in fig. 1). This file is a type of raw datum produced by Next-Generation Sequencing (NGS) platforms that contains reads of DNA fragments coming from a complex sample (metagenome). In paired-end sequencing, each DNA fragment is read from both ends, generating two “reads” (one for each end) that are correlated to each other. This improves the assembly 120 because it provides information about the distance and orientation between the two reads. The use of paired-end reads in metagenomics helps to reconstruct longer (contig) and more reliable genomic sequences, thanks to the overlap and to the distance constraint between the two ends. In addition, it reduces ambiguity in the repeated regions of the DNA, since the two reads “anchor” the fragment to univocal positions. In addition, it increases the overall quality of the assembly 120, which is critical when a reference genome is not available.
[0033] For example, an assembler software, or simply assembler, can be used for the assembly 120, which reconstructs longer genomic sequences (contig) by putting together overlapping reads.
[0034] In particular, a de novo assembler can be used, defined as such because the assembly 120 does not utilize any guide sequence or external model genome during reconstruction. An example of a usable assembler is MEGAHIT, which is an ultrafast graph-based assembler by De Bruijn, designed to handle metagenomic datasets of hundreds of Gbps on a single node, reducing time and the memory required. A metagenomic dataset is a set of sequences (reads) obtained by shotgun sequencing from a complex environmental sample, containing mixed DNA from multiple organisms. Generally, metagenomic data starts with millions of mixed reads from various species and an assembler overlaps them to create longer contigs, facilitating the analysis of partial genomes present in the microbial community. In summary, a De Bruijn graph is constructed in which the nodes represent k-mers (sub-sequences of length k) and the arcs represent overlaps between them; a path through the graph corresponds to a reconstructed contig. The assembler MEGAHIT to obtain the contigs can be used with potentially adjustable parameters for the k-mer (kmin, kmax, kstep).
[0035] This approach makes it possible to discover genomes of non-catalogued organisms and genetic variants that are not present in known reference databases, useful when working on complex and poorly characterized microbial communities.
[0036] The filtering 130 of the reads removes from the FASTQ files those reads that do not meet quality criteria (e.g. low average quality score) or minimum length criteria, so as to prevent incorrect or too short sequences from compromising subsequent analyses. For example, the filtering 130 is useful for increasing quality, discarding very short (typically < 300 bp) and potentially low-reliability contigs. This reduces noise and artifacts, improving the subsequent gene prediction 140.
[0037] For example, for the filtering 130 it is possible to use a toolkit, that is, a collection of software for manipulating sequence files; for example seqtk is a lightweight toolkit for operations on FASTQ such as subsampling, format conversion and filtering, which allows to select reads above a certain length or quality threshold and can be used in the embodiments described here.
[0038] Assembly 120 and filtering 130 ensure that only data of good quality and adequate length (contig) are passed to the next steps, improving the reliability of the functional analyses.
[0039] Gene prediction (ORF)
[0040] In some embodiments, the method then comprises gene prediction 140 and generation of ORF (Open Reading Frames) sequences, which recognize ORFs and coding signals, using a statistical model (e.g. HMM model).
[0041] Then, once the assembled contigs are obtained, the method searches within them for possible gene sequences. In other words, the method identifies areas of the DNA that appear to encode proteins (ORFs). These probable sequences are extracted for analysis.
[0042] In particular, once the contigs have been obtained, the method identifies the ORF coding regions by applying a statistical model. For this purpose, for example, an HMM is used. An HMM (Hidden Markov Model) is a statistical model in which a series of hidden states evolve according to Markov rules, useful for modeling biological sequences with local variability.
[0043] The statistical model (e.g. HMM) scans the contigs to identify coding regions - portions of DNA that can become proteins (ORFs) - and extracts them into a protein file. In this way, the useful parts (ORFs) containing instructions for
[0044] constructing enzymes are found in the contigs, extracting them for functional annotation analysis 150. The protein sequence to be used for the following functional annotation 150 is then extracted from each ORF.
[0045] Functional annotation
[0046] In some embodiments, the method then comprises functional annotation 150 of the predicted proteins, for example using an aligner software. The purpose is to search for similarities. This is useful, in fact, for assigning a biological function to the predicted proteins and is generally done by means of an alignment against a complete reference database 160. Each gene sequence identified is compared to databases of known genes and proteins. Each gene is assigned a function or a value or Enzyme Code (EC) on the basis of the best similarity with sequences already studied. An EC (Enzyme Commission) number is a four-level numeric code that classifies enzyme reactions (not proteins), based on the type of catalyzed reaction; each EC number corresponds to a specific reaction.
[0047] Each predicted protein (ORF) is then compared, that is, aligned, with a complete reference database 160 using an aligner software. Each protein is assigned an EC number, which describes exactly what enzymatic reaction it can catalyze. Knowing the EC number of each protein allows to understand what chemical reactions (e.g. production of fatty acids, vitamins, neurotransmitters) the microbiome is potentially capable of carrying out.
[0048] For example, the functional annotation 150 described here can be based on a complete reference database 160 that includes the bacterial enzymes (encoded by EC numbers) present on UniProt, as well as sequences of particular interest coming from Swissprot and RefSeq. This guarantees more accurate annotations of enzyme functions. The result is a table or archive of abundances of functional genes (EG) in the sample. The use of a complete reference database 160 with stringent inclusion criteria integrating EC numbers from UniProt and sequences from Swissprot and RefSeq guarantees greater accuracy in the functional annotation 150.
[0049] An example of an aligner software is DIAMOND, which is an ultra- fast protein aligner, compatible with BLASTP, optimized for large metagenomic datasets; it offers an excellent compromise between speed and sensitivity. BLASTP is a variant of BLAST that aligns protein sequences against a protein database,
[0050] calculating statistical significance and allowing functions to be inferred by homology.
[0051] This step assigns one or more EC numbers to each ORF, which identify the catalyzed enzymatic reaction.
[0052] Possibly, in the functional annotation 150 step, after alignment with, for example, DIAMOND BLASTP, it is possible to apply a filter (so-called databasesequence coverage filter) that requires an alignment to cover at least a certain percentage of a database sequence (subject coverage) before it is considered valid. This filter, for example, discards all alignments in which the portion of the database sequence (subject) covered by the alignment is less than a predetermined threshold (e.g. 50%). Using a database sequence coverage threshold can ensure that functional profiles are comparable between samples, regardless of sequencing depth or sample complexity. Applying this filter reduces false positives due to partial or spurious alignments, improving the specificity of the functional annotation 150, without significantly compromising overall sensitivity. In addition, only alignments significantly covering the reference sequences are considered for subsequent normalization on Gigabases.
[0053] Aggregation of EC values and normalization
[0054] In some embodiments, the method comprises aggregation 170 of the EC number values to estimate 190 the potential production of molecules of interest, that is, quantify the metabolic potential. If necessary, the EC number values are also normalized 180.
[0055] For example, molecules of interest can be fatty acids, neurotransmitters, tryptophan derivatives, polyphenols, vitamins, or others.
[0056] Since each gene annotated as an enzyme (EC) contributes to potential metabolic pathways, the method sums and normalizes gene levels to estimate how much capacity there is in the microbiome to produce certain bioactive molecules (e.g., fatty acids, neurotransmitters, vitamins). The more genes of a certain type, the greater the potential overall metabolic production. The normalized summation of the EC number values thus provides a quantitative estimate of the production capacity of bioactive molecules.
[0057] EC number counts are possibly normalized 180 with respect to the amount of raw data (Gigabases) in order to compare different samples.
[0058] The values of EC numbers associated with known metabolic pathways (e.g. synthesis of polyphenols, tryptophan derivatives, vitamins) are thus added, obtaining a numerical profile of the productive potential of the microbiome.
[0059] In other words, the functional annotations obtained are processed to estimate the potential production of compounds, which can even be a few thousand. This approach is not limited to traditionally analyzed compounds (e.g. fatty acids, neurotransmitters, tryptophan derivatives, etc.), but covers a broad spectrum of bioactive metabolites, offering a comprehensive view of the metabolic functionality of the microbiome.
[0060] After the functional annotation 150, the EC numbers detected are therefore quantified to estimate the ability of the microbiome to produce molecules of interest, that is, the metabolic potential. Abundance or EC number count values are aggregated by known metabolic pathways (e.g. synthesis of fatty acids, neurotransmitters, tryptophan derivatives, polyphenols, vitamins).
[0061] The aggregation 170 and possible normalization 180 of EC number values is therefore useful for estimating the overall metabolic potential. Each EC number corresponds to an enzyme reaction in a metabolic pathway; summing the EC numbers associated with the production pathways of target molecules results in a quantification of the microbiome’s ability to synthesize those metabolites.
[0062] The result is a quantitative profile of the microbiome’s functional capacity. This profile is a set of numerical values that describe how abundant or active certain functional components are within a microbial community. In this context, these components are the enzymatic activities identified by means of the EC numbers, whose abundance normalized on sequencing Gigabases reflects the microbiome’s potential to carry out specific metabolic reactions. The quantitative profile is therefore not limited to indicating the presence or absence of a function, but it quantifies its extent, allowing to rigorously compare different samples in terms of overall metabolic potential. The quantitative profile can be correlated to phenotypes of interest, so as to assess whether and how any changes in the microbiome’s functional capacity are associated with observable host characteristics, such as clinical parameters, body weight, or disease state. Establishing a robust correlation between the quantitative profile and the phenotypes makes it possible to identify microbial biomarkers predictive of healthor disease conditions. These biomarkers can then be used for early diagnosis, therapeutic monitoring or to design interventions aimed at modifying the microbiome in order to obtain a clinical benefit.
[0063] Hybrid selection of functional markers (ML)
[0064] In some embodiments, the method can comprise selection 200 of relevant variables, implemented with a hybrid strategy by means of several machine learning models to identify the most significant functional markers.
[0065] Among the various genes noted, some will be “functional markers”, that is, key functions correlated to certain biological conditions or characteristics (e.g. host health). In order to find them, the method described here uses machine learning algorithms that choose the most informative genes (or enzyme classes) in a guided manner, discarding the least relevant ones. This results in a reduced set of functions to focus on for biological interpretations.
[0066] Examples of machine learning models can be LASSO (regulated logistic regression for variable selection), Random Forest (assessment of the importance of variables with stringent threshold based on percentile), and linear SVM (extraction of predictor weights to highlight the most influential variables). The hybrid approach compares the results of the models, keeping only the markers that are reproducible and stable through cross-validation techniques.
[0067] The hybrid machine learning strategy is useful to identify the enzyme classes (ECs) most representative or predictive of a biological condition. Methods such as, for example, LASSO, Random Forest, and linear SVM are applied in parallel. These algorithms select the variables (EC) with the greatest statistical importance in the dataset. The combined use of multiple models improves the robustness and significance of the chosen markers.
[0068] In these possible embodiments, several machine learning models (e.g. LASSO, Random Forest, Linear SVM) are applied to select the most informative EC number values, that is, those functional markers that best distinguish groups (e.g. healthy vs sick). The results are cross referenced and only those markers identified by all models are kept, ensuring robustness and reproducibility. Such a hybrid approach ensures that the chosen markers are stable and truly discriminating.
[0069] In summary, the embodiments of the method described here transform raw sequencing data into a set of robust functional markers, quantifying the
[0070] microbiome’s productive capacity and selecting the most significant signals for diagnostic or therapeutic applications.
[0071] In a possible implementation of the embodiments described here, the method can be provided in the form of a software application installable on a local server or on a dedicated workstation. The software comprises a pre-processing module that receives metagenomic sequencing files at input, such as for example in FASTQ format, and initiates the sequence of steps and operations performed by the computer-implemented method to generate contigs using assemblers (e.g. MEGAHIT), apply length filters (e.g. with seqtk), predict ORF with statistical model (e.g. HMM), perform functional annotation 150 with aligner software (e.g. with DIAMOND BLASTP) using a complete reference database 160, and perform the aggregation 170 of the EC number values. In this implementation, the user interface can be a command-line interface or a graphical interface that allows the user to define input file paths, optional assembly parameters 120, and normalization criteria 180 of the EC number value counts.
[0072] In another possible implementation of the embodiments described here, the method can be provided as a web-accessible cloud service. In this configuration, the sequencing data is loaded into a cloud infrastructure that scalably deploys the method’s operations on virtual machine clusters. The architecture can, for example, provide an orchestration component that divides the data into batches, assigns each batch to computation nodes for parallel execution of assembly 120, filtering 130, and functional annotation 150, and an aggregation component 170 that combines the partial results, normalizes 180 the counts on a Gigabase basis, and applies the machine learning models for the selection 200 of functional markers.
[0073] In yet another possible implementation of the embodiments described here, the method provides hardware acceleration of the most burdensome computational steps. In this configuration, protein assembly and alignment modules can be implemented on GPUs or FPGAs via optimized libraries, reducing computational time. The HMM model for ORF prediction 140 can be run on dedicated hardware with streaming pipelines for real-time contig processing, while the machine learning step can leverage GPU cards. This embodiment guarantees a high productivity for the analysis of large cohorts of samples, keeping the sequence ofsteps and operations performed by the method described constant.
[0074] The method described here can therefore work as standalone software, cloud service or on accelerated hardware (GPU / FPGA), allowing performance optimizations based on available resources.
[0075] In all embodiments, the general structure of the complete reference database 160 is the same. Typically, the complete reference database 160 can be a data archive of protein sequences annotated with EC numbers, constructed from UniProt, SwissProt and RefSeq according to criteria of completeness and non-hypotheticity. The database is organized in relational tables or in a format optimized for highspeed searches, and includes indexes on EC numbers, species of origin, and sequence quality parameters. This structure effectively supports lookup steps during functional annotation and subsequent aggregation 170 of metabolic data.
[0076] The method in accordance with the embodiments described here allows for a more complete coverage of microbial functions, and for this purpose comprises a step of de novo assembly of the metagenomic reads in contig, allowing to reconstruct ORFs even not present in the reference databases. The de novo assembly 120 followed by ORF prediction 140 allows to analyze samples with species not represented in known pangenomes.
[0077] The method in accordance with the embodiments described here allows for better accuracy in the annotation of the enzyme functions (EC) and for this purpose uses a complete reference database 160 that includes only complete and not hypothetical sequences, for example EC numbers from UniProt, and relevant sequences from SwissProt and RefSeq, in order to reduce false positives in the assignment of EC numbers and increase accuracy.
[0078] Thanks to the de novo assembly and the complete reference database 160, the method described here allows for greater sensitivity and functional specificity, since it can capture ORFs not present in traditional databases, increasing sensitivity in detecting rare or new functions.
[0079] The method in accordance with the embodiments described here allows for an integrated and normalized quantification of the metabolic potential, since it favorably normalizes the EC number counts on sequencing Gigabase basis, and aggregates on specific metabolic pathways in order to estimate the productive potential of a high number, in the order of a few thousand, of bioactive compounds.
[0080] Moreover, it is possible to obtain a broad-spectrum profiling of bioactive metabolites since the aggregation 170 of EC number values on metabolic pathways allows to estimate not only canonical metabolic pathways (fatty acids, vitamins), but also the synthesis capacity of neurotransmitters and polyphenols, or other, offering a much broader picture than prior art approaches.
[0081] The method in accordance with the embodiments described here possibly allows for a robust selection 200 of functional markers thanks to the fact that it implements a hybrid approach based on several machine learning models, cross referencing the results to identify only the reproducible and discriminating EC number markers, improving the reliability of the extracted biomarkers. This therefore results in biomarker robustness, in particular for diagnostic applications, thanks to the combination of several machine learning models, which reduces the risk of selecting noisy variables, increasing the reliability of functional markers for diagnosis or stratification of patients.
[0082] The method in accordance with the embodiments described here allows for comparability between samples thanks to the use of database-sequence coverage filters, both in the protein alignment step and also in the normalization on Gigabases, which ensures that the functional profiles are comparable regardless of the sequencing depth or the complexity of the sample.
[0083] The method also achieves inter-study comparability, since normalization on Gigabases and the use of a stable and reproducible database facilitate comparison between different studies, reducing variability due to changing database versions or alignment parameters.
[0084] The method in accordance with the embodiments described here overcomes the limitations of the prior art approaches in terms of completeness of functional coverage, accuracy of enzymatic quantification and robustness in marker selection, offering an end-to-end method that integrates de novo assembly, complete reference database, and possibly hybrid selection 200 of functional markers using machine learning, to obtain microbiome metabolic profiles that are more accurate and comparable between different samples.
[0085] EXPERIMENTAL DATA
[0086] Applying the method described here to 80 samples (40 healthy, 40 patients with pancreatic cancer), a Principal coordinate analysis (PCoA) diagram based on theBray-Curtis distance was generated, which shows a clear separation between the groups, supported by very high statistical significance (adonis P < 2-1 O '6). The PCoA diagram is shown in fig. 2, where the two groups of samples, healthy and sick, are visible.
[0087] It was found that the functional profiles of the microbiome differ to such an extent between healthy and sick, that they can be distinguished with great reliability, suggesting possible diagnostic markers.
[0088] It is clear that modifications and / or additions may be made to the present invention, without thereby departing from the field and scope of protection, as defined by the claims.
Claims
CLAIMS1. Computer-implemented method for the functional profiling of the microbiome, said method comprising:assembly (120) and filtering (130) of metagenomic sequencing data of a biological sample;gene prediction (140) and generation of ORF sequences, using a statistical model; functional annotation (150), using an aligner software, of the predicted sequences, on the basis of a complete reference database (160);aggregation (170) of the EC number values, possibly normalizing said EC number values (180), to estimate (190) the potential production of molecules of interest and therefore the metabolic potential.
2. Computer- implemented method as in claim 1, wherein said assembly (120) receives at input paired-end metagenomic sequencing files in FASTQ format.
3. Computer-implemented method as in claim 1 or 2, wherein said metagenomic sequencing data are reads obtained from a next-generation sequencing run, in particular shotgun sequencing, and said assembly (120) assembles the reads by overlapping them, obtaining longer and more continuous sequences, called contigs.
4. Computer- implemented method as in any claim from 1 to 3, wherein said statistical model comprises a HMM model.
5. Computer- implemented method as in claims 3 and 4, wherein, once the contigs are obtained, said HMM model scans the contigs to identify the ORF coding regions and extracts them into a protein file for the functional annotation analysis (150).
6. Computer-implemented method as in any claim from 1 to 5, wherein said complete reference database (160) includes bacterial enzymes encoded by EC numbers present in the UniProt database, as well as sequences of particular interest from the Swissprot and RefSeq databases.
7. Computer- implemented method as in any claim from 1 to 6, wherein said EC number values are normalized with respect to the amount of raw data Gigabases.
8. Computer-implemented method as in any claim from 1 to 7, wherein to estimate the potential production of molecules of interest and therefore the metabolic potential, EC number values associated with known metabolic pathways,optionally normalized, are aggregated or added, obtaining a quantitative, or numerical, profile of the production potential of bioactive compounds by the microbiome.
9. Computer-implemented method as in claim 8, wherein said quantitative or numerical profile is a set of numerical values describing how abundant or active certain functional components, in particular enzymatic activities identified by means of EC number values, are within a microbial community, wherein an abundance normalized on sequencing Gigabases of said enzymatic activities reflects the microbiome’s potential to carry out specific metabolic reactions.
10. Computer-implemented method as in any claim from 1 to 9, further comprising selection (200) of variables, that is, enzymatic classes of relevant EC numbers, implemented with a hybrid strategy by means of several machine learning models to identify the most significant functional markers.
