Identification and prediction of metabolic pathways from correlation-based metabolite networks

1. A method for determining a likelihood of a metabolic pathway existing in an organism, implemented by a computer processor having an associated memory, the memory including instructions that when executed by the computer processor implement the method comprising:

calculating a network feature vector for each metabolic pathway of first and second sets of metabolic pathways, wherein the first set is composed of metabolic pathways known to exist in the organism, wherein the second set is composed of metabolic pathways known to not exist in the organism, wherein elements of the network feature vectors are network properties of subgraphs of a metabolite concentration correlation network (MCCN), wherein the MCCN links together nodes representing metabolites of the organism, wherein the links between nodes represent correlations between concentration profiles of the metabolites; wherein each subgraph is a mapping of metabolites of the existing or not existing metabolic pathways to the nodes representing those metabolites in the MCCN;

receiving the calculated network feature vectors at a supervised machine learning (SML) model, to train the SML model for classifying metabolic pathways as existing or not existing in the organism according to network feature vectors determined from MCCN subgraph mappings;

mapping to the MCCN a proposed metabolic pathway to determine a network feature vector of the proposed metabolic pathway, wherein the proposed metabolic pathway is not previously known to exist in the organism; and

processing the network feature vector determined for the proposed metabolic pathway, by the trained SML model, to determine a likelihood of the proposed metabolic pathway existing in the organism.

2. The method of claim 1 , further comprising setting a model threshold of the SML model, wherein the likelihood is positive when the SML model indicates a prediction value above the threshold, and the likelihood is negative when the SML model indicates a prediction value below the threshold.

3. The method of claim 2 , further comprising performing in vivo testing of the proposed metabolic pathway when the likelihood is positive.

4. The method of claim 1 , wherein training the SML model comprises training on at least two machine learning models and selecting a model that provides results having the greatest area under a receiver operating characteristic curve (AUC).

5. The method of claim 4 , wherein the machine learning models comprise at least one algorithm from a set including: random forest (RF), random forest with reduced feature set (RF red), AdaBoost (AB), XGBoost, random tree (RT), support vector machine (SVM), and naïve Bayes (nB) algorithms.

6. The method of claim 1 , wherein the network feature vectors include highest ranking features according to an InfoGain algorithm.

7. The method of claim 1 , wherein training the SML model comprises applying the network feature vectors of the first and second sets of metabolic pathways to train a first SML model, determining a subset of features of the network feature vectors that most contribute to the classifying capability of the first SML model, and training a second SML model by applying network feature vectors of the first and second sets that include only the subset of features determined to contribute the most to the classifying capability.

8. The method of claim 7 , wherein the subset of features are the features listed in Table 2 of the specification.

9. A system for identifying previously unknown metabolic pathways in an organism, comprising a computer processor and an associated memory, the memory comprising instructions that when executed by the computer processor implement steps comprising:

calculating a network feature vector for each metabolic pathway of first and second sets of metabolic pathways, wherein the first set is composed of metabolic pathways known to exist in the organism, wherein the second set is composed of metabolic pathways known to not exist in the organism, wherein elements of the network feature vectors are network properties of subgraphs of a metabolite concentration correlation network (MCCN), wherein the MCCN links together nodes representing metabolites of the organism, wherein the links between nodes represent correlations between concentration profiles of the metabolites; wherein each subgraph is a mapping of metabolites of the existing or not existing metabolic pathways to the nodes representing those metabolites in the MCCN;

receiving the calculated network feature vectors at a supervised machine learning (SML) model, to train the SML model for classifying metabolic pathways as existing or not existing in the organism according to network feature vectors determined from MCCN subgraph mappings;

mapping to the MCCN a proposed metabolic pathway to determine a network feature vector of the proposed metabolic pathway, wherein the proposed metabolic pathway is not previously known to exist in the organism; and

processing the network feature vector, determined for the proposed metabolic pathway, by the trained SML model, to determine a likelihood of the proposed metabolic pathway existing in the organism.