Towards the Integration of Metabolic Network Modelling and Machine Learning for the Routine Analysis of High-Throughput Patient Data

The decreasing cost of high-throughput technologies allows to consider their use in healthcare and medicine. To prepare for this upcoming revolution, the community is assembling large disease-dedicated datasets such as TCGA or METABRIC. These datasets will serve as references to compare new patient samples to in order to assign them to a predefined category (i.e. ‘patients associated with poor prognosis’). Some problems affecting the downstream analysis remain to be solved, the bottleneck is no longer data generation itself but the integration of the existing datasets with the present knowledge. Constraint-based modelling, that only requires the setting of a few parameters, became popular for the integration of high-throughput data in a metabolic context. Notably, context-specific building algorithms that extract a subnetwork from a reference network are largely used to study metabolic changes in various diseases. Reference networks are composed of canonical pathways while extracted subnetworks include only active pathways in the context of interest based on high-throughput data. Even though these algorithms can be part of automated pipelines, to be applied by clinicians, the model-building pipelines must be coupled to a standardized semi-automated analysis workflow based on machine learning approaches to avoid bias and reduce the cost of diagnostics.