Machine learning methods for detecting structure in metabolic flow networks
Metabolic flow networks are large scale, mechanistic biological models with good predictive power. However, even when they provide good predictions, interpreting the meaning of their structure can be very difficult, especially for large networks which model entire organisms. This is an underaddressed problem in general, and the analytic techniques that exist currently are difficult to combine with experimental data. The central hypothesis of this thesis is that statistical analysis of large datasets of simulated metabolic fluxes is an effective way to gain insight into the structure of metabolic networks. These datasets can be either simulated or experimental, allowing insight on real world data while retaining the large sample sizes only easily possible via simulation. This work demonstrates that this approach can yield results in detecting structure in both a population of solutions and in the network itself.
This work begins with a taxonomy of sampling methods over metabolic networks, before introducing three case studies, of different sampling strategies. Two of these case studies represent, to my knowledge, the largest datasets of their kind, at around half a million points each. This required the creation of custom software to achieve this in a reasonable time frame, and is necessary due to the high dimensionality of the sample space.
Next, a number of techniques are described which operate on smaller datasets. These techniques, focused on pairwise comparison, show what can be achieved with these smaller datasets, and how in these cases, visualisation techniques are applicable which do not have simple analogues with larger datasets.
In the next chapter, Similarity Network Fusion is used for the first time to cluster organisms across several levels of biological organisation, resulting in the detection of discrete, quantised biological states in the underlying datasets. This quantisation effect was maintained across both real biological data and Monte-Carlo simulated data, with related underlying biological correlates, implying that this behaviour stems from the network structure itself, rather than from the genetic or regulatory mechanisms that would normally be assumed.
Finally, Hierarchical Block Matrices are used as a model of multi-level network structure, by clustering reactions using a variety of distance metrics: first standard network distance measures, then by Local Network Learning, a novel approach of measuring connection strength via the gain in predictive power of each node on its neighbourhood. The clusters uncovered using this approach are validated against pre-existing subsystem labels and found to outperform alternative techniques.
Overall this thesis represents a significant new approach to metabolic network structure detection, as both a theoretical framework and as technological tools, which can readily be expanded to cover other classes of multilayer network, an under explored datatype across a wide variety of contexts. In addition to the new techniques for metabolic network structure detection introduced, this research has proved fruitful both in its use in applied biological research and in terms of the software developed, which is experiencing substantial usage.