Statistical normalisation of network propagation methods for computational biology
- Autores: Sergio Picart Armada
- Directores de la Tesis: Alexandre Perera Lluna (dir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2020
- Idioma: español
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
The advent of high-throughput technologies and their decreasing cost have fostered the creation of a rich ecosystem of public database resources. In an era of affordable data acquisition, the core challenge has shifted to improve data interpretation, in order to understand normal and disease states. To that end, leveraging the current contextual knowledge in the form of annotations and biological networks is a powerful data amplifier to elucidate novel hypotheses.
Label propagation and diffusion are the linchpin of the state of the art in network algorithms. In its simplest form, label propagation predicts the labels of a given node (for instance a gene, protein or metabolite) using those of its interactors. More elaborated approaches propagate beyond direct interactors, with robust performance in many computational biology domains.
It has been pointed out that the topological structure of biological networks can bias propagation algorithms. Poorly known entities are overlooked and harder to link to experimental findings, which in turn keeps them barely annotated. Some efforts try to break this circularity by statistically normalising the topological bias, but the properties of the bias and the real benefit of its removal are yet to be carefully examined.
This thesis covers two blocks. First, a characterisation of the bias in diffusion-based algorithms, with the implementation of statistical normalisations. Second, the application of such normalisation in classical computational biology problems: pathway analysis for metabolomics data and target gene prediction for drug development. In the first block, the presence of the bias is confirmed and linked to the network topology, albeit dependent on which nodes have labels. Equivalences are proven between diffusion processes with variations on their definitions, thus easing its choice. Closed forms on the first and second statistical moments of the null distributions of the diffusion scores are provided and linked to the spectral features of the network. The normalisation can be detrimental if the bias favours nodes with positive labels. An ad-hoc study of the data and the expected properties of the findings is recommended for an optimal choice. To that end, this thesis contributes the diffuStats software package, easing the computation and benchmark of several normalised and unnormalised diffusion scores.
The second block starts with pathway analysis for metabolomics data. This choice is driven by the relative lack of computational solutions for metabolomics, whose output still requires an effortful interpretation. Here, a knowledge graph is conceived to connect the metabolites to the biological pathways through intermediate entities, like reactions and enzymes. Given the metabolites of interest, a propagation process is run to prioritise a relevant sub-network, suitable for manual inspection. The statistical normalisation is required due to the network design and properties. The usefulness of this approach is proven not only regarding pathway findings, but also examining the metabolites and reactions within the suggested sub-networks. The knowledge network construction and the propagation algorithm are distributed in the FELLA software package.
The second practical application is the prediction of plausible gene targets in disease. Besides benchmarking the effect of the statistical normalisation, particular care is put into obtaining meaningful performance estimates for practical drug development. Target data is usually known at the protein complex level, which leads to performance over-estimation if ignored. Here, this effect is corrected in a varied comparison of prioritisation algorithms, networks, performance metrics and diseases. The results support that the statistical normalisation has a small but negative impact. After correcting for the protein complex structure, network-based algorithms are still deemed useful for drug discovery.
