An Active Learning Approach to Network Inference

The understanding of pathway structures is crucial to our understanding of the functional organisation of genes and proteins. Abnormally functioning pathways underlie many human diseases. Given the extent of noise in biological datasets, and limited amounts of data available, unsupervised determination of network topology is a substantially under-determined inference problem. Consequently, in this project, we focus on network completion, which is a supervised learning task involving multiple types of data. Starting from a set of known links and non-links we construct a classifier which predicts and ranks further possible functionmal links for experimental validation. There will be a parallel experimental programme following an active learning approach to network inference, that is, predictions of functional links will be investigated which will, in turn, be used to improve the predictor. To provide a focus for the accompanying laboratory work, by the medical researchers associated with the project, our principal aim will be the discovery and validation of pathway structures associated with hypertension. Hypertension is the most common cause of preventable disease in the developed world.

The project has a wide range of potential beneficiaries. The proposed innovations in machine learning, in multiple kernel and active learning, could be applied across a wide range of disciplines, from recognizing hand-written digits, to face identification, text categorisation, bioinformatics and database marketing, for example. Thus the proposed contributions have potential applications well beyond the indicated biomedical application. Other topics we cover are also of interest within machine learning, and associated application domains, such as network inference, data cleaning, outlier detection and learning with label noise, for example. The project would also be of interest to bioinformatics researchers and medical researchers interested in network inference. The inference of transcriptional regulatory networks is very important for advancing our understanding of the complex regulatory mechanisms within cellular systems: many diseases are associated with abnormally functioning pathways. Finally, the project would be of interest to medical researchers with an interest in the hypertensive state. Though the emphasis of the project is algorithmic innovation and a novel approach to network inference, we have necessarily had to focus on a specific biomedical context. Professor Murphy's prime interest is the hypertensive state and so the associated biological context is the understanding of the cellular regulatory networks associated with this condition. This, in itself, is a very important goal given that hypertension is a significant contributor to premature mortality and morbidity and, for the vast majority of cases (about 95%), the cause is unknown (essential hyperytension). Hypertension is a major econmic burden for society since it is a major cause of stroke, heart failure, renal
failure, blindness and myocardial infarction.