Novel Approaches to Network Inference

Abstract: this talk will have two parts. In Part 1 (30 mins) we consider unsupervised inference of network topology. In Part 2 (20 mins) we give brief overviews of other past or current research themes we are following within machine learning and bioinformatics. The understanding of pathway structures is crucial to our comprehension of the functional organisation of genes and proteins in the cell. In Part 1 we consider a new method (called GLAMA) based on a sparse graphical lasso approach (work with Yiming Ying, University of Exeter). This approach is based on the observation that if the ij^th component of the precision matrix (estimating the inverse covariance matrix) is zero, then variables i and j are conditionally independent. The proposed method uses a projected gradient descent procedure to minimise a regularised Bregman divergence between the estimated precision matrix and inverse covariance matrix derived from the data. We apply the method to a dataset derived from induced perturbations of the ERK pathway. The ERK pathway is well studied due to its importance in cancer research and other areas and it acts as an appropriate ground truth. By adjusting the sparsity parameter in GLAMA we can obtain a sparse network with 100% accuracy for the predicted functional links or a denser network with many more predicted links but with an accuracy less than 50%. In Part 2 of the talk we outline four other themes of past or current interest related to multiple kernel learning, data integration, metric learning and applications in bioinformatics.