In this installment of the Artificial Intelligence Seminar Series, we will hear short talks from two students from Prof. Mark Craven's lab. The group develops and applies machine learning to biomedical problems; for example, extracting structured information from scientific literature, inferring and modelling gene interaction networks, and modelling, classifying, and aligning temporal gene expression data. After the talks, we'll have time for informal discussion with Prof. Craven and the speakers, Deborah Muganda-Rippchen and Deborah Chasman. Refreshments are provided!

Deborah Muganda-Rippchen: Computing Multi-Level Clustered Alignments of Gene-Expression Time Series

Identifying similarities and differences in expression patterns across multiple time series can provide a better understanding of the relationships among various normal biological and experimentally induced conditions such as chemical treatments or the effects induced by a gene knockout/suppression. We consider the task of identifying sets of genes that have a high degree of similarity both in their (i) expression profiles within each condition, and (ii) changes in expression responses across conditions. Previously, we developed an approach for aligning time series that computes clustered alignments. In this approach, an alignment represents the correspondences between two gene expression time series. Portions of one of the time series may be compressed or stretched to maximize the similarities between the two series. A clustered alignment groups genes such that the genes within a cluster share a common alignment, but each cluster is aligned independently of the others. Unlike standard gene-expression clustering, which groups genes according to the similarity of their expression profiles, the clustered-alignment approach clusters together genes that have similar changes in expression responses across treatments. We have now extended the clustered alignment approach to produce multi-level clusterings that identify subsets of genes that have a high degree of similarity both in their (i) expression profiles within each treatment, and (ii) changes in expression responses across treatments. We examine the validity of this multi-level clustering method by performing a GO-term enrichment analysis of the clusters. Additionally, we use permutation testing to determine if our clusters have alignment scores that are unlikely to occur by chance.

Deborah Chasman: Inferring host subnetworks involved in viral replication

Viruses require host machinery to complete nearly every step of their life cycle. An understanding of the interactions between viruses and their hosts is needed for the development of antiviral therapeutics that target specific processes. High-throughput, genome-wide screens can identify host genes involved in viral replication; however, these screens generally produce a large amount of unstructured data, which can be difficult to interpret. We propose a method to hypothesize which specific host pathways and interactions are relevant to viral replication. We view the gene interaction network as a graph, in which proteins or genes are represented as nodes, and various pairwise interactions identified by experimental techniques are represented as edges. We use a mixed integer linear programming approach to identify subnetworks that are consistent with the biological data. The predicted subnetworks distill available knowledge into a comprehensible format, make predictions about which host factors are most proximal to direct interactions with viral components, and may be used to guide further biological research.