Search

We investigate the application of distributional semantics models for facilitating unsupervised extraction of biomedical terms from unannotated corpora.\ud Term extraction is used as the first step of an ontology learning process that aims to (semi-)automatic annotation of biomedical concepts and relations from more than 300K PubMed titles and abstracts. We experimented with both traditional distributional semantics methods such as Latent Semantic Analysis (LSA) and Latent Dirichlet Allocation (LDA) as well as the neural language models CBOW and Skip-gram from Deep Learning. The evaluation conducted concentrates on sepsis, a major life-threatening condition, and shows that Deep Learning models outperform LSA and LDA with much higher precision.

I present a method for inferring a protein interdependency network, based on correlated evolution between proteins on a large phylogeny of 72 diverse eukaryotic species. My original contribution is in the span of the phylogenetic tree used to generate the network: similar studies have concentrated on more localised regions in the tree of life and have been undertaken with more limited intent. I show that the whole-eukaryotic correlated evolution network is a real network and has interesting features of its own. The method can be broken down into three major, sequential parts: binary trait derivation, phylogenetic inference and likelihood analysis. I describe the implementation of a reciprocal BLAST protocol for the inference of a binary trait matrix corresponding to the\ud presence or absence of orthologues in each species in the analysis, based on a reference human proteome. Rows in the matrix correspond to the reference proteins, columns to species: entries are 1, denoting presence, or 0, denoting absence. The matrix that I derive is mapped onto a phylogeny of the same set of species, to facilitate the detection of correlated evolution between proteins, or\ud orthologous sets thereof, based on the pattern of gains and losses. The phylogeny is inferred from a set of genes, which are selected according to the criterion that they be present in all 72 species. 15 genes meeting this criterion are identified from the trait matrix; 14 of them are aligned and used for phylogenetic inference. The inference itself is performed using the program BayesPhylogenies, which implements a phylogenetic mixture model using a Markov Chain Monte Carlo (MCMC) method. A consensus phylogeny (tree) is calculated after the chain has been run for many millions of iterations; trees based on the genes individually were also inferred, for purposes of comparison. I use the program BayesTraits to perform a likelihood analysis on pairs of proteins from the trait matrix. This method detects correlated evolution by means of a likelihood ratio statistic, relating the likelihood of the two proteins having evolved independently, to the likelihood of their having\ud evolved in an interdependent, or correlated, fashion. If the likelihood ratio statistic exceeds a certain threshold, this is interpreted as the signature of correlated evolution. Using presumptively interacting protein pairs from the Human Protein Reference Database, and a control (or null) set of pairs where no interaction is expected, I present evidence for the efficacy of the method in detecting correlated evolution. I proceed to infer a network based on correlated evolution, wherein\ud each link represents an instance of pairwise correlation, and demonstrate that a power law gives a good fit to the distribution of nodal degree within the network, which is also the case for a network of presumptive protein interactions with no filter for correlated evolution. Finally, I infer a new equation to characterise the evolutionary rules which fashioned the network. I propose a method for testing the equation, and discuss future directions.

Background: The motivation for the BioHub project is to create an Integrated Knowledge Management System (IKMS) that will enable chemists to source ingredients from bio-renewables, rather than from non-sustainable sources such as fossil oil and its derivatives. Method: The BioHubKB is the data repository of the IKMS; it employs Semantic Web technologies, especially OWL, to host data about chemical transformations, bio-renewable feedstocks, co-product streams and their chemical components. Access to this knowledge base is provided to other modules within the IKMS through a set of RESTful web services, driven by SPARQL queries to a Sesame back-end. The BioHubKB re-uses several bio-ontologies and bespoke extensions, primarily for chemical feedstocks and products, to form its knowledge organisation schema. Results: Parts of plants form feedstocks, while various processes generate co-product streams that contain certain chemicals. Both chemicals and transformations are associated with certain qualities, which the BioHubKB also attempts to capture. Of immediate commercial and industrial importance is to estimate the cost of particular sets of chemical transformations (leading to candidate surfactants) performed in sequence, and these costs too are captured. Data are sourced from companies' internal knowledge and document stores, and from the publicly available literature. Both text analytics and manual curation play their part in populating the ontology. We describe the prototype IKMS, the BioHubKB and the services that it supports for the IKMS. Availability: The BioHubKB can be found via http://biohub.cs.manchester.ac.uk/ontology/biohub-kb.owl.