Place: CS Conference Room in MUDD
Abstract:
Much of knowledge modeling in the molecular biology domain involves interactions between proteins, genes, various forms of RNA, small molecules, etc. Interactions between these substances are typically extracted and codified manually, increasing the cost and time for modeling and substantially limiting the coverage of the resulting knowledge base. In this presentation, I will describe an automatic system that learns interaction verbs from text. We investigate text features relating verbs with genes and proteins, and apply statistical tests and a linear regression model to determine whether a given verb belongs to the class of interaction verbs. Our system, AVAD, achieves over 87% precision and 82% recall when tested on an 11 million word corpus of journal articles. In addition, we compare the automatically obtained results with a manually constructed database of interaction verbs and show that the automatic approach can significantly enrich the manual list by detecting rarer interaction verbs that were omitted from the database.
In the second part of presentation, I will introduce my ongoing research work, pattern learning. The learned interaction verbs and their inflections are used as keywords to detect binary protein-protein interaction patterns, which will be expressed in the Lexico-Syntactic style: First, a full parser is used to parse sentences, and then the lowest-level node which covers two proteins and at least one of the keywords is found. After that, an algorithm proposes candidate patterns under the node automatically. Finally, several possible ranking methods will be used to sort the patterns. Some preliminary results will be shown.
ps: The paper about verb detection, "Learning Anchor Verbs for Biological Interaction Patterns from Published Text Articles", can be found on my webpage: www.cs.columbia.edu/~wubin.