A figure-based system for extracting, archiving, and retrieving protein-protein interactions (PPIS) from biomedical literature.

摘要

Proteins are complex biological polymers that are commonly considered the workhorses of cells. They mediate virtually all the cellular functions. Correctly identifying and characterizing Protein-Protein interactions (PPIs) is an important task for thoroughly understanding the molecular mechanisms within cells. Despite great efforts that have been made by life science researchers to identify PPIs through experiments and then document them through publications, there still lacks an effective means for retrieving PPI data from literature: manual curation of documents provides accurate results, but is a slow and tedious task that requires a large amount of effort. In this thesis we present a comprehensive system to automatically extract PPI-related information from biomedical articles by mining both textual and graphical information. Our framework aims to assist life scientist to accurately and efficiently curate relevant PPI information from literature.;We first develop a solution for robustly harvesting figure-caption pairs from biomedical literature. Our approach relies on the idea that the PDF specification of the document layout can be used to identify encoded figures and figure boundaries within the PDF and enforce constraints among figure-regions. This allows us to harvest fragmented figures from the PDF, correctly identify subfigures that belong to the same figure, and identify the captions associated with each figure. Our method simultaneously recovers figures and captions and applies additional filtering process to remove irrelevant figures such as logos, to eliminate text passages that were incorrectly identified as captions, and to re-group subfigures to generate a putative figure.;We then present a robust solution for automatically segmenting each figure into unimodal panels. Our approach analyzes figure captions to estimate the number of panels and then combines it with geometric constrains obtained from the position of the panel labels for conducting robust panel extraction. We further develop a hybrid image-text based classification scheme to automatically identify experimental evidence (methods) of PPIs in each panel. We store all processed results (the raw documents, their figures and captions, PPI methods, etc.) in a relational database and construct a new content-based image retrieval (CBIR) system called ePPI (Experimental PPIs). Life scientists can use cascaded queries in ePPI to easily and effectively retrieve PPI-related figures within each article as well as capture essential facts (e.g., experimental methods) regarding specific pairs of PPIs.