Identifying Genes in Published Pathway Figure Images

Abstract

BACKGROUNDPathway figures are commonly found in the biomedical literature providing intuitive models of complex processes in a visually concise format. The contents of a pathway figure often reflect the key findings and relevant context of an article. Unfortunately, the vast majority of pathway figures are drawn as one-off static images despite freely available pathway tools and resources, thus rendering their contents inaccessible to search, data mining and downstream analysis.\n\nAPPROACHLeveraging advances in optical character recognition and domain expertise in pathway modeling, we devised an approach to identify genes in published pathway figures. The approach was optimized against a set of figure images obtained from PubMed Central and tested against a set of 400 curated pathways with known content from WikiPathways (F-measure 95.2%).\n\nRESULTSApplied to 3982 published pathway figures spanning a four year period, our approach identified 29,189 gene symbols representing 4159 unique gene identifiers. The gene content unlocked from just this small sample of published figures includes novel and diverse pathway associations unmatched by any pathway database. Our approach over doubled the number of genes associated with the articles containing these figures as compared to combined annotations available from PubMed and PubTator. Encouraged by these initial results, we plan to scale the approach to make the molecular contents of the continuing stream of published pathway figures more accessible.