Interaction Variability of Human Protein Isoforms Identified through Biomedical Literature Mining

ABSTRACT : Over the last decade, advances achieved in genomic technologies have led to uncover vast amount of protein-protein interaction data. Nevertheless, the existing protein-protein interaction databases cover the interactions related only to a part of the proteome and protein isoform interaction databases are sparsely populated. Such isoforms are generated through transcript diversity mechanisms (e.g. alternative splicing) and could exhibit functional differences. Protein-protein interaction data on isoforms is necessary for analysing their functional similarities and understanding the effects of transcript diversity on protein-protein interaction networks. Biomedical literature is an invaluable complementary resource to experimental data. Automated tools are required to gather, view and analyse the isoform interactions from the biomedical literature. This study presents a comprehensive automated text mining based analysis, which extracts protein interactions from the biomedical literature for human protein isoforms linked to the transcripts clustered in HumanSDB3 (an alternative splicing database of the human transcriptome). Extracted protein-protein interaction data is delivered to public through a new database called TBIID which stands for Transcript Based Isoform Interactions Database. TBIID contains a total number of 31,819 interactions between 7,161 unique proteins. The interaction data is automatically gathered from a subset of 205,207 interaction abstracts, which are selected from about 4 million Medline abstracts belonging to the isoforms in HumanSDB3. The automatic extraction methods achieve state-of-the-art performance (53.22% precision, 68.94% recall, 60.07% F1-score). TBIID is utilised to quantify the variability in the isoform interactions based on their shared and unique interactions. Results reveal that almost all clusters analysed (99%) contain isoforms interacting with unique protein partners, with an average unique to shared interaction rate of ~5. Similar results are obtained by analysing the data from public protein-protein interaction databases. These findings are significant in that they demonstrate that isoforms tend to interact with unique partners, indicating that they could be involved in different interaction networks potentially for performing different functions. Hence, it can be concluded that transcript diversity has a potential to generate a significantly diverse interactome. The literature analysis presented here gives access to protein interactions that are not yet contained in public resources and in particular, that are linked to transcript isoforms generated by alternative splicing and stored in HumanSDB3. TBIID is accessible at http://tbiid.emu.edu.tr serving as an up to date and comprehensive resource for future experiments on isoform interactions. Keywords: alternative splicing, protein isoforms, biomedical text mining, abstract retrieval, interaction abstract selection, protein-protein interaction extraction, machine learning, interaction variability analysis. ……………………………………………………………………………………………………………………………………………………………………………………………………………………