Scalable Text Mining Assisted Curation of PTM Proteoforms in the Protein Ontology

Karen E. Ross, Darren A. Natale, Cecilia N. Arighi, Sheng-Chih Chen, Hongzhan Huang, Gang Li, Jia Ren, Michael Wang, K. Vijay-Shanker, Cathy H. WuICBO/BioCreative 2016 · OpenReview

Abstract

Post-translational modifications (PTMs) are critical regulators of protein function, yet the manual curation of these "proteoforms" into structured ontologies like the Protein Ontology (PRO) remains a significant bottleneck. This paper presents an automated pipeline that leverages advanced text mining tools (RLIMS-P and eFIP) and the iPTMnet database to identify and curate PTM proteoforms at scale. By applying stringent filtering and entity normalization, the system achieved a 50% increase in PRO's coverage of species-specific PTM forms with 91% precision, demonstrating a scalable path for turning unstructured scientific literature into high-quality ontological data.

Introduction

Proteins rarely function in a static state; they are frequently modified after translation through processes such as phosphorylation. These modifications act as molecular switches regulating cell signaling and disease progression. To capture this complexity, the Protein Ontology (PRO) defines and categorizes specific "proteoforms." However, the sheer volume of scientific literature makes manual curation unsustainable. Prior to this study, PRO contained approximately 2,550 PTM proteoform classes, many of which were imported in bulk. There is a critical need for automated, scalable methods to extract these specific biological details from unstructured text.

Methodology

The researchers developed an automated pipeline designed to bridge the gap between raw text mining and standardized curation. The workflow integrates several specialized components to ensure data integrity and biological relevance.

Overview of the PTM-focused text mining and curation pipeline.

Key components of the pipeline include:

  • RLIMS-P & eFIP: Specialized text mining tools used to identify mentions of kinases, substrates, and specific phosphorylation sites. eFIP specifically detects if a phosphorylation event alters protein-protein interactions.
  • iPTMnet Integration: Acts as a central hub, merging text mining results with curated data from sources like PhosphoSitePlus.
  • Entity Normalization: Protein names are mapped to unique UniProtKB identifiers using tools like PubTator to resolve nomenclature ambiguity.
  • Automated Stanza Generation: Validated PTM sites (cross-referenced against protein sequences) are used to automatically generate PRO "stanzas"—structured definitions containing logical relationships and literature evidence.
  • Stringent Filtering: To maintain high quality, the pipeline discards approximately 85% of results, focusing on high-confidence, singly-phosphorylated sites that are less prone to machine-parsing errors.

Datasets

The study utilized a massive corpus of approximately 25 million PubMed abstracts and PMC full-length open-access articles. iPTMnet served as the primary integration platform, while UniProtKB provided the necessary protein sequence validation and identifier mapping. The resulting terms were targeted for inclusion in the Protein Ontology (PRO).

Key Results and Discussion

The pipeline successfully identified approximately 820 new substrate-site pairs for inclusion in PRO, representing a 50% increase in the ontology's coverage of species-specific PTM proteoforms. A manual review of 91 randomly selected results confirmed a precision of 91% (83/91).

Beyond simple identification, the pipeline provides rich functional context. Of the new proteoforms, 89% were linked to specific kinases, and 11% included data on phosphorylation-dependent interactions. The power of this approach is exemplified by the mapping of the PIN1 protein network, which illustrates how multiple kinases and interactions across species can be automatically synthesized into a functional network.

Automated mapping of the PIN1 proteoform network across human and mouse models.

The "Open World" Advantage

The authors utilize the "Open World Assumption," meaning that defining a protein as phosphorylated at one site does not preclude modifications at others. This allows the ontology to grow incrementally as new evidence emerges, rather than requiring a complete, exhaustive picture of a protein before any data can be added.

Limitations and Future Directions

While highly effective for single modifications, the pipeline faces challenges with "multi-site" modifications—determining if two different modifications mentioned in a paper occur on the same physical molecule. Additionally, while the pipeline is robust, approximately 90% of phosphorylation sites are mentioned only in the full text of articles rather than abstracts, highlighting the need for broader access to full-length, often paywalled, literature. Future work may also involve integrating "bottom-up" proteomics data to validate complex multi-site modifications that text mining alone cannot yet confirm.

Summary updated

BibTeX
@inproceedings{ross2016scalable,
  author = {Karen E. Ross and Darren A. Natale and Cecilia N. Arighi and Sheng-Chih Chen and Hongzhan Huang and Gang Li and Jia Ren and Michael Wang and K. Vijay-Shanker and Cathy H. Wu},
  title = {{Scalable Text Mining Assisted Curation of PTM Proteoforms in the Protein Ontology}},
  booktitle = {ICBO/BioCreative},
  year = {2016},
  url = {https://openreview.net/forum?id=93HOmtmWNMB},
}
Figure