IDPpub

IDPpub mines MEDLINE abstracts using BioBERT-based natural language processing to extract protein substrates and phosphorylation site positions for compiling evidence-backed phosphoproteome annotations.

Key Features:

• Deep learning–based NLP: Employs BioBERT fine-tuned for biomedical text mining to extract phosphorylation-related information from abstracts.
• Training and performance: Optimized on over 3,000 training abstracts with average precision 0.93 and recall 0.94, and an independent assessment reporting precision 0.91 and recall 0.77 after normalization and mapping.
• Entity normalization: Normalizes extracted proteins to gene symbols using NCBI gene queries.
• Site mapping: Maps phosphorylation sites to human UniProt sequences via ProtMapper and to mouse UniProt sequences via direct matching.
• Repository content: Contains 18,458 unique human phosphorylation sites supported by 58,227 abstracts and 5,918 mouse sites supported by 14,610 abstracts, including 1,803 CPTAC sites not present in PhosphoSitePlus manual curation.
• Validation with proteomics datasets: Applied to CPTAC pan-cancer phosphoproteomics datasets and compared with PhosphoSitePlus to identify previously unannotated sites.

Scientific Applications:

• Phosphosite discovery: Facilitates identification of novel phosphorylation sites with primary literature evidence.
• Phosphoproteome annotation: Provides evidence-based annotations to supplement databases such as PhosphoSitePlus.
• CPTAC dataset analysis: Supports interpretation and cross-referencing of CPTAC pan-cancer phosphoproteomics data.
• Functional and disease studies: Aids investigation of phosphosite functional roles, disease mechanisms, and potential therapeutic targets.

Methodology:

BioBERT-based NLP models were fine-tuned on >3,000 MEDLINE abstracts to extract substrate–site pairs; proteins are normalized via NCBI gene queries; sites are mapped to UniProt using ProtMapper for human sequences and direct matching for mouse sequences; performance was evaluated using precision and recall metrics including independent assessment.