LBD-COVID19

LBD-COVID19 applies literature-based discovery and knowledge graph completion to identify and prioritize potential drug repurposing candidates for COVID-19.

Key Features:

• Literature-Based Discovery: Extracts semantic triples using SemRep via the SemMedDB database from PubMed and other COVID-19 research sources to identify relationships between drugs and disease mechanisms.
• Semantic Triple Extraction and Filtering: Applies filtering rules and an accuracy classifier based on a BERT variant (PubMedBERT) to refine semantic predications, with the classifier achieving an F1 score of 0.854.
• Knowledge Graph Construction and Completion: Constructs a knowledge graph from filtered semantic triples and applies five neural knowledge graph completion algorithms—TransE, RotatE, DistMult, ComplEx, and STELP—with TransE reporting MR = 0.923 and Hits@1 = 0.417.
• Time Slicing for Model Training and Assessment: Uses a time slicing approach to train and evaluate models to ensure temporal relevance of predictions.
• Discovery Pattern-Based Approach: Utilizes discovery patterns alongside knowledge graph completion to identify additional candidate drugs and generate hypotheses.

Scientific Applications:

• Drug Repurposing Identification: Identifies known drugs linked to COVID-19 in the literature and proposes novel candidates including paclitaxel, SB 203580, alpha 2-antiplasmin, metoclopramide, and oxymatrine, with associated mechanistic explanations.
• Hypothesis Generation: Generates plausible hypotheses about drug–COVID-19 interactions to support further experimental validation and research.

Methodology:

Semantic triple extraction with SemRep via SemMedDB from PubMed; filtering rules and an accuracy classifier based on PubMedBERT (F1 = 0.854); knowledge graph construction; neural knowledge graph completion using TransE, RotatE, DistMult, ComplEx, and STELP; time slicing for training and evaluation; and discovery-pattern analysis.