Automated Knowledge Discovery in Bioscience

Abstract

Biomedical literature doubles every 4-5 years, with over 1.5 million PubMed-indexed articles published annually and over 38 million articles in the current database. No researcher can read even a fraction of the literature relevant to their field, creating a knowledge fragmentation problem: critical connections between findings in different subfields remain undiscovered for years. Automated knowledge discovery (AKD) uses natural language processing, knowledge graph construction, and reasoning algorithms to extract, integrate, and generate novel hypotheses from the biomedical literature and structured databases. Swanson's (1986) pioneering work demonstrated that literature-based discovery could identify non-obvious connections (fish oil treats Raynaud's disease) years before experimental confirmation. Modern AKD systems leverage large language models (LLMs), biomedical knowledge graphs, and multi-modal data integration. We present the Automated Knowledge Discovery Assessment Framework (AKDAF), evaluating five AKD approaches -- biomedical NLP entity extraction, knowledge graph link prediction, LLM-powered hypothesis generation, literature-based discovery systems, and multi-modal knowledge fusion -- across four bioscience discovery tasks (drug repurposing, gene-disease association, pathway cross-talk identification, and biomarker discovery). Our Knowledge Discovery Effectiveness Score (KDES) measures extraction precision, novel hypothesis quality, experimental validation rate, discovery speed, and interpretability. LLM-powered hypothesis generation achieves the highest KDES (0.922) through contextual reasoning over biomedical corpora, while knowledge graph link prediction achieves the highest experimental validation rate (0.940) through structured relational inference