HypoChainer: A Collaborative System Combining LLMs and Knowledge Graphs for Hypothesis-Driven Scientific Discovery.

Area of Science:

• * Computational biology
• * Artificial intelligence in science
• * Knowledge discovery

Background:

• * Modern scientific discovery faces challenges due to vast data and human cognitive limits.
• * Deep learning models like graph neural networks (GNNs) generate numerous predictions, overwhelming manual validation.
• * Large language models (LLMs) show promise but struggle with reliability issues like hallucinations and lack of structured knowledge grounding.

Purpose of the Study:

• * To introduce HypoChainer, a collaborative visualization framework for enhancing scientific discovery.
• * To address the limitations of manual filtering of GNN predictions and unreliable LLM outputs.
• * To integrate human expertise, LLM reasoning, and knowledge graphs (KGs) for visual hypothesis generation.

Main Methods:

• * Retrieval-augmented LLMs (RAGs) and visualizations for contextual exploration of GNN predictions.
• * Iterative hypothesis construction using KG exploration and LLM-guided refinement.
• * Visual analytics for prioritizing predictions based on refined hypothesis chains and KG evidence for validation selection.

Main Results:

• * HypoChainer facilitates efficient filtering of GNN predictions for experimental validation.
• * The framework supports domain experts in generating and refining novel scientific hypotheses.
• * Visual exploration and LLM-driven reasoning improve the reliability and focus of scientific inquiry.

Conclusions:

• * HypoChainer offers a novel approach to overcome challenges in large-scale scientific discovery.
• * Integrating human expertise with AI tools like LLMs and KGs is crucial for advancing research.
• * The visual framework enhances the efficiency and effectiveness of hypothesis generation and validation selection.