From General-Purpose to Disease-Specific Features: Aligning LLM Embeddings on a Disease-Specific Biomedical Knowledge...

Area of Science:

• Biomedical informatics
• Computational pharmacology
• Neuroscience

Background:

• Drug repurposing for neurodegenerative diseases like Alzheimer disease and related dementias (ADRD) is challenging due to limited, sparse, and heterogeneous data.
• General Large Language Model (LLM) embeddings lack specific biomedical context for tasks like drug repurposing.

Purpose of the Study:

• To introduce CLEAR (Contextualizing LLM Embeddings via Attention-based gRaph learning), a novel framework to improve computational drug repurposing.
• To align LLM embeddings with knowledge graph structures for enhanced biomedical inference.

Main Methods:

• Developed CLEAR, a multimodal representation-fusion framework.
• Integrated LLM embeddings with the topological structure of context-specific Knowledge Graphs (KGs) using attention-based graph learning.
• Validated CLEAR across five benchmark datasets.

Main Results:

• CLEAR achieved state-of-the-art results, improving predictive performance (e.g., F1 score) by up to 30% over existing methods.
• Applied CLEAR to identify potential drug repurposing candidates for ADRD, including Parkinson disease-related dementia and Lewy Body dementia.
• Demonstrated CLEAR's ability to learn a biologically coherent embedding space and prioritize drug candidates.

Conclusions:

• Grounding biomedical LLM embeddings with KG signals significantly improves drug repurposing in data-sparse settings.
• CLEAR offers a promising approach for identifying new therapeutic uses for existing drugs, particularly for complex conditions like ADRD.
• The framework successfully identified leading ADRD drug candidates and summarized therapeutic relationships.