AI Drug Repurposing System: Knowledge Graph + Large Model Helps Old Drugs Find New Uses

This article introduces an end-to-end drug repurposing system based on knowledge graphs and large language models. Through CrewAI multi-agent collaboration, it quickly discovers new indications for existing drugs, aiming to address the pain points of traditional new drug development—long cycle (10-15 years) and high cost (over $2.6 billion)—and unlock the untapped therapeutic value of already marketed old drugs.

Background

Developing a new drug takes an average of 10-15 years and $2.6 billion, while drug repurposing (old drugs for new uses) is an efficient alternative path. Typical cases include sildenafil (Viagra) switching from a cardiovascular drug to a treatment for erectile dysfunction, and thalidomide changing from a teratogenic drug to a treatment for multiple myeloma.

Traditional Method Dilemmas

• Phenotypic screening: High cost, low efficiency, difficult to explain mechanisms
• Target screening: Relies on understanding of disease mechanisms; most disease targets are unknown
• Clinical observation: Relies on luck, cannot be carried out systematically
• Literature mining: Limited ability of manual processing of massive literature

Traditional methods can no longer meet the needs of the explosive growth of biomedical data.

Core Technologies

1. Knowledge Graph: Integrates multi-source data such as DrugBank and OMIM, constructs a drug-disease-gene association network, and reveals potential therapeutic paths through entity nodes and relationship edges
2. Large Language Model: Processes unstructured literature, extracts entity relationships, understands context, infers hypotheses, and generates verifiable scientific conclusions
3. Multi-agent Collaboration: Under the CrewAI framework, agents for data collection, knowledge integration, hypothesis generation, evidence evaluation, and priority ranking work collaboratively

End-to-end System Architecture

• Data Layer: Integrates structured (database) and unstructured (literature) data and updates in real time
• Knowledge Graph Construction Layer: Entity recognition, relationship extraction, graph embedding (GNN)
• Reasoning Layer: Link prediction, path reasoning, large model enhancement, multi-agent collaboration
• Evaluation Layer: Evaluation of literature evidence, mechanism rationality, safety, and clinical value
• Output Layer: Candidate drug list, mechanism explanation, evidence report, visualization interface

Taking Alzheimer's disease as an example, the system process is as follows:

1. Knowledge graph query: Obtain information such as pathogenic genes (APP/PSEN1/PSEN2), pathological features, and signaling pathways
2. Candidate identification: Find drugs acting on relevant pathways (e.g., anti-inflammatory drugs)
3. Literature mining: Retrieve studies supporting the protective effect of the drug in Alzheimer's disease models
4. Mechanism reasoning: The drug reduces amyloid toxicity by inhibiting neuroinflammation
5. Safety evaluation: Focus on side effects in elderly patients
6. Priority ranking: Generate candidate list and report

This process takes only a few minutes to hours, much faster than the weeks/months required by traditional methods.

Challenges and Responses

1. Data Quality: Noise/inconsistency/missing → Cleaning and standardization + entity linking + quality assessment
2. Graph Sparsity: Few known relationships → Large model supplements literature knowledge + GNN topological reasoning
3. Causal Identification: Correlation ≠ causation → Causal inference + time series data + mechanism knowledge
4. Clinical Translation: Laboratory results are difficult to reproduce → Multi-level verification (computational → in vitro → animal → clinical) + real-world data
5. Interpretability: Black-box prediction → Interpretable reasoning paths + confidence scores + human-machine collaboration

Future Outlook

• Larger-scale knowledge graphs: Cover more entities and relationships
• Multimodal models: Integrate text/molecule/protein/image data
• Agent collaboration upgrade: More autonomous handling of complex tasks
• Dry-wet experiment closed loop: Accelerate candidate verification
• Personalized repurposing: Based on individual patient characteristics

Conclusion

AI technologies (knowledge graph + large model + multi-agent) provide an efficient path for drug repurposing. Although experimental verification and expert judgment are needed, it has become a powerful assistant in drug development, accelerating the transformation from "ten years to sharpen a sword" to "old drugs for new uses", and bringing hope to patients for faster access to effective treatment.