Alzheimer’s, a form of dementia that affects memory, thinking, and behaviour, remains a colossal challenge. In the US alone, direct annual costs are estimated at over $360 billion. But that figure barely scratches the surface of the broader economic burden, which may reach as high as $832 billion when factoring in indirect costs and societal impact.
As with many of today’s most pressing challenges, there’s growing hope that Artificial Intelligence (AI) could help uncover better ways to prevent and treat Alzheimer’s. A team of researchers at Cedars-Sinai Medical Center, one of the largest nonprofit academic medical centers in the US, may be uncovering promising clues through their innovative work developing advanced AI to predict risk and discover new drugs for Alzheimer’s. The results have been game-changing.
Crucially, they’re not doing this through one AI solution, but by combining AI with advanced software tools and database technologies, particularly knowledge graphs and GraphRAG (Graph Retrieval-Augmented Generation).
Aiming to directly drive innovation in treatment, the team is leveraging graph technology to advance what it describes as “Knowledge-Aware” automated machine learning. The team’s efforts centre on two main fronts: revealing overlooked genetic markers that may play a role in the condition, and pinpointing already-FDA-approved drugs that could be repurposed to expand current treatment options.
Building a knowledge graph with over 1.6 million edges
One of the first fruits of this work is the Alzheimer’s Disease Knowledge Base (AlzKB) using a graph database as the core framework. This knowledge graph integrates over 20 biomedical sources, covering genetic, drug, and disease relationships through a comprehensive ontology. It is already guiding machine learning systems to insights that would have otherwise been missed.
The current version contains 234,000 nodes and 1.67 million edges, and to make AlzKB accessible to non-technical users, the team has developed KRAGEN, a knowledge graph-enhanced RAG (Retrieval-Augmented Generation) system, tailored specifically for Alzheimer’s.
KRAGEN encodes graph data into vectors, enabling LLMs to generate contextually accurate responses by drawing directly from AlzKB’s structured knowledge. This means researchers and clinicians don’t need to speak a specialist language like Cypher to extract insights–they can simply ask questions in plain English.
KRAGEN, AlzKB, and Memgraph together power the automated ML pipelines that the Cedars-Sinai team uses to predict Alzheimer’s risk and identify potential drug candidates. KRAGEN combines semantic relationships within AlzKB with large language models (LLMs), allowing nuanced question-answering. And with Graph of Thoughts (GoT), another tool inspired by recent AI-knowledge graph fusion work out of Hugging Face, integrated into the system, compound questions are split into manageable parts, which the ML models answer before combining for a final response. The GoT capabilities come from a purpose-built AI agent built by the team, ESCARGOT, which generates Cypher queries in real time. ESCARGOT enables on-the-fly decomposition of complex queries, dynamically retrieving data from Memgraph for precise, compound answers.
“Our goal is to inform machine learning so that it can do a better job of things like feature selection, model selection, and model interpretation,” says Jason Moore, who leads Cedars-Sinai’s Department of Computational Biomedicine. “And we want our biology and clinical users to be able to query the knowledge base without having to learn to program to query the knowledge graph.”
He adds: “Essentially, ESCARGOT makes the ‘graph of thoughts’ process more dynamic. It breaks down the query into components and then generates Python code to execute a Cypher query directly in Memgraph, bypassing the vector database and RAG approach.”
Essentially, this combination will allow researchers to understand, with relative ease, how different entities connect, which drugs may treat Alzheimer’s, how one gene regulates another, and how these influences cascade.
This, the team believes, will surface new genetic risk factors—and, ultimately, new drugs for Alzheimer’s. But rather than tread well-worn paths where global research has already intensely focused on the top 100 genes linked to the disease and a catalogue of known drug leads, the team is deliberately looking elsewhere. “Lots of people are looking under the lamppost of a known gene,” Moore says. “What I’m interested in is the novel discoveries over ‘there’, out in the dark.”
Using graphs and AI to shine a light
From the outset, Moore knew that an off-the shelf LLM simply wouldn’t be able to handle, process, and understand the amount of complex data they were working with: “ChatGPT can answer a question about a gene or a drug or Alzheimer’s, but it cannot really put all those entities together and understand their relationships in any kind of a complete way. ChatGPT was no better right out of the box than flipping a coin at answering questions.”
Instead, a knowledge graph offers a higher-order synthesis of countless pieces of data—connections, context, and relationships. As a result, Cedars-Sinai has radically improved the effectiveness of its natural language interface, unlocking far more of the insight and intelligence AlzKB was built to provide. For Moore, the significance of this combination of graph, advanced data analysis and AI itself has big implications.
Why? Because this is translating to a stronger and more-nuanced biological understanding of Alzheimer’s: “I think,” he says, “we’ve shown how we can use the knowledge in the knowledge graph to inform machine learning to give us new ideas for treating this disease.”
Although still in its early phases, the research is already shedding light on genetic factors linked to Alzheimer’s that have gone unnoticed until now. It raises an intriguing prospect: that familiar, widely used drugs—such as Temazepam, commonly prescribed for insomnia (a symptom often seen in Alzheimer’s patients), and even Ibuprofen, a standard remedy for headaches—could potentially be repurposed or modified to help in much wider areas of Alzheimer’s care.
They also report the additional benefit of vastly improved reasoning. The system is achieving a 94.2% accuracy rate for multi-hop reasoning—far exceeding ChatGPT’s standalone performance of 49.9%.
Are graph-based technologies the key to unlocking AI-driven, personalized medical research and development?
The emerging goal is a fully automated machine learning platform, once again powered by a knowledge graph. Soon, users of AlzKB will be able to input queries like, “Identify the genes associated with this drug and disease,” or “Generate a dataset limited to these specific genes,” and the AI will handle the rest—executing algorithms, surfacing key variables, and delivering results that are both clear and actionable for researchers.
This capability promises to fast-track meaningful advances in Alzheimer’s research, and not in the distant future, but today.
Across a range of innovative initiatives, developers are embracing graph technology and GraphRAG to fully leverage the potential of big medical data and artificial intelligence.
