Let Me Cypher That for You: A Multi-Graph Agent for Deciphering Biomedical Knowledge

The human body is a vast, interconnected network where a problem in one area can ripple out to affect distant systems. For years, scientists have suspected that the lingering effects of the SARS-CoV-2 virus, known as long COVID, might do more than just cause fatigue or breathing issues; they might also accelerate the decline of the brain, contributing to conditions like Alzheimer's or Parkinson's disease. To investigate these complex connections, researchers build massive digital maps called knowledge graphs. Imagine a giant web where every node is a piece of biological information—a gene, a protein, a symptom, or a disease—and every line connecting them represents a known relationship. These maps hold the key to understanding how a virus might trigger a chain reaction leading to neurodegeneration. However, for most doctors and biologists, these maps are locked behind a high wall of technical language. To ask a question of the map, a user must write a specific, complex code known as Cypher, a skill that requires specialized training and is often too difficult for the very people who need the answers most.

A team of researchers at the Fraunhofer Institute for Algorithms and Scientific Computing in Germany has built a new tool designed to tear down that wall. They created an interactive system called CGEx, which acts as a translator between human curiosity and these complex digital maps. Instead of forcing researchers to learn a new programming language, the system allows them to simply type a question in plain English, such as "How does the virus affect the nervous system?" The software then takes that question, figures out the structure of the specific map being used, and automatically writes the correct code to find the answer. It then runs that code, gathers the results, and translates the raw data back into a clear, detailed explanation. The researchers tested this system on two different, massive maps containing thousands of connections between COVID-19 and neurodegenerative diseases. They found that the system could successfully generate the correct code and provide accurate answers for a significant portion of questions asked, even when the questions required tracing a path through multiple steps of biological relationships, though performance varied depending on the specific map and system version.

The core of this work is a system that learns and adapts as it works. When a user asks a question, the software first looks at the specific map to understand what kinds of data it contains, much like a librarian checking the catalog before searching for a book. It then uses a powerful language model to construct a query, but it does not guess blindly. The system is guided by strict rules and examples that teach it how to handle the specific quirks of the biological data, such as knowing that "neurodegeneration" and "nervous system" might refer to the same concept in different parts of the map. If the system gets stuck or produces a result that doesn't make sense, it can ask the user for feedback. If a user approves a result, the system saves that success as a lesson for the future; if they disapprove, the system tries again, using the feedback to refine its approach. This creates a cycle where the tool becomes smarter and more accurate with every interaction.

The researchers evaluated their system using fifty-two carefully crafted questions that covered a wide range of biological scenarios, from simple connections to complex, multi-step reasoning. They compared the performance of their original system against an upgraded version that used a more advanced language model. The results showed a significant improvement in accuracy. The original system was able to generate correct answers for about 83 percent of the questions on one map and roughly 56 percent on the other. The upgraded version, however, achieved a success rate of 94 percent on the first map and 83 percent on the second. This jump in performance demonstrates that the system is not just a simple translator but a robust tool capable of handling the messy, inconsistent nature of real-world biological data.

One of the most important findings is how the system handles the gaps and inconsistencies that exist in scientific data. Biological knowledge is often stored in different ways across different databases, and sometimes the information a researcher needs simply isn't there. The researchers found that their system could navigate these challenges by looking for alternative paths or synonyms. For instance, when asked about therapies that target a specific protein, the system learned to look for words like "inhibitor," "blocker," or "antagonist" if the exact word "therapy" was missing from the map. In cases where the data was truly missing, the system did not simply blend its own training knowledge to fabricate an answer; instead, it identified the likely cause of the failure, such as missing nodes or absent intermediate relations, effectively surfacing gaps in the knowledge graph while still offering insights grounded in both the retrieved evidence and broader clinical understanding.

The study also highlighted the value of keeping different maps separate rather than trying to merge them into one giant, uniform database. The researchers found that their system could query two different maps simultaneously, using the strengths of one to fill in the weaknesses of the other. This approach proved more practical and flexible than the traditional method of trying to force all data into a single, rigid structure. By allowing the system to access diverse sources of evidence directly, the researchers were able to uncover connections that might have been missed if the data had been forced into a single format. The system successfully traced complex pathways, such as how a virus might enter the brain, trigger inflammation, and eventually lead to cognitive decline, providing a visual and textual explanation of these mechanisms.

Ultimately, this work represents a shift in how scientists interact with their data. It moves away from the idea that only experts in computer science can unlock the secrets of biological networks. By removing the technical barrier of writing code, the system opens the door for clinicians, biologists, and other researchers to explore complex hypotheses directly. The tool does not replace the need for human judgment; rather, it acts as a powerful assistant that handles the heavy lifting of data retrieval, allowing experts to focus on interpreting the results and generating new ideas. The researchers conclude that this approach offers a practical and scalable way to explore the links between infectious diseases and long-term health conditions, potentially leading to new insights into how to prevent or treat neurodegenerative diseases in the post-pandemic world.