🤖 AI

MultiCNKG: Integrating Cognitive Neuroscience, Gene, and Disease Knowledge Graphs Using Large Language Models

This paper introduces MultiCNKG, a novel framework that leverages large language models to integrate Cognitive Neuroscience, Gene, and Disease ontologies into a cohesive, multi-layered knowledge graph, demonstrating high precision and robustness to advance applications in personalized medicine and cognitive disorder diagnostics.

Original authors: Ali Sarabadani, Kheirolah Rahsepar Fard

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ali Sarabadani, Kheirolah Rahsepar Fard

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a massive, three-dimensional puzzle, but the pieces are scattered across three different rooms, and each room speaks a slightly different language.

Room 1 (Genes): Contains pieces about how our DNA works (like the instruction manual for the body).
Room 2 (Diseases): Contains pieces about what goes wrong when the body gets sick (like the "error reports").
Room 3 (Cognitive Neuroscience): Contains pieces about how our brains think, remember, and pay attention (like the "user experience" of the mind).

For a long time, scientists had to manually try to connect these pieces. It was slow, and they often missed how a specific gene might lead to a specific disease that then affects a specific way of thinking.

Enter "MultiCNKG": The Super-Translator and Puzzle Master

This paper introduces a new system called MultiCNKG. Think of it as a super-smart robot librarian (powered by a Large Language Model, or "LLM," like the technology behind advanced chatbots) that can walk into all three rooms at once.

Here is how it works, using simple analogies:

1. The Collection (Gathering the Pieces)

The researchers started with three existing "libraries" of information:

Gene Ontology: A huge list of genetic functions (43,000 items).
Disease Ontology: A list of illnesses (11,200 items).
Cognitive Neuroscience KG: A map of brain processes like memory and attention (2,900 items).

2. The Translation (Making Them Speak the Same Language)

The problem was that "Alzheimer's" in the disease room might be written differently than in the gene room. The LLM acts like a universal translator. It reads the descriptions of every item and says, "Ah, these two are actually the same thing!" or "These two are very closely related."

It cleans up the mess, removes duplicates, and aligns the pieces so they fit together perfectly.

3. The Expansion (Finding Hidden Connections)

This is the magic part. A normal database only shows connections that humans have already written down. But this AI is smart enough to guess new connections based on what it knows.

Analogy: Imagine you have a map of a city. A normal map shows the roads that exist. This AI looks at the map and says, "Based on how these neighborhoods are built, there must be a hidden path connecting this park to that hospital, even if no one has drawn it yet."
The system then checks these new guesses with human experts to make sure they make sense.

4. The Result: A Unified Map

The final product is a single, giant map (Knowledge Graph) with:

6,900 nodes (the puzzle pieces: genes, diseases, brain processes, pathways, and treatments).
11,300 edges (the connections: things that "cause" other things, things that "regulate" others, etc.).

This map allows you to trace a path from a tiny genetic code all the way to a complex human behavior, seeing the whole story in one place.

How Good Is It? (The Report Card)

The researchers tested this new map to see if it was accurate and useful.

Accuracy: It got about 85% of the connections right (Precision) and found 87% of the connections it was supposed to find (Recall).
Expert Approval: When real brain and disease experts looked at the new connections the AI found, 89.5% of them said, "Yes, this makes scientific sense."
Prediction Power: The system was tested on its ability to guess missing links (like a detective solving a crime). It performed very well, beating or matching other famous systems used in computer science.

The Bottom Line

The paper claims that by using a powerful AI to merge these three separate worlds of knowledge, they created a tool that is more complete and coherent than any of the original parts alone. It successfully bridges the gap between our DNA, our diseases, and our minds, creating a clearer picture for researchers to study.

The authors note that while the system is powerful, it currently relies on expensive, proprietary AI tools, and they hope to make it more open and scalable in the future.

Technical Summary: MultiCNKG – Integrating Cognitive Neuroscience, Gene, and Disease Knowledge Graphs Using Large Language Models

Problem Statement

The integration of biomedical and cognitive data faces significant challenges due to the fragmentation of knowledge across distinct domains: genetic mechanisms (Gene Ontology), neurological disorders (Disease Ontology), and cognitive processes (Cognitive Neuroscience Knowledge Graph). Traditional machine learning methods, particularly those relying on shallow embeddings like Bag-of-Words or Word2Vec, struggle to capture the polysemous nature of biomedical terms and the deep semantic links required to connect molecular data with behavioral outcomes. Furthermore, existing Knowledge Graphs (KGs) often lack the cohesive structure necessary to trace pathways from genes to diseases to cognitive functions. The paper posits that Large Language Models (LLMs), with their extensive pre-trained knowledge and context-aware capabilities, offer a solution to overcome the semantic limitations of traditional Graph Neural Networks (GNNs) and facilitate the integration of these disparate data sources.

Methodology

The authors propose MultiCNKG, a framework designed to merge three distinct knowledge sources into a unified, multi-layered KG. The methodology proceeds through four primary stages:

Data Collection and Preprocessing:
- Sources: The framework integrates the Cognitive Neuroscience Knowledge Graph (CNKG), Gene Ontology (GO), and Disease Ontology (DO).
- Preprocessing: Raw text descriptions of entities are tokenized and normalized. LLMs (specifically GPT-4 and BioGPT) are utilized to map entities with multiple forms (e.g., "Alzheimer's disease" vs. "AD") to canonical forms using embedding similarity. Noise removal is performed via threshold-based filtering on similarity scores.
Knowledge Graph Representation:
- The integrated graph is modeled as a directed multigraph $G = (V, E, R)$ , where nodes represent genes, diseases, and cognitive concepts, and edges represent relations such as "causes," "associated with," and "regulates."
- The structure is encoded using adjacency matrices and knowledge triples $(h, r, t)$ .
Graph Alignment and Integration:
- Node Alignment: Nodes from the three source graphs are merged if their embedding similarity exceeds a predefined threshold $\tau$ .
- Edge Alignment: Equivalent relations (e.g., "causes" vs. "induces") are unified through LLM-based semantic comparison.
- Adjacency Update: The final adjacency matrix combines the source matrices plus a delta matrix ( $\Delta A$ ) representing newly discovered relations.
Graph Expansion using LLMs:
- Relation Prediction: LLMs estimate the probability of new relations between nodes ( $P(r_{new}|h, t)$ ). If the probability exceeds a threshold, the relation is added.
- Iterative Expansion: The graph evolves iteratively, with newly discovered entities and edges ( $\Delta G$ ) added to the existing structure.
- Validation: New relations are assigned confidence scores and validated via external databases or manual review by domain experts. Relations with confidence dropping below a threshold are removed.

Key Contributions

MultiCNKG Framework: The development of a novel framework that successfully integrates CNKG, GO, and DO into a single cohesive graph containing 6.9K nodes and 11.3K edges.
LLM-Driven Integration: The application of LLMs (GPT-4) not just for text processing, but for critical graph operations including entity alignment, semantic similarity computation, and the discovery of novel, previously undiscovered connections between genetic, disease, and cognitive domains.
Multi-Layered Architecture: The creation of a graph that explicitly links three layers: the Genetic Layer (GO), the Disease Layer (DO), and the Cognitive Neuroscience Layer (CNKG), enabling a holistic view from molecular mechanisms to behavioral constructs.
Comprehensive Evaluation: A rigorous evaluation of the integrated graph using traditional metrics (Precision, Recall, F1-Score) and specialized KG metrics (Graph Consistency, Novelty Detection, Expert Validation), alongside link prediction benchmarks.

Results

The performance of MultiCNKG was evaluated against its source components and standard benchmarks:

Integration Metrics:
- Precision: 85.20%
- Recall: 87.30%
- Coverage: 92.18% (indicating successful transfer of data from source graphs)
- Graph Consistency: 82.50%
- Expert Validation: 89.50% (confirming biological and cognitive plausibility of new edges)
- Novelty Detection: 40.28% (representing the proportion of edges discovered by the LLM that were not in the original graphs).
- Comparison: MultiCNKG demonstrated improved recall and coverage compared to the individual source graphs (GO, DO, CNKG), though precision was slightly lower than GO alone, likely due to the inclusion of novel, less certain connections.
Link Prediction:
- The graph was tested using models such as TransE and RotatE.
- TransE: Mean Rank (MR) of 391, Mean Reciprocal Rank (MRR) of 0.411.
- RotatE: MR of 263, MRR of 0.395.
- These results show competitive performance against standard benchmarks like FB15k-237 and WN18RR, validating the graph's structural integrity and predictive capability.

Significance and Claims

The paper claims that MultiCNKG represents a significant advancement in biomedical knowledge integration by leveraging the semantic depth of LLMs to bridge gaps between molecular genetics, neurological disorders, and cognitive functions.

Scientific Utility: The framework facilitates a multi-layered view that allows researchers to trace pathways from genes to diseases to cognitive processes, supporting applications in personalized medicine, cognitive disorder diagnostics, and hypothesis formulation.
Methodological Shift: The study demonstrates that LLMs can effectively compensate for the contextual and semantic shortcomings of traditional GNN pipelines, particularly in tasks requiring implicit graph structure reasoning and multi-hop inference.
Future Potential: While acknowledging challenges regarding scalability and reliance on proprietary models (GPT-4), the authors posit that MultiCNKG sets a foundation for interdisciplinary research. They suggest future work could involve integrating additional ontologies (e.g., DrugBank) and utilizing open-source LLMs to enhance transparency and support drug repurposing and early detection of cognitive disorders.