MultiCNKG: Unified LLM-Assisted Knowledge Graph

• MultiCNKG is a unified knowledge graph integrating CNKG, GO, and DO to connect genes, diseases, and cognitive functions.
• It employs LLM-assisted semantic alignment to standardize nodes and relations, enabling coherent cross-domain reasoning.
• Quantitative evaluation shows high precision, recall, and expert validation, supporting applications like personalized medicine and hypothesis generation.

MultiCNKG is a unified, LLM-assisted knowledge graph that integrates the Cognitive Neuroscience Knowledge Graph (CNKG), Gene Ontology (GO), and Disease Ontology (DO) in order to bridge molecular mechanisms, diseases, and cognitive processes within a single resource. Its stated purpose is to overcome semantic and structural fragmentation across neuroscientific, genetic, and disease knowledge sources and to support end-to-end reasoning across the molecular-to-behavioral spectrum, including hypothesis generation, diagnostic support, and personalized medicine (Sarabadani et al., 8 Oct 2025).

1. Definition, scope, and source graphs

MultiCNKG is designed to harmonize three heterogeneous source KGs: CNKG, which emphasizes cognitive constructs and tasks; GO, which encodes gene functions, processes, and components; and DO, which organizes disease concepts. The integrated graph is intended to interconnect genes, biological pathways, diseases, and cognitive processes so that multi-hop reasoning can proceed from genotype to phenotype and behavior (Sarabadani et al., 8 Oct 2025).

Graph	Size	Schema
CNKG	2.9K nodes, 4.3K edges	9 node types, 20 edge types
GO	43K nodes, 75K edges	3 node types, 4 edge types
DO	11.2K nodes, 8.8K edges	1 node type, 2 edge types
MultiCNKG	6.9K nodes, 11.3K edges	5 node types, 7 relation types

The final integrated MultiCNKG contains 6.9K nodes across five node types: Genes, Diseases, Cognitive Processes, Biological Pathways, and Therapeutic Targets. Its 11.3K edges span seven relation types: Causes, Associated with, Regulates, Involved in, Treated by, Influences, and Linked to (Sarabadani et al., 8 Oct 2025).

The paper describes MultiCNKG as a multi-layered graph that enables translational applications such as personalized medicine through gene–disease–target paths, cognitive disorder diagnostics through symptom-to-mechanism mapping, and hypothesis generation through plausible gene–disease–cognition links not present in any single source KG. A plausible implication is that the system is positioned not merely as a merged ontology, but as a reasoning substrate that connects levels of analysis that are usually dispersed across separate biomedical and cognitive resources.

2. Schema harmonization and representational design

MultiCNKG standardizes both node classes and relation labels across the three source graphs. Semantically equivalent relations were unified through LLM-assisted semantic comparison together with ontology constraints; the reported examples include standardizing causes versus induces and associated_with versus related_to to canonical labels (Sarabadani et al., 8 Oct 2025).

Node-type consolidation reduced source-specific classes to five representative categories. GO’s molecular function, process, and component terms were aligned to Genes and Biological Pathways; DO’s disease concepts were aligned to Diseases; and CNKG constructs were mapped to Cognitive Processes. This normalization step is central to the graph’s claim of semantic coherence because it constrains downstream reasoning to a compact schema rather than preserving source-specific heterogeneity (Sarabadani et al., 8 Oct 2025).

The graph itself is represented as a directed multigraph $G = (V,E,R)$ with triples $(h,r,t)$, where $h,t \in V$ and $r \in R$. Reported examples include $$(\mathrm{APOE4}, \mathrm{causes}, \mathrm{Alzheimer's\ disease})$$ and $$(\mathrm{Alzheimer's\ disease}, \mathrm{impairs}, \mathrm{Memory})$$. Adjacency is encoded by a matrix $A$ with $A_{ij}=1$ if a relation exists between $v_i$ and $v_j$ (Sarabadani et al., 8 Oct 2025).

This representational choice matters because MultiCNKG is not presented as a simple table of aligned terms. It is explicitly a graph whose semantics are carried by typed nodes, typed edges, and graph-structural update rules, allowing the system to support cross-layer traversals from genes to disorders to cognition.

3. Integration pipeline and LLM-mediated augmentation

The integration pipeline is described as a sequence of ingestion, normalization, alignment, and LLM-based augmentation. During ingestion and normalization, entities’ textual descriptions were tokenized as $(h,r,t)$0. Canonicalization and synonym resolution were performed by embedding similarity, and semantic closeness was quantified by cosine similarity,

$(h,r,t)$1

Duplicates and low-confidence entities or edges were filtered by threshold-based noise removal (Sarabadani et al., 8 Oct 2025).

Graph alignment merged entities across CNKG, DO, and GO when the similarity score satisfied $(h,r,t)$2, where $(h,r,t)$3 was set empirically. Equivalent relations were unified across sources through LLM-guided semantic comparison, and the final adjacency was updated by combining source adjacencies with $(h,r,t)$4, where $(h,r,t)$5 captures newly discovered relations from LLM augmentation (Sarabadani et al., 8 Oct 2025).

The augmentation stage uses LLMs to propose new relations between candidate node pairs. For nodes $(h,r,t)$6, the LLM estimates the probability of a new relation,

$(h,r,t)$7

and accepts a proposal if $(h,r,t)$8. An alternative or complementary similarity-based probabilistic model is also reported:

$(h,r,t)$9

Iterative expansion is expressed as

$h,t \in V$0

with $h,t \in V$1 (Sarabadani et al., 8 Oct 2025).

GPT-4 is identified as the core model for alignment, semantic comparison, and relation proposal, while BioGPT assists with preprocessing for biomedical terminology. The LLMs were instructed to decide entity equivalence and propose canonical labels, map relation semantics to the unified schema, and suggest plausible cross-ontology relations with supporting rationale. The exact prompt templates and numeric thresholds are not disclosed. To reduce hallucinations, the paper reports confidence-based thresholds $h,t \in V$2, OWL-based consistency checks, external database cross-references, and expert-in-the-loop validation (Sarabadani et al., 8 Oct 2025).

A notable feature of this pipeline is that validation is iterative rather than terminal. Confidence scores are assigned, checked against external sources or domain experts, and low-confidence relations can be removed in the next iteration. This suggests a graph curation process in which LLM proposals remain provisional until they survive semantic and expert scrutiny.

4. Quantitative evaluation and predictive performance

MultiCNKG is evaluated along three axes: alignment quality, structural health, and predictive power. Precision, recall, and F1 are defined in the standard way,

$h,t \in V$3

Coverage is defined as the proportion of source content preserved in the merged KG, graph consistency is the proportion of triples passing semantic and structural checks, and novelty detection is the fraction of edges newly discovered by LLM augmentation (Sarabadani et al., 8 Oct 2025).

The reported headline results are precision 85.20%, recall 87.30%, coverage 92.18%, graph consistency 82.50%, novelty detection 40.28%, and expert validation 89.50%. These figures are used to support the claims of high-quality alignment, broad source preservation, semantic coherence, and biologically and cognitively plausible augmentation (Sarabadani et al., 8 Oct 2025).

For link prediction, the paper reports Mean Rank and Mean Reciprocal Rank,

$h,t \in V$4

together with $h,t \in V$5. TransE uses the translational score

$h,t \in V$6

while RotatE represents relations as rotations in complex space and scores triples by

$h,t \in V$7

On MultiCNKG, the reported results are TransE: MR = 391, MRR = 0.411; RotatE: MR = 263, MRR = 0.395. Additional MRRs are DistMult 0.418, ComplEx 0.301, ConvE 0.321, and HolmE 0.211. For MultiCNKG P@10, the reported values are TransE 0.518, RotatE 0.571, DistMult 0.519, ComplEx 0.487, ConvE 0.493, and HolmE 0.403 (Sarabadani et al., 8 Oct 2025).

The training setup follows standard filtered evaluation and negative sampling with margin-based ranking loss typical for KG embeddings, but the paper does not disclose embedding dimensionality, batch sizes, optimizer, or related hyperparameters. The claim of competitive performance is contextualized against benchmarks such as FB15k-237 and WN18RR; for example, RotatE MRR 0.336 is cited for FB15k-237, whereas MultiCNKG reports RotatE MRR 0.395 and P@10 up to 0.571 on its domain-specific graph (Sarabadani et al., 8 Oct 2025).

5. Reasoning patterns, applications, and validation

MultiCNKG’s practical value is illustrated through cross-layer reasoning paths. A reported gene-to-disease-to-cognitive-process chain is $h,t \in V$8. A gene-to-pathway-to-cognitive-process path is $h,t \in V$9. A disease-to-therapeutic-target-to-cognitive-process path is $r \in R$0 (Sarabadani et al., 8 Oct 2025).

These examples are used to motivate three application domains. In personalized medicine, the graph supports gene–disease–target reasoning. In cognitive disorder diagnostics, it supports mapping observed symptoms to candidate molecular mechanisms. In hypothesis generation, it proposes plausible gene–disease–cognition links not present in any single source KG (Sarabadani et al., 8 Oct 2025).

Validation is explicitly multi-layered. Experts in cognitive neuroscience and biomedicine reviewed LLM-proposed edges, yielding an expert validation score of 89.50%. Reported failure modes include misalignment, semantic drift or relation mislabeling, and LLM hallucinations. The corresponding mitigations are high-similarity thresholds and LLM adjudication, ontology-informed standardization with OWL-based consistency checks, and confidence scoring plus external cross-references and removal of low-confidence edges after expert feedback (Sarabadani et al., 8 Oct 2025).

The paper does not report formal ablations, but it states that the metric profile suggests LLM augmentation increases coverage and consistency, while novelty detection is substantial but balanced against precision and expert plausibility constraints. This suggests that novelty is treated as useful only insofar as it survives semantic and expert validation, rather than as an end in itself.

6. Relation to adjacent multi-KG research, limitations, and prospective extensions

Within the broader literature, MultiCNKG can be situated alongside multi-KG completion and alignment methods that formalize fused graphs and cross-graph transfer. CKGC-CKD constructs a fused/global KG by combining local triple sets with seed alignments, treats alignment links as special edges in a global encoder, and couples local and global models through mutual knowledge distillation (Zhang et al., 2023). JMAC jointly addresses multilingual knowledge graph completion and alignment through separate relation-aware encoders, structural inconsistency reduction, and an entropy-mediated mechanism for alignment seed enlargement and triple transferring (Tong et al., 2022).

These systems are not part of the reported MultiCNKG pipeline, which instead centers on LLM-assisted entity alignment, semantic similarity computation, and graph augmentation. However, they provide methodological context for understanding MultiCNKG as part of a larger family of systems that attempt to convert heterogeneous KGs into a coherent structure amenable to completion, alignment, and reasoning. A plausible implication is that future MultiCNKG variants could incorporate more explicit graph-neural or distillation-based transfer mechanisms in addition to LLM mediation.

The reported limitations of MultiCNKG are dependency on proprietary LLMs such as GPT-4, scalability challenges for very large KGs, potential biases inherited from both source KGs and LLM training corpora, and coverage gaps in specific biomedical subdomains (Sarabadani et al., 8 Oct 2025).

Planned enhancements include integration of additional KGs such as DrugBank with 7.4K nodes and 366K edges and PharmKG with 7.6K nodes and 500K edges; adoption of open-source LLMs including BioGPT and LLaMA variants; dynamic updates via federated learning; enhanced validation through crowdsourcing; richer provenance tracking; and advanced causal reasoning to move beyond association to mechanism (Sarabadani et al., 8 Oct 2025).

The paper also notes practical reproducibility limits. Although the graph is defined as a directed multigraph and supports adjacency and triple forms, and although OWL reasoning implies compatibility with RDF/OWL serializations while property-graph representations are described as natural fits, the text does not disclose a public repository, API, exact serialization, specific GO or DO versions, licensing details, or a formal update schedule (Sarabadani et al., 8 Oct 2025).