Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record, with a small difference in title or year; 95 of the 95 references it cites were found in a registry or an index; 6 of the 11 sources its claims cite were checked for support.

GOLDEN fusion: a graph-oriented learning with domain-embedding network fusion for generating super gene sets in functional genomics

Open access
Li, Qi, Nichols, Cody, Welner, Robert S, Chen, Jake Y, Ku, Wei-Shinn, Yue, Zongliang · 2026 · Briefings in Bioinformatics

Abstract

The integrative analysis of gene sets, networks, and pathways is pivotal for deciphering omics data in translational biomedical research. To significantly increase gene coverage and enhance the utility of gene sets from diverse sources, we introduced pathways, annotated gene lists, and gene signatures (PAGs) enriched with metadata to represent biological functions. Furthermore, we established PAG-PAG networks by leveraging gene member similarity and gene regulations. However, in practice, high similarity in descriptions and gene membership often produces redundant, lengthy PAG lists, leading to gene set enrichment results that are difficult to interpret. We present Graph-Oriented Learning with Domain-Embedding Network (GOLDEN) fusion, an integrative framework that jointly leverages (i) connection-based embeddings derived from PAG-PAG relationships and (ii) semantic-based embeddings learned from PAG descriptions with a large language model (LLM). The two representations are combined via early fusion with a tunable weighting to produce a unified embedding on which clustering identifies concise, higher-level super-PAG. To assess when clustering is appropriate, we introduce a Connection Disparity Index, trained on synthetic stochastic block models, as a proxy for network "clusterability." We further optimize the number of clusters with consensus clustering. On Gene Ontology Annotation biological process benchmarks, GOLDEN fusion recovers is-a structure more accurately than either connection-only or semantic-only baselines, demonstrating consistent gains in Adjusted Rand Index and Normalized Mutual Information. Finally, we generate summaries for each super-PAG by synthesizing its member PAG descriptions using a comparative analysis of LLMs. GOLDEN fusion provides an integrated framework for interpreting omics results.

Overview

CoChat Summary

TL;DR

GOLDEN fusion combines graph embeddings from PAG–PAG networks with LLM-derived semantic embeddings to generate higher-level super-PAGs, using a Connection Disparity Index (CDI) to assess clusterability. On Gene Ontology Annotation benchmarks it outperforms graph-only and semantic-only baselines and provides LLM-generated summaries for super-PAGs.

Key findings

  • GOLDEN fusion (weighted graph + semantic fusion) achieved higher ARI than either single-source baseline on GOA SPAGs 5–10 and SPAGs 10–100; the best configurations favored the semantic channel at α ≈ 0.8, and 256-d Node2Vec + 512-d OpenAI embeddings were recommended for the accuracy–efficiency trade-off.
  • Graph-only clustering was best with Agglomerative Clustering (mean ARI 0.51 for SPAGs 5–10; 0.29 for SPAGs 10–100), while semantic-only clustering was also best with Agglomerative Clustering (mean ARI 0.66 and 0.55, respectively); HDBSCAN was lowest (0.01–0.12). GOLDEN fusion outperformed REVIGO, GOSemSim, ReCiPa, and clusterProfiler across both benchmarks.
  • The CDI predictive model classified synthetic networks into unclusterable and clusterable categories with accuracy 0.88 and F1 score 0.9; random forest achieved F1 = 0.90. In SPAG 5–10 pairs, weak functional relevance accounted for 40.0% of CDI-high pairs versus 9.1% of CDI-low pairs, while moderate relevance was 54.5% versus 30.0%.
  • GPT-4.1 gave the highest cosine similarity for PAG summarization: 0.71 ± 0.10 on SPAGs 5–10 and 0.72 ± 0.13 on SPAGs 10–100. In a CML TKI-resistant versus sensitive case study, GOLDEN fusion stabilized at k = 2 with CDI = 0.997.

Methods

  • Inputs are PAG IDs or gene lists processed through PAGER 3.0; the method builds an m-type PAG–PAG graph weighted by hypergeometric gene-overlap significance, learns Node2Vec graph embeddings, generates semantic embeddings from PAG descriptions using OpenAI text-embedding-3-small, aligns them by PCA, L2-normalizes, and fuses them with a tunable weight α. Clustering is performed on the fused representation, with consensus clustering to choose k, followed by LLM summarization of each super-PAG.
  • Benchmarks include the GOA biological process m-type PAG–PAG network with GOA is-a relations as ground truth: SPAGs 5–10 (300 two-member groups) and SPAGs 10–100 (127 two-member groups). KEGG and Reactome pathway datasets were also used with repeated random sampling (300 trials, 2 categories, up to 20 pathways per category). Synthetic SBMs and DC-SBMs were used to train CDI models.
  • Performance was evaluated mainly with ARI and NMI against ground-truth labels; CDI was trained on clustering coefficient distribution features using classifiers including logistic regression, SVM, KNN, random forest, and XGBoost for CDI-β.

Limitations

  • Stated: LLM-generated summaries lack human curation and are intended to support interpretation/navigation, not replace curated ontology definitions or expert annotation; super-PAGs are currently flattened one-level structures; higher edge thresholds can improve pattern representation but reduce network coverage; integrating multi-modality relationships increases modeling complexity; and beyond CML, additional experimental or literature-based validation of identified super-PAGs is needed.
  • Evident from the provided text: several method and result sections are empty or fragmented, so details of CDI-based separation and PAG summarization are incomplete; benchmark ground truth relies on GOA is-a and pathway hierarchies rather than independent experimental validation.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1093/bib/bbag244

Cites

95 of 95 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 The framework of GOLDEN fusion. Input: The input can be either a PAG list identified by PAG IDs or a gene list identified by gene Entrez IDs. The system queries the PAGER database for m-type PAG–PAG relationship info and PAG description info. Graph-based embeddings: Converts a graph (m-type PAG–PAG relationship) into PAG embeddings, applies clustering methods to generate super-PAG, and uses a clusterability prediction model to evaluate whether the graph should be subdivided into meaningful subgraphs. Density-based embeddings: Encodes descriptions with an LLM encoder to create embeddings, which are then clustered. Output: Produces a CDI score, super-PAG, and summary descriptions for each super-PAG.
Figure 1 The framework of GOLDEN fusion. Input: The input can be either a PAG list identified by PAG IDs or a gene list identified by gene Entrez IDs. The system queries the PAGER database for m-type PAG–PAG relationship info and PAG description info. Graph-based embeddings: Converts a graph (m-type PAG–PAG relationship) into PAG embeddings, applies clustering methods to generate super-PAG, and uses a clusterability prediction model to evaluate whether the graph should be subdivided into meaningful subgraphs. Density-based embeddings: Encodes descriptions with an LLM encoder to create embeddings, which are then clustered. Output: Produces a CDI score, super-PAG, and summary descriptions for each super-PAG.CC BY · PubMed Central
Figure 2 Heatmap of ARI comparing clusterability across three settings. The node composition of two subgraphs is presented with proportions of (a) 0.5 and 0.5, (b) 0.7 and 0.3, and (c) 0.9 and 0.1 generated by an SBM. Higher ARI values indicate more accurate recovery of ground-truth clusters. (d) The two cases illustrate the distribution of clustering coefficients, comparing a clusterable scenario (Case 1) with an unclusterable scenario (Case 2) using Spectral Clustering. An ARI of 0.7 is used as the cutoff for determining clusterability to train the classification models.
Figure 2 Heatmap of ARI comparing clusterability across three settings. The node composition of two subgraphs is presented with proportions of (a) 0.5 and 0.5, (b) 0.7 and 0.3, and (c) 0.9 and 0.1 generated by an SBM. Higher ARI values indicate more accurate recovery of ground-truth clusters. (d) The two cases illustrate the distribution of clustering coefficients, comparing a clusterable scenario (Case 1) with an unclusterable scenario (Case 2) using Spectral Clustering. An ARI of 0.7 is used as the cutoff for determining clusterability to train the classification models.CC BY · PubMed Central
Figure 3 Recovery performance of parent GOA terms measured by ARI, comparing three clustering algorithms—HDBSCAN, agglomerative, and k-means—using graph-based embeddings on (a) SPAGs 5–10 and (b) SPAGs 10–100. Numeric labels indicate mean values.
Figure 3 Recovery performance of parent GOA terms measured by ARI, comparing three clustering algorithms—HDBSCAN, agglomerative, and k-means—using graph-based embeddings on (a) SPAGs 5–10 and (b) SPAGs 10–100. Numeric labels indicate mean values.CC BY · PubMed Central
Figure 4 Recovery performance of parent GOA terms measured by ARI, comparing three clustering algorithms—HDBSCAN, agglomerative, and k-means—using semantic-based embeddings on (a) SPAGs 5–10 and (b) SPAGs 10–100. Numeric labels indicate mean values.
Figure 4 Recovery performance of parent GOA terms measured by ARI, comparing three clustering algorithms—HDBSCAN, agglomerative, and k-means—using semantic-based embeddings on (a) SPAGs 5–10 and (b) SPAGs 10–100. Numeric labels indicate mean values.CC BY · PubMed Central
Figure 5 Coefficient optimization of parent GOA term recovery, measured by ARI, comparing three clustering algorithms—HDBSCAN, agglomerative, and k-means—using embedding fusion on (a) SPAGs 5–10 and (b) SPAGs 10–100. Numeric labels indicate mean values.
Figure 5 Coefficient optimization of parent GOA term recovery, measured by ARI, comparing three clustering algorithms—HDBSCAN, agglomerative, and k-means—using embedding fusion on (a) SPAGs 5–10 and (b) SPAGs 10–100. Numeric labels indicate mean values.CC BY · PubMed Central
Figure 6 Visualization of the two super-PAG generated by GOLDEN fusion, with colored dots representing two distinct clusters. The GOs are enriched from a list of differentially expressed genes (DEGs) in a case study comparing CML TKI resistant versus sensitive samples.
Figure 6 Visualization of the two super-PAG generated by GOLDEN fusion, with colored dots representing two distinct clusters. The GOs are enriched from a list of differentially expressed genes (DEGs) in a case study comparing CML TKI resistant versus sensitive samples.CC BY · PubMed Central

Verification

Claims the paper makes

3 Valid · 0 Not valid · 1 Not checked · 74 Own results and interpretations

40 more claims

Create a free account to see all 78 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 26
  • Probably not a substantial claim. 13
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 81 references

Research works printed but not linked: 3

  1. Zhang, 2013, Breast cancer subtyping from plasma proteins, BMC Med Genomics, 2013, 1
  2. Lin, 2004, ROUGE: A Package for Automatic Evaluation of Summaries
  3. OpenAi, 2026, ChatGPT (April Version) [Large Language AI Model]

Metadata

Authors
Li, Q., Nichols, C., Welner, R. S., Chen, J. Y., Ku, W.-S., & Yue, Z.
Affiliations
  • Fisk University (US)
  • Auburn University (US)
  • University of Alabama at Birmingham (US)
Corresponding author
  • Chen, Jake Y
  • Ku, Wei-Shinn
  • Yue, Zongliang
Venue
Briefings in Bioinformatics
Published
2026-05
Volume
27
Issue
3
ISSN
  • 1467-5463 (ISSN-L)
  • 1477-4054
OpenAlex
W7162300909
Cited by
0
References
81
FWCI
0.00
Type
journal article
Language
en
Open access
Locations
  • Briefings in Bioinformatics (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
  • University of Alabama at Birmingham Center for Clinical and Translational Science Pilot Award (UM1TR004771)
  • Gulf Research Program of the National Academies of Sciences (SCON-10001538)
  • University of Alabama at Birmingham (UC267384)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Bioinformatics and Genomic Networks
  • Health Sciences › Medicine › Health Informatics › Biomedical Text Mining and Ontologies
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Genomics and Phylogenetic Studies
Keywords
gene expression signatures, graph embedding, domain embedding, large language models, semantic embeddings, feature-level fusion, consensus clustering, Gene Ontology biological process, omics data interpretation, functional genomics
MeSH terms
Large Language Models, Humans, Genomics, Gene Regulatory Networks

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: GOLDEN fusion: a graph-oriented learning with domain-embedding network fusion for generating super gene sets in functional genomics. https://app.cochat.ai/papers/10.1093/bib/bbag244
The paper
Li, Q., Nichols, C., Welner, R. S., Chen, J. Y., Ku, W.-S., & Yue, Z. (2026). GOLDEN fusion: a graph-oriented learning with domain-embedding network fusion for generating super gene sets in functional genomics. Briefings in Bioinformatics. https://doi.org/10.1093/bib/bbag244

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.