# Optimized Dynamic Network Biomarker Deciphers a High-Resolution Heterogeneity Within Thyroid Cancer Molecular Subtypes
Researcher: Ge Zhang
Chinese name: 张格
ORCID identity anchor: https://orcid.org/0000-0002-3116-3246
Canonical researcher: https://drgezhang.com/#person
Journal: Med Research
Year: 2025
Type: Article
DOI: 10.1002/mdr2.70004
Canonical page: https://drgezhang.com/papers/doi-10-1002-mdr2-70004.html
Markdown record: https://drgezhang.com/papers/doi-10-1002-mdr2-70004.md
## Full Authors
1. Haonan Zhang
2. Ge Zhang
3. Pengyuan Xu
4. Fengyi Yu
5. Liwen Li
6. Runzhi Huang
7. Pengpeng Zhang
8. Kaisaierjiang Kadier
9. Yuanyuan Wang
10. Qiqi Gu
11. Yalei Ding
12. Tianshu Gu
13. Hao Chi
14. Shiqian Zhang
15. Ruhao Wu
16. Yudi Xu
17. Shanshan Zhu
18. Honglin Zheng
19. Taiqi Zhao
20. Qi He
21. Xinguang Qiu
## Official Abstract
The progression of differentiated thyroid carcinoma (DTC) poses significant clinical challenges, especially in determining the optimal time for intervention. To capture early signals of disease progression, we employed an optimized dynamic network biomarker (DNB) method—a systems biology approach that detects abrupt molecular changes indicating a critical transition signal. This analysis revealed that Stage II marks a critical transition in the disease trajectory. We further developed a scoring system called TCPSLevel (Thyroid Carcinoma Progression Signature Level), which quantifies individual progression risk based on gene expression profiles. TCPSLevel showed strong associations with clinical features and prognosis across multiple datasets. Our ensemble consensus clustering approach uncovered three robust DTC molecular subtypes, which demonstrated distinct clinical outcomes, immune microenvironments, regulatory landscapes, and therapeutic agents. A clinically applicable classifier (miniPC) was constructed using machine learning to facilitate subtype prediction. We also identified ASPH as a key regulator driving progression and validated its expression and function experimentally. Together, these findings offer new insights and practical tools for early risk assessment and personalized management of thyroid cancer.
Text reproduced verbatim from Version of Record ([source](https://onlinelibrary.wiley.com/doi/full/10.1002/mdr2.70004)); [DOI](https://doi.org/10.1002/mdr2.70004); [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/).
## Research & review context
### Research fields
- Cancer systems biology
- Thyroid cancer molecular research
### Review topics
- Molecular heterogeneity and cancer subtyping
- Expression-network analysis of disease progression
### Scientific role
Stage-associated network and molecular-subtype analysis with external classifier evaluation and separate ASPH tissue and cell follow-up.
### Evidence position
Public transcriptomics, human tissue observations and cell experiments; not physiological circadian measurement, individual transition-time prediction or clinical drug efficacy.
### Related concepts
- Network-based cancer subtyping
- Classifier versus outcome evidence
### Possible citation contexts
- Methods: expression-network approaches to thyroid-cancer heterogeneity.
- Evidence synthesis: computational subtype characterization alongside candidate experimental follow-up, without treating a subtype classifier as a survival model.
Possible literature-review contexts, not evidence of existing citations. Read the findings and Evidence Scope before citing the original article.
## Evidence Scale
**Optimized Dynamic Network Biomarker Deciphers a High-Resolution Heterogeneity Within Thyroid Cancer Molecular Subtypes**
Stage-associated DNB analysis, TCPSLevel, a compact subtype classifier and ASPH follow-up are presented as distinct evidence layers.
| Evidence | Value |
| --- | --- |
| GEO microarray cohorts | 21 cohorts, plus TCGA |
| TCGA count in Methods | 544 |
| TCGA count in GRN analysis | 484 |
| scRNA cells in Methods | 17,709 |
| scRNA cells in Results | 17,518 |
| Reported single-cell donors | 1 |
| Human tissue specimens | 48 tumor and 48 adjacent specimens |
### Counting note
TCGA counts of 544 in Methods and 484 in GRN analysis describe different reported analytical stages; no reason for excluding the numerical difference is inferred. Methods reports 17,709 cells and Results 17,518, without a reconstructed QC-exclusion process. One reported donor is not a large participant cohort; 48 tumor and 48 adjacent specimens do not mean 96 independent patients. Cohorts, cells, paired specimens and cultured-cell experiments are not added into a unique-participant total. External validation means retrospective expression-template/classifier evaluation, not prospective clinical utility.
## Research Question
Can stage-associated transcriptional networks support thyroid-cancer subtyping and nominate experimentally assessable progression-related candidates?
## Author Evidence Summary
The study connects stage-associated computational analysis with subtype classification and experimental candidate assessment. A stage-ordered network signal is not direct observation of an individual patient's future transition. Human tissue associations, cell perturbation and computational therapeutic hypotheses are presented separately.
## Key Findings
### KF1: Stage II had the highest reported network instability signal.
Context: Stage-associated cross-sectional network analysis.
| Evidence | Value |
| --- | --- |
| DNB genes | 36 |
Source locator: Methods 2.2; Results 3.1; Fig. 1
### KF2: TCPSLevel was associated with clinical and outcome differences.
Context: Expression-score construction and retrospective association.
| Evidence | Value |
| --- | --- |
| Expression bins | 100 |
| Background samplings | 1,000 |
| Reported survival cutoff | 0.4 |
Source locator: Methods 2.4; Results 3.2; Fig. 2
### KF3: The ensemble analysis yielded three molecular subtypes.
Context: Consensus framework selection and external NTP evaluation.
| Evidence | Value |
| --- | --- |
| Candidate combinations | 20 |
| Selected combination | CV-PAM |
| Reported external NTP datasets | 6 |
Source locator: Methods 2.6-2.7; Results 3.3; Fig. 3
### KF4: miniPC classified the consensus subtype labels using a compact gene set.
Context: Gene selection and external subtype-classifier evaluation.
| Evidence | Value |
| --- | --- |
| Candidate genes | 296 |
| Final genes | 12 |
| Final classifier | Neural network |
| External classifier cohorts | 5 |
Source locator: Methods 2.8; Results 3.4; Fig. 4
### KF5: Single-cell analysis localized expression patterns across cellular compartments.
Context: Single-donor single-cell analysis with separately reported Methods and Results counts.
| Evidence | Value |
| --- | --- |
| Reported donors | 1 |
| Cells in Methods | 17,709 |
| Cells in Results | 17,518 |
| Annotated populations | 7 |
Source locator: Methods 2.1.1 and 2.3; Results 3.7; Fig. 5K-N
### KF6: ASPH was evaluated in human tissues and through TPC-1 silencing.
Context: Human tissue association and separate cultured-cell perturbation.
| Evidence | Value |
| --- | --- |
| Tumor specimens | 48 |
| Adjacent specimens | 48 |
| TPC-1 silencing | Reduced proliferation and migration |
Source locator: Methods: tissue and cell experiments; Results 3.6 and 3.8; Fig. 6
### KF7: IPS and CMap supplied candidate immune/drug-response hypotheses.
Context: Computational treatment-related hypothesis generation.
| Evidence | Value |
| --- | --- |
| Evidence | Computational scoring and signature matching |
Source locator: Methods 2.10 and 2.20; Results 3.5 and 3.9; Figure 7G-I; Discussion
## Study Design & Analytical Framework
### Study profile
- Study Design: Public bulk/single-cell analysis with human tissue assessment and cell experiments
- Evidence Type: Retrospective molecular subtyping with experimental candidate follow-up
- Population: Thyroid-cancer datasets and tissue specimens, with separate cultured-cell experiments
- Primary Endpoint: DNB-associated expression structure and molecular subtype assignment
- Secondary Endpoint: Retrospective outcome associations and ASPH-related experimental readouts
- External dataset evaluation: Yes
- Data modalities: bulk transcriptomics, single-cell RNA sequencing, network/classifier analysis, human immunohistochemistry, in-vitro gene perturbation
## What This Study Adds
- Connects stage-associated network analysis with molecular subtyping.
- Distinguishes subtype discovery from subtype-classifier evaluation.
- Separates human tissue, cultured-cell and computational drug-hypothesis evidence.
## Evidence Scope
### Supports
- Stage II had the highest reported network instability signal.
- TCPSLevel was associated with clinical and outcome differences.
- The ensemble analysis yielded three molecular subtypes.
- miniPC classified the consensus subtype labels using a compact gene set.
- Single-cell analysis localized expression patterns across cellular compartments.
- ASPH was evaluated in human tissues and through TPC-1 silencing.
- IPS and CMap supplied candidate immune/drug-response hypotheses.
### Does Not Establish
- Longitudinal prediction of an individual patient's transition time.
- Clinical intervention benefit.
- miniPC as an established survival model rather than a subtype classifier.
- Cell counts as patient counts or paired-specimen counts as independent participants.
- Complete demonstration of direct ASCL2 regulation or binding.
- Clinical efficacy of CMap-nominated compounds.
- Classification of this study as a physiological circadian-rhythm experiment.
- Resolution of all Methods/Results counts or accession differences.
### Limitations
- TCGA counts of 544 in Methods and 484 in GRN analysis describe different reported analytical stages; no reason for excluding the numerical difference is inferred. Methods reports 17,709 cells and Results 17,518, without a reconstructed QC-exclusion process. One reported donor is not a large participant cohort; 48 tumor and 48 adjacent specimens do not mean 96 independent patients. Cohorts, cells, paired specimens and cultured-cell experiments are not added into a unique-participant total. External validation means retrospective expression-template/classifier evaluation, not prospective clinical utility.
- Cross-sectional stage comparisons do not establish longitudinal prediction of each patient's future deterioration time.
- The reported cutoff is not a cross-cohort locked clinical threshold. Grouping rules used in other analyses are not automatically equivalent.
- Methods 2.6 prints GSE333630 whereas other locations use GSE33630. These are not silently rewritten as a fully verified accession inventory.
- This is subtype classification, not prospective individual survival-benefit validation. Unverified figure-only accuracy or AUC values are not reconstructed.
- A single donor and a large cell count do not establish large-population replication. No explanation for the count difference is invented.
- Paired tissues are not 96 independent patients. ASCL2-ASPH network inference is not direct binding validation, and cell effects do not establish human targeted-treatment efficacy.
- Predicted compounds are not proven effective treatments. IPS is not an observed outcome in patients receiving immunotherapy.
- Longitudinal prediction of an individual patient's transition time.
- Clinical intervention benefit.
- miniPC as an established survival model rather than a subtype classifier.
- Cell counts as patient counts or paired-specimen counts as independent participants.
- Complete demonstration of direct ASCL2 regulation or binding.
- Clinical efficacy of CMap-nominated compounds.
- Classification of this study as a physiological circadian-rhythm experiment.
- Resolution of all Methods/Results counts or accession differences.
## When This Study Is Useful to Cite
### CU1: Which thyroid studies applied dynamic network biomarkers to stage-associated expression?
Stage II had the highest reported network instability signal. DNB genes: 36. Cross-sectional stage comparisons do not establish longitudinal prediction of each patient's future deterioration time. TCPSLevel was associated with clinical and outcome differences. Expression bins: 100; Background samplings: 1,000; Reported survival cutoff: 0.4. The reported cutoff is not a cross-cohort locked clinical threshold. Grouping rules used in other analyses are not automatically equivalent.
Evidence: [KF1](#kf1), [KF2](#kf2)
### CU2: What evidence supports ensemble thyroid-cancer molecular subtyping?
The ensemble analysis yielded three molecular subtypes. Candidate combinations: 20; Selected combination: CV-PAM; Reported external NTP datasets: 6. Methods 2.6 prints GSE333630 whereas other locations use GSE33630. These are not silently rewritten as a fully verified accession inventory.
Evidence: [KF3](#kf3)
### CU3: How was a compact classifier developed for consensus thyroid subtypes?
miniPC classified the consensus subtype labels using a compact gene set. Candidate genes: 296; Final genes: 12; Final classifier: Neural network; External classifier cohorts: 5. This is subtype classification, not prospective individual survival-benefit validation. Unverified figure-only accuracy or AUC values are not reconstructed.
Evidence: [KF4](#kf4)
### CU4: What evidence distinguishes ASPH expression association from experimental perturbation?
Single-cell analysis localized expression patterns across cellular compartments. Reported donors: 1; Cells in Methods: 17,709; Cells in Results: 17,518; Annotated populations: 7. A single donor and a large cell count do not establish large-population replication. No explanation for the count difference is invented. ASPH was evaluated in human tissues and through TPC-1 silencing. Tumor specimens: 48; Adjacent specimens: 48; TPC-1 silencing: Reduced proliferation and migration. Paired tissues are not 96 independent patients. ASCL2-ASPH network inference is not direct binding validation, and cell effects do not establish human targeted-treatment efficacy.
Evidence: [KF5](#kf5), [KF6](#kf6)
### CU5: What computational rather than clinical therapeutic hypotheses were generated?
IPS and CMap supplied candidate immune/drug-response hypotheses. Evidence: Computational scoring and signature matching. Predicted compounds are not proven effective treatments. IPS is not an observed outcome in patients receiving immunotherapy.
Evidence: [KF7](#kf7)
## What This Study Should Not Be Cited to Claim
- Longitudinal prediction of an individual patient's transition time.
- Clinical intervention benefit.
- miniPC as an established survival model rather than a subtype classifier.
- Cell counts as patient counts or paired-specimen counts as independent participants.
- Complete demonstration of direct ASCL2 regulation or binding.
- Clinical efficacy of CMap-nominated compounds.
- Classification of this study as a physiological circadian-rhythm experiment.
- Resolution of all Methods/Results counts or accession differences.
## Evidence Matrix
### Stage-associated DNB inference
- Context: Stage-associated cross-sectional network analysis.
- Finding: Stage II had the highest reported network instability signal. DNB genes: 36
- Evidence Level: Retrospective expression and computational evidence
- Scope: Cross-sectional stage comparisons do not establish longitudinal prediction of each patient's future deterioration time.
- Source Locator: Methods 2.2; Results 3.1; Fig. 1
- Evidence refs: [KF1](#kf1)
### TCPSLevel construction and association
- Context: Expression-score construction and retrospective association.
- Finding: TCPSLevel was associated with clinical and outcome differences. Expression bins: 100; Background samplings: 1,000; Reported survival cutoff: 0.4
- Evidence Level: Retrospective expression and computational evidence
- Scope: The reported cutoff is not a cross-cohort locked clinical threshold. Grouping rules used in other analyses are not automatically equivalent.
- Source Locator: Methods 2.4; Results 3.2; Fig. 2
- Evidence refs: [KF2](#kf2)
### Consensus subtype framework
- Context: Consensus framework selection and external NTP evaluation.
- Finding: The ensemble analysis yielded three molecular subtypes. Candidate combinations: 20; Selected combination: CV-PAM; Reported external NTP datasets: 6
- Evidence Level: Retrospective expression and computational evidence
- Scope: Methods 2.6 prints GSE333630 whereas other locations use GSE33630. These are not silently rewritten as a fully verified accession inventory.
- Source Locator: Methods 2.6-2.7; Results 3.3; Fig. 3
- Evidence refs: [KF3](#kf3)
### miniPC subtype classifier
- Context: Gene selection and external subtype-classifier evaluation.
- Finding: miniPC classified the consensus subtype labels using a compact gene set. Candidate genes: 296; Final genes: 12; Final classifier: Neural network; External classifier cohorts: 5
- Evidence Level: Retrospective expression and computational evidence
- Scope: This is subtype classification, not prospective individual survival-benefit validation. Unverified figure-only accuracy or AUC values are not reconstructed.
- Source Locator: Methods 2.8; Results 3.4; Fig. 4
- Evidence refs: [KF4](#kf4)
### Single-cell context
- Context: Single-donor single-cell analysis with separately reported Methods and Results counts.
- Finding: Single-cell analysis localized expression patterns across cellular compartments. Reported donors: 1; Cells in Methods: 17,709; Cells in Results: 17,518; Annotated populations: 7
- Evidence Level: Retrospective expression and computational evidence
- Scope: A single donor and a large cell count do not establish large-population replication. No explanation for the count difference is invented.
- Source Locator: Methods 2.1.1 and 2.3; Results 3.7; Fig. 5K-N
- Evidence refs: [KF5](#kf5)
### ASPH experimental follow-up
- Context: Human tissue association and separate cultured-cell perturbation.
- Finding: ASPH was evaluated in human tissues and through TPC-1 silencing. Tumor specimens: 48; Adjacent specimens: 48; TPC-1 silencing: Reduced proliferation and migration
- Evidence Level: Human tissue association and in-vitro perturbation
- Scope: Paired tissues are not 96 independent patients. ASCL2-ASPH network inference is not direct binding validation, and cell effects do not establish human targeted-treatment efficacy.
- Source Locator: Methods: tissue and cell experiments; Results 3.6 and 3.8; Fig. 6
- Evidence refs: [KF6](#kf6)
### Computational therapeutic hypotheses
- Context: Computational treatment-related hypothesis generation.
- Finding: IPS and CMap supplied candidate immune/drug-response hypotheses. Evidence: Computational scoring and signature matching
- Evidence Level: Retrospective expression and computational evidence
- Scope: Predicted compounds are not proven effective treatments. IPS is not an observed outcome in patients receiving immunotherapy.
- Source Locator: Methods 2.10 and 2.20; Results 3.5 and 3.9; Figure 7G-I; Discussion
- Evidence refs: [KF7](#kf7)
## Q&A
### What does the Stage II DNB signal establish?
Stage II had the highest reported network instability signal. DNB genes: 36. Cross-sectional stage comparisons do not establish longitudinal prediction of each patient's future deterioration time.
Evidence: KF1
### Is TCPSLevel already a clinically validated prospective risk tool?
TCPSLevel was associated with clinical and outcome differences. Expression bins: 100; Background samplings: 1,000; Reported survival cutoff: 0.4. The reported cutoff is not a cross-cohort locked clinical threshold. Grouping rules used in other analyses are not automatically equivalent.
Evidence: KF2
### How were the thyroid molecular subtypes evaluated externally?
The ensemble analysis yielded three molecular subtypes. Candidate combinations: 20; Selected combination: CV-PAM; Reported external NTP datasets: 6. Methods 2.6 prints GSE333630 whereas other locations use GSE33630. These are not silently rewritten as a fully verified accession inventory.
Evidence: KF3
### Does miniPC classify subtypes or directly prove improved survival?
miniPC classified the consensus subtype labels using a compact gene set. Candidate genes: 296; Final genes: 12; Final classifier: Neural network; External classifier cohorts: 5. This is subtype classification, not prospective individual survival-benefit validation. Unverified figure-only accuracy or AUC values are not reconstructed.
Evidence: KF4
### How many independent donors support the single-cell analysis?
Single-cell analysis localized expression patterns across cellular compartments. Reported donors: 1; Cells in Methods: 17,709; Cells in Results: 17,518; Annotated populations: 7. A single donor and a large cell count do not establish large-population replication. No explanation for the count difference is invented.
Evidence: KF5
### What human and experimental evidence supports ASPH?
ASPH was evaluated in human tissues and through TPC-1 silencing. Tumor specimens: 48; Adjacent specimens: 48; TPC-1 silencing: Reduced proliferation and migration. Paired tissues are not 96 independent patients. ASCL2-ASPH network inference is not direct binding validation, and cell effects do not establish human targeted-treatment efficacy.
Evidence: KF6
### Were CMap-nominated compounds shown to benefit patients?
IPS and CMap supplied candidate immune/drug-response hypotheses. Evidence: Computational scoring and signature matching. Predicted compounds are not proven effective treatments. IPS is not an observed outcome in patients receiving immunotherapy.
Evidence: KF7
## Concepts & Entities
### Disease
differentiated thyroid carcinoma
### Entities
DNB, TCPSLevel, miniPC, ASPH, ASCL2
### Methods
network analysis, consensus clustering, NTP, subtype classification
### Evidence
tissue association, in-vitro perturbation
## Related Research
- [Smooth muscle cell fate decisions decipher a high-resolution heterogeneity within atherosclerosis molecular subtypes](https://drgezhang.com/papers/smc-fate.html) — Expression-state and subtyping methods in a different disease, not the same mechanism or participant population. (DOI: 10.1186/s12967-022-03795-9)
- [Machine learning combined with omics-based approaches reveals T-lymphocyte cellular fate imbalance in abdominal aortic aneurysm](https://drgezhang.com/papers/doi-10-1186-s12915-025-02400-x.html) — Single-cell/transcriptomic stratification in another study; this is not independent replication of an ASPH mechanism. (DOI: 10.1186/s12915-025-02400-x)
## Publication & Provenance
- DOI: https://doi.org/10.1002/mdr2.70004
- Doi Url: https://doi.org/10.1002/mdr2.70004
- Publisher: https://onlinelibrary.wiley.com/doi/full/10.1002/mdr2.70004
- Review Basis: Evidence synthesis from the cited main text, figure legends and designated primary records. No independent reanalysis of participant-level data. Unavailable supplementary or figure-only details are not reconstructed.
## Evidence-page notice
Author-controlled evidence synthesis, distinct from the published abstract and not a replacement for the version of record. Source-specific counting and interpretation limits are retained.
## Links
- DOI: https://doi.org/10.1002/mdr2.70004
- HTML page: https://drgezhang.com/papers/doi-10-1002-mdr2-70004.html
- Google Scholar query: https://scholar.google.com/scholar?q=%22Optimized%20Dynamic%20Network%20Biomarker%20Deciphers%20a%20High-Resolution%20Heterogeneity%20Within%20Thyroid%20Cancer%20Molecular%20Subtypes%22
- ORCID: https://orcid.org/0000-0002-3116-3246