# Advancing predictive markers in lung adenocarcinoma: A machine learning-based immunotherapy prognostic prediction signature

Researcher: Ge Zhang
Chinese name: 张格
ORCID identity anchor: https://orcid.org/0000-0002-3116-3246
Canonical researcher: https://drgezhang.com/#person

Journal: Environmental Toxicology
Year: 2024
Type: Article
DOI: 10.1002/tox.24284
Canonical page: https://drgezhang.com/papers/doi-10-1002-tox-24284.html
Markdown record: https://drgezhang.com/papers/doi-10-1002-tox-24284.md

## Full Authors

1. Zhongyan Li
2. Shengbin Pei
3. Yanjuan Wang
4. Ge Zhang
5. Haoran Lin
6. Shiyang Dong

<a id="paper-discovery"></a>
## Research & review context

### Research fields

- Computational oncology
- LUAD prognostic modeling

### Review topics

- Immune-associated transcriptomic signatures
- Validation leakage in molecular survival models

### Scientific role

A retrospective 13-gene model-development study with outcome-based screening, multi-algorithm selection and computational drug hypotheses.

### Evidence position

Development and retrospective association evidence; no untouched external or prospective immunotherapy-benefit validation.

### Related concepts

- IPPS
- Outcome-based selection
- Computational drug sensitivity

### Possible citation contexts

- Methods reviews describing the Lasso plus SuperPC model selected from 101 algorithm combinations.
- Critical appraisal of cohort reuse, outcome-based immunotherapy sample selection and the distinction between prognostic AUCs and treatment-response evidence.

Possible literature-review contexts, not evidence of existing citations. Read the findings and Evidence Scope before citing the original article.

## Evidence Scale

**IPPS: retrospective LUAD prognostic modeling and its validation limits**

A 13-gene LUAD prognostic model was selected from 101 algorithm combinations using public transcriptomic cohorts. Gene screening and model selection reused evaluation cohorts; survival-defined groups and unresolved reporting discrepancies prevent treating this study as independent immunotherapy-response validation.

| Evidence | Value |
| --- | --- |
| final signature genes | 13 |
| candidate model combinations | 101 |
| GEO cohorts used in screening and model selection | 5 |

### Counting note

Methods 2.1 reports 577 LUAD cases from OAK/POPLAR, whereas Results 3.1 and Figure 1D describe a selected group of 296. The retrieved article does not fully reconcile those denominators; no pooled participant count is inferred. TCGA and five GEO cohorts entered survival-based gene screening, and the five GEO cohorts also determined model selection. They are distinct datasets, not independent held-out validation after model selection.

## Research Question

How was an immunotherapy-associated gene signature developed for retrospective LUAD survival stratification, and what limits the interpretation of its reported performance?

## Author Evidence Summary

The study used OAK/POPLAR data for an initial expression contrast and screened candidate genes against survival in TCGA-LUAD and five GEO cohorts. It constructed 101 model combinations in TCGA and selected Lasso plus SuperPC using the mean C-index across the five GEO cohorts. The final IPPS includes 13 genes. The TCGA time-dependent AUCs reported consistently in the text and Figure 2H are 0.80, 0.79 and 0.78 at 1, 3 and 5 years; these are training-cohort estimates. The GEO datasets contributed to gene screening and model selection, so they are not untouched independent tests. The initial groups were defined using overall-survival thresholds despite being labelled CR/PR and PD/SD. Drug-sensitivity results are computational hypotheses. Differences in the reported OAK/POPLAR denominators and in GSE31210 AUCs remain unresolved.

## Key Findings

<a id="kf1"></a>
### KF1: The reported IPPS contains 13 genes and uses Lasso plus SuperPC selected from 101 candidate combinations.

Context: Retrospective development in TCGA-LUAD. Candidate-gene survival screening included TCGA and five GEO cohorts; selection of the final algorithm used the mean C-index in those same five GEO cohorts.

| Evidence | Value |
| --- | --- |
| signature genes | 13 |
| candidate combinations | 101 |
| selected model | Lasso + SuperPC |

Source locator: Methods 2.2, author-uploaded publisher-layout pp. 2-3; Results 3.1-3.2, pp. 3-5; Figure 1D-E; Figure 2A.

<a id="kf2"></a>
### KF2: The article reports TCGA training-cohort AUCs of 0.80, 0.79 and 0.78 for 1-, 3- and 5-year survival.

Context: These are apparent performance estimates in the TCGA development cohort, not independent external or prospective validation. The GSE31210 text-versus-figure discrepancy is excluded from performance claims.

| Evidence | Value |
| --- | --- |
| 1-year TCGA AUC | 0.80 |
| 3-year TCGA AUC | 0.79 |
| 5-year TCGA AUC | 0.78 |

Source locator: Results 3.2, author-uploaded publisher-layout p. 5; Figure 2H, p. 4. Text and displayed TCGA panel agree.

<a id="kf3"></a>
### KF3: The initial OAK/POPLAR comparison used overall-survival extremes and does not by itself establish an objective immunotherapy-response classifier.

Context: Figure 1D and Results 3.1 define groups by survival greater than 12 months or less than 6 months, while labelling them CR/PR and PD/SD. Those labels must not be treated as verified RECIST response classifications.

| Evidence | Value |
| --- | --- |
| longer-survival definition | overall survival >12 months |
| shorter-survival definition | overall survival <6 months |

Source locator: Results 3.1 and Figure 1D, author-uploaded publisher-layout p. 3; compare Methods 2.1, p. 2.

<a id="kf4"></a>
### KF4: The drug-sensitivity analysis generates computational treatment hypotheses rather than observations of drug benefit in patients.

Context: The Methods names oncoPredict and pRRophetic, and Results 3.7 uses CTRP and PRISM drug-sensitivity resources. No new drug-treatment experiment is established by this analysis.

| Evidence | Value |
| --- | --- |
| inference tools | oncoPredict and pRRophetic |
| drug-response resources | CTRP and PRISM |

Source locator: Methods 2.4, author-uploaded publisher-layout pp. 2-3; Results 3.7, p. 7; Discussion, pp. 10-12. No compound-specific effect estimate is reproduced.

## Study Design & Analytical Framework

### Study profile

- Study Design: Retrospective public-cohort gene screening and survival-model selection with computational immune and drug analyses
- Evidence Type: Retrospective transcriptomic associations and model-development performance
- Population: LUAD records from TCGA, five GEO datasets and an OAK/POPLAR-derived discovery comparison
- Primary Endpoint: Overall survival association and time-dependent discrimination
- Secondary Endpoint: Inferred immune infiltration, genomic associations and computational drug sensitivity
- External dataset evaluation: No
- Data modalities: bulk transcriptomics, clinical survival data, mutation and copy-number data, computational immune and drug-response estimates

## What This Study Adds

- A documented 13-gene prognostic-model development workflow linking an OAK/POPLAR discovery contrast to TCGA and GEO survival analyses.
- A case for separating retrospective prognostic discrimination, outcome-driven model selection and computational treatment hypotheses from independent clinical validation.

## Evidence Scope

### Supports

- Description of the reported 13-gene Lasso-plus-SuperPC development procedure and 101 candidate combinations.
- The text-and-figure-consistent TCGA training AUCs, explicitly labelled as apparent development-cohort performance.
- A source-based distinction between survival-defined discovery groups, prognostic modeling and computational drug-response inference.

### Does Not Establish

- An independently validated predictor of RECIST response or differential benefit from immunotherapy.
- Prospective clinical utility or improved outcomes from IPPS-guided treatment.
- Unbiased external performance after gene and algorithm selection using the same GEO cohorts.
- A reconciled OAK/POPLAR denominator or a resolved GSE31210 AUC estimate.
- Experimental drug efficacy or causal functions of the model genes.

### Limitations

- The study is retrospective. Survival-based gene selection used all six TCGA/GEO cohorts, and five GEO cohorts were reused to choose the algorithm. Their results cannot be treated as untouched external validation.
- Methods 2.1 states 577 OAK/POPLAR LUAD cases, while Results 3.1 and Figure 1D state a selected 296. The retrieved version does not sufficiently specify the full filtering relationship; no reconciled or pooled denominator is asserted.
- Results 3.1 and Figure 1D use OS >12 months and OS <6 months but name the groups CR/PR and PD/SD. Survival strata do not establish objective response categories or differential immunotherapy benefit.
- GSE31210 AUCs differ between Results 3.2 (0.67, 0.76, 0.72 at 1, 3, 5 years) and Figure 2L (0.67, 0.73, 0.75). Neither series is adopted as a verified performance estimate.
- The TCGA AUCs quoted here are apparent training-cohort estimates. Individual participant data and code were not reanalysed, and no uncertainty interval is invented.
- Drug-response analyses are computational. The Discussion uses experimental language, but Methods 2.4 and Results 3.7 describe inference from drug-sensitivity databases.
- The author-uploaded publisher-layout text was readable publicly on ResearchGate, and Figures 1 and 2 were visually checked there. Direct PDF download was blocked in this environment; the archived source fingerprint identifies the acquired web transcription, not PDF bytes. No claim is based on an unseen supplementary result.

<a id="citation-use-cases"></a>
## When This Study Is Useful to Cite

### CU1: How was the LUAD IPPS signature developed?

Thirteen genes and Lasso plus SuperPC were selected in a retrospective workflow examining 101 combinations; five GEO cohorts contributed to model selection.

Evidence: [KF1](#kf1)

### CU2: What apparent TCGA discrimination did IPPS report?

The text and Figure 2H agree on TCGA AUCs of 0.80, 0.79 and 0.78, which are development-cohort estimates.

Evidence: [KF2](#kf2)

### CU3: Can IPPS be cited as independent clinical immunotherapy-response validation?

The paper describes survival-defined discovery groups and computational drug inference; these do not establish independent objective-response or treatment-benefit validation.

Evidence: [KF1](#kf1), [KF3](#kf3), [KF4](#kf4)

<a id="citation-boundaries"></a>
## What This Study Should Not Be Cited to Claim

- Independent external validation, prospectively demonstrated clinical utility or immunotherapy treatment-effect prediction.
- Verified RECIST response categories inferred only from the CR/PR and PD/SD labels used for survival extremes.
- A corrected GSE31210 AUC, a deduplicated patient denominator, or experimentally validated drug efficacy.

<a id="evidence-matrix"></a>
## Evidence Matrix

<a id="evidence-kf1"></a>
### The reported IPPS contains 13 genes and uses Lasso plus SuperPC selected from 101 candidate combinations.

- Context: Retrospective development in TCGA-LUAD. Candidate-gene survival screening included TCGA and five GEO cohorts; selection of the final algorithm used the mean C-index in those same five GEO cohorts.
- Finding: signature genes: 13; candidate combinations: 101; selected model: Lasso + SuperPC
- Evidence Level: Retrospective transcriptomic associations and model-development performance
- Scope: Bounded retrospective development evidence; no independent treatment-response validation.
- Source Locator: Methods 2.2, author-uploaded publisher-layout pp. 2-3; Results 3.1-3.2, pp. 3-5; Figure 1D-E; Figure 2A.
- Evidence refs: [KF1](#kf1)

<a id="evidence-kf2"></a>
### The article reports TCGA training-cohort AUCs of 0.80, 0.79 and 0.78 for 1-, 3- and 5-year survival.

- Context: These are apparent performance estimates in the TCGA development cohort, not independent external or prospective validation. The GSE31210 text-versus-figure discrepancy is excluded from performance claims.
- Finding: 1-year TCGA AUC: 0.80; 3-year TCGA AUC: 0.79; 5-year TCGA AUC: 0.78
- Evidence Level: Retrospective transcriptomic associations and model-development performance
- Scope: Bounded retrospective development evidence; no independent treatment-response validation.
- Source Locator: Results 3.2, author-uploaded publisher-layout p. 5; Figure 2H, p. 4. Text and displayed TCGA panel agree.
- Evidence refs: [KF2](#kf2)

<a id="evidence-kf3"></a>
### The initial OAK/POPLAR comparison used overall-survival extremes and does not by itself establish an objective immunotherapy-response classifier.

- Context: Figure 1D and Results 3.1 define groups by survival greater than 12 months or less than 6 months, while labelling them CR/PR and PD/SD. Those labels must not be treated as verified RECIST response classifications.
- Finding: longer-survival definition: overall survival >12 months; shorter-survival definition: overall survival <6 months
- Evidence Level: Retrospective transcriptomic associations and model-development performance
- Scope: Bounded retrospective development evidence; no independent treatment-response validation.
- Source Locator: Results 3.1 and Figure 1D, author-uploaded publisher-layout p. 3; compare Methods 2.1, p. 2.
- Evidence refs: [KF3](#kf3)

<a id="evidence-kf4"></a>
### The drug-sensitivity analysis generates computational treatment hypotheses rather than observations of drug benefit in patients.

- Context: The Methods names oncoPredict and pRRophetic, and Results 3.7 uses CTRP and PRISM drug-sensitivity resources. No new drug-treatment experiment is established by this analysis.
- Finding: inference tools: oncoPredict and pRRophetic; drug-response resources: CTRP and PRISM
- Evidence Level: Retrospective transcriptomic associations and model-development performance
- Scope: Bounded retrospective development evidence; no independent treatment-response validation.
- Source Locator: Methods 2.4, author-uploaded publisher-layout pp. 2-3; Results 3.7, p. 7; Discussion, pp. 10-12. No compound-specific effect estimate is reproduced.
- Evidence refs: [KF4](#kf4)

## Q&A

### What model did the IPPS study select?

A 13-gene Lasso-plus-SuperPC model was selected from 101 candidate combinations. The five GEO evaluation cohorts also contributed to gene screening and algorithm selection.

Evidence: KF1

### What survival performance can be quoted without resolving the GSE31210 discrepancy?

The text and Figure 2H agree on TCGA AUCs of 0.80, 0.79 and 0.78 at 1, 3 and 5 years. They describe the development cohort and must not be presented as independent external performance.

Evidence: KF2

### Were the five GEO cohorts untouched independent test sets?

No. Results 3.1 uses the six TCGA/GEO cohorts for survival-based gene screening, and Results 3.2 selects the final algorithm using the mean C-index across the five GEO cohorts.

Evidence: KF1

### Does the OAK/POPLAR analysis verify an objective immunotherapy-response classifier?

No. The reported contrast uses overall survival greater than 12 months versus less than 6 months, despite CR/PR and PD/SD labels. This does not establish independently verified objective-response prediction or treatment benefit.

Evidence: KF3

### Were the proposed drugs tested as IPPS-guided treatments?

The reported analysis uses oncoPredict, pRRophetic, CTRP and PRISM to infer drug sensitivity. It does not demonstrate patient benefit from selecting a drug using IPPS.

Evidence: KF4

## Concepts & Entities

### Conditions

lung adenocarcinoma

### Methods

survival modeling, Lasso, supervised principal components, time-dependent ROC analysis

### Models & Tools

IPPS, oncoPredict, pRRophetic

### Datasets

TCGA-LUAD, GSE13213, GSE26939, GSE30219, GSE31210, GSE42127, OAK, POPLAR

## Related Research

- [Deciphering the prognostic role of endoplasmic reticulum stress in lung adenocarcinoma: integrating prognostic prediction and immunotherapy strategies](https://drgezhang.com/papers/doi-10-1007-s10238-024-01439-4.html) — A separate retrospective LUAD stress-related prognostic signature; it does not independently validate IPPS. (DOI: 10.1007/s10238-024-01439-4)

## Publication & Provenance

- DOI: https://doi.org/10.1002/tox.24284
- Publisher: https://onlinelibrary.wiley.com/doi/10.1002/tox.24284
- PubMed: https://pubmed.ncbi.nlm.nih.gov/38591820/
- Author Uploaded Fulltext Url: https://www.researchgate.net/publication/379695827_Advancing_predictive_markers_in_lung_adenocarcinoma_A_machine_learning-based_immunotherapy_prognostic_prediction_signature
- Figure1 Url: https://www.researchgate.net/figure/Comparative-Study-of-Immunotherapy-Cohorts-and-Model-Gene-Selection-A-The-volcano-plot_fig1_379695827
- Figure2 Url: https://www.researchgate.net/figure/Model-Construction-and-Evaluation-A-Comparison-of-C-index-values-for-101-predictive_fig2_379695827
- Version: Author-uploaded publisher-layout early-online version, 2024, pp. 1-13, DOI 10.1002/tox.24284; Wiley download stamp 12 April 2024. The current publisher record lists Environmental Toxicology 39(10), 4581-4593, first online 9 April 2024.
- Evidence Basis: The complete article transcription through its references and final p. 13 footer was read on ResearchGate; Figures 1 and 2 were visually inspected on 26 October 2026. The formal title, ordered six-author list and DOI match the publisher record. The current issue PDF was not obtained, so complete byte-for-byte or post-pagination textual equivalence is not claimed.
- Uploader: Ge Zhang; ResearchGate author-content statement dated 14 April 2024.
- Accessed On: 2026-10-26
- Acquired Format: ResearchGate HTML full-text transcription and public figure-page screenshots; no PDF binary acquired.
- Fulltext Transcription Sha256: 3f27b079adb01e49b923a198959cd57d96201251458c3c47bb0005cd3a22daee
- License: Copyright 2024 Wiley Periodicals LLC. No open licence is asserted. Full text and screenshots are retained only in the private audit directory and are not distributed with the website.
- Reporting Discrepancies: OAK/POPLAR 577 versus selected 296; OS-defined CR/PR and PD/SD labels; GSE31210 AUC disagreement between Results 3.2 and Fig. 2L. These are disclosed without choosing an unsupported correction.

## Evidence-page notice

Retrospective model-development evidence with cohort reuse and unresolved reporting discrepancies. This page does not establish independent immunotherapy-response prediction or clinical treatment benefit.

## Links

- DOI: https://doi.org/10.1002/tox.24284
- HTML page: https://drgezhang.com/papers/doi-10-1002-tox-24284.html
- Google Scholar query: https://scholar.google.com/scholar?q=%22Advancing%20predictive%20markers%20in%20lung%20adenocarcinoma%3A%20A%20machine%20learning-based%20immunotherapy%20prognostic%20prediction%20signature%22
- ORCID: https://orcid.org/0000-0002-3116-3246