What No One Has Done Yet
The gaps identified below are not merely absences in the literature -- they are the strongest original contributions this review can make to the field. A review that only summarizes existing knowledge is a textbook. A review that maps the uncharted territory is a roadmap.
The Three Highest-Priority Gaps
These findings emerged from systematic analysis across multiple themes and the author's census of 25 multi-omics lung cancer studies. They represent the review's most powerful original contributions.
1. Feature Selection Philosophy Determines Everything -- And the Field Has Not Confronted This
A systematic census of 25 multi-omics lung cancer studies revealed that the single most consequential methodological choice -- how features are selected before integration -- determines downstream results more than the integration algorithm itself. Of 18 MOVICS-based studies, 13 use Cox regression-based (supervised) feature selection, filtering to ~2,000-3,500 survival-associated features and invariably yielding 2 subtypes (high-risk vs. low-risk). Only 3 studies use variance-based (unsupervised) selection with ~10,000 features, yielding 4-5 biologically distinct subtypes.
The hidden problem is that prognosis-first approaches collapse biologically distinct subtypes with similar survival into a single "high-risk" group. Two tumors may both have poor survival but for completely different reasons -- one is immune-excluded (potentially responsive to angiogenesis inhibitors to improve immune access), the other is metabolically dysfunctional (potentially responsive to metabolic inhibitors). Binary risk stratification merges them, losing the therapeutic information needed for precision medicine.
Compounding this, clinical features (age, stage, smoking status, sex, performance status) are almost never integrated WITH molecular features -- they are used for post-hoc validation but not as input features. No study has systematically tested whether multi-omics molecular features add incremental predictive value over clinical features alone. Feature selection is almost exclusively univariate (one gene at a time), ignoring multivariate methods like elastic net, WGCNA network modules, or stability selection that could identify more robust and parsimonious feature sets.
What the manuscript should argue: The field needs a hierarchical approach -- biological clustering first to preserve mechanistic heterogeneity, followed by within-subtype risk stratification using both molecular and clinical features. Feature selection methodology should be reported with the same rigor as integration algorithm choice. See Chapter 13 for the full analysis.
2. Sex-Stratified Multi-Omics Analyses Are Nearly Absent
Across themes 01 (Molecular Heterogeneity), 03 (Multi-Omics Applications), 04 (AI/ML), and 05 (Sex/Gender), we found that virtually no multi-omics study has systematically stratified results by biological sex. This is not a niche concern -- it is a blind spot affecting the validity of subtype classifications, drug response predictions, and biomarker panels that are implicitly trained on sex-mixed cohorts.
The evidence for sex-linked biological differences in lung cancer is strong: women have higher EGFR mutation rates, estrogen receptor signaling modulates tumor biology [PMID: 11905727; PMID: 38607364], and immunotherapy outcomes may differ by sex [PMID: 29778737]. Yet the major proteogenomic studies (CPTAC, East Asian cohorts), the AI/ML models for subtype classification, and the multi-omics integration benchmarks all report sex-pooled results. The consequence is that molecular subtypes, prognostic signatures, and drug response models may perform differently in men and women -- but we have no way of knowing, because the stratified analyses have not been done.
What the manuscript should argue: Sex-stratified multi-omics analysis should be a standard requirement, not an optional subgroup analysis. The pharmacogenomic databases (GDSC, CCLE) should report sex-stratified drug response data. AI/ML models should report performance metrics separately by sex.
3. No Multi-Omics Study Has Focused on NSCLC in Never-Smokers
Despite never-smoker lung cancer accounting for 10-25% of all lung cancers globally -- and rising in incidence -- the intersection of multi-omics integration (Theme 03) with never-smoker biology (Theme 06) is an empty set. Every existing multi-omics NSCLC study pools smokers and never-smokers together, diluting the signal from a population with fundamentally different driver mutations, mutational signatures, and tumor microenvironment characteristics.
Zhang et al. [PMID: 34493867] demonstrated that never-smoker LUAD comprises three genomically distinct subtypes (piano, mezzo, forte). Yet none of the CPTAC proteogenomic studies, none of the multi-omics subtyping studies, and none of the spatial omics studies have been designed specifically around these subtypes. The proteogenomic characterization of East Asian never-smokers by Chen et al. [PMID: 32649875] comes closest but was not a dedicated multi-omics integration study.
What the manuscript should argue: A dedicated, multi-ethnic, multi-omics study of never-smoker NSCLC -- integrating genomics, transcriptomics, proteomics, epigenomics, and spatial profiling -- is among the most impactful studies that could be conducted in lung cancer research today.
High-Priority Gaps
| Gap | Where it emerged | Why it matters for the field |
|---|---|---|
| Exposome-to-molecular-subtype mapping is absent | Themes 06, 07 | We know PM2.5 causes lung cancer. We know never-smoker LUAD has three subtypes. No one has connected which exposures produce which subtypes. The Martinez-Ruiz PM2.5-EGFR promotion model [PMID: 37046093] is the only molecular bridge between environmental exposure and subtype-specific biology. |
| AI fairness/bias assessment rarely performed in oncology | Theme 04 | Foundation models trained predominantly on Western cohorts are being deployed globally. Sex-stratified AI model evaluation is virtually absent. AI tools for never-smoker NSCLC do not exist. The generalizability crisis in medical AI is well-documented but unresolved in oncology specifically. |
| Few ML models validated on multi-ethnic, multi-institutional cohorts | Themes 04, 10 | External validation studies consistently show performance degradation. Most published AI models have been developed and validated within single institutions or single ancestry populations. |
| Composite immune biomarker panels not standardized | Theme 09 | Despite a decade of research showing that PD-L1, TMB, gene expression signatures, and microbiome composition each contribute partial predictive information, no composite panel has reached guideline status for clinical use. |
| Drug repurposing predictions rarely translate to clinical trials | Theme 11 | Computational drug repurposing generates many candidates. Very few reach clinical validation. The translation bottleneck is not computational -- it is experimental and regulatory. |
Moderate-Priority Gaps
| Gap | Where it emerged | Why it matters |
|---|---|---|
| Cost-effectiveness of multi-omics precision oncology not established | Theme 10 | Without health-economic evidence, payers will not cover multi-omics testing at scale. The clinical utility argument is necessary but not sufficient. |
| Single-cell epigenomics in lung cancer is limited | Themes 08, 12 | Bulk epigenomic studies cannot resolve cell-type-specific regulatory mechanisms. The intersection of single-cell resolution with epigenomic profiling in lung cancer is nearly empty. |
| Spatial omics in never-smoker lung cancer is absent | Themes 06, 12 | The spatial architecture of the tumor microenvironment in the most common form of never-smoker LUAD has never been mapped. |
| Pre-neoplastic lesion molecular evolution poorly understood | Theme 01 | We classify lung cancer subtypes at diagnosis, but the molecular events that determine subtype commitment during premalignant evolution are largely unknown. |
| Longitudinal molecular profiling through treatment is sparse | Themes 01, 08, 12 | Most molecular studies are cross-sectional snapshots. Understanding treatment-driven evolution requires serial sampling and multi-omics profiling. |
| miRNA/lncRNA prognostic signatures have reproducibility concerns | Theme 08 | The 76 retractions in our epigenetics collection are quantifiable evidence of a reproducibility crisis. Many published non-coding RNA biomarker associations may not hold up under independent validation. |
| Pregnancy-associated lung cancer molecularly understudied | Theme 05 | An important and understudied clinical scenario where hormonal, immunological, and molecular factors intersect. |
The Retraction Signal as a Gap
The 208 retracted papers identified across this collection are not just a quality control statistic -- they point to specific areas of the literature where evidence quality should be treated with caution.
| Theme | Retracted | Interpretation |
|---|---|---|
| 08 Epigenetics | 76 | Crisis-level. Concentrated in miRNA/lncRNA prognostic signatures. The field needs pre-registered validation studies. |
| 01 Molecular Heterogeneity | 50 | Elevated but spread across subtypes -- no single hotspot. |
| 04 AI/ML | 30 | Moderate. Some in radiomics, some in prognostic modeling. |
| 11 Drug Repurposing | 28 | Moderate. Network pharmacology area flagged. |
| 09 Immune Biomarkers | 14 | Expected background rate for a large corpus. |
| All others | 10 total | Clean. |
What the manuscript should argue: The concentration of retractions in non-coding RNA biomarker studies should be explicitly discussed as evidence for the need for rigorous, pre-registered validation frameworks. The review should recommend specific reproducibility standards (independent validation cohorts, pre-registered analysis plans, code/data sharing) for future biomarker studies.