Marc E. Lenburg

Background: Household air pollution (HAP) from biomass combustion is a major health burden and suspected lung carcinogen. The airway transcriptional response to wood smoke (WS) remains poorly understood.Methods: We analyzed nasal epithelial RNA-seq data from 29 samples collected from 25 never-smoking, nonpregnant women in a Guatemalan ancillary study of the Household Air Pollution Intervention Network (HAPIN) trial. Baseline gene expression was evaluated against continuous PM₂.₅ exposure and urinary 1-hydroxypyrene (1-OHP). Four women also provided paired samples before and after randomized transition to liquefied petroleum gas (LPG). Gene set enrichment analysis assessed overlap with smoky coal and cigarette smoke signatures and hallmark pathways; over-representation analysis evaluated 1-OHP-associated genes.Results: Higher 1-OHP was associated with 203 differentially expressed genes (FDR < 0.25); no genes were associated with PM₂.₅ at this threshold. Genes downregulated in smoky coal users were enriched among genes downregulated with increasing PM₂.₅ and 1-OHP. Higher 1-OHP was associated with enrichment of inflammatory and immune pathways, including TNFα signaling via NFκB and IL2-STAT5 signaling, and negative enrichment of oxidative phosphorylation and MYC targets. After LPG transition, genes previously identified as upregulated in active smokers in an independent study were enriched among genes downregulated in the paired samples from these never-smoking women.Conclusions: WS exposure, particularly as measured by urinary 1-OHP, was associated with inflammatory airway transcriptional changes and overlap with other combustion-related signatures. Some transcriptional changes may be reversible after exposure reduction.

Archetypes are associated with the somatic mutation landscape. A, Boxplot of MATH scores for Lung PCA PML samples with archetype assignments (n = 6 normal-like, n = 10 inflammation, n = 6 cell adhesion, n = 9 proliferation, and n = 38 secondary) and Lung PCA tumor samples (tumor samples were not included in PML archetype analysis, n = 21 samples). PML samples assigned to the proliferation archetype had significantly higher MATH scores compared with all other PMLs (linear mixed-effects model with archetype vs. not as binary fixed effect and patient as a random effect, P < 0.0001) but were not significantly different from tumor samples (linear mixed-effects model with archetype/tumor category as a fixed effect and patient as a random effect, contrasts were used to compare individual levels). B, Heatmap of z-statistic for the absence/presence of each driver mutation (rows) modeled as a function of distance from archetype (columns) in a combined analysis of Lung PCA (n = 57 AAH and n = 12 AIS/MIA samples) and GSE102511 (n = 13 AAH samples) samples with available exome sequencing data. EGFR driver mutations are significantly associated with the proliferation archetype. *, P < 0.05 (logistic mixed-effects model with distance from each archetype and study as fixed effects and patient as a random effect).

Gene modules recur across studies of LUAD premalignancy. A, Stacked bar plot of sample histology per patient from the Lung PCA cohort (n = 40 patients, n = 177 samples following filtering). B, Stacked bar plot showing the histology distribution for each cohort utilized in this analysis. The cohorts included were the Lung PCA (n = 80 PML, n = 48 normal samples, and n = 49 tumor samples), GSE102511 which contained paired normal/PML/tumor samples from 17 patients (n = 17 PML, n = 15 normal samples, and n = 15 tumor samples), GSE166720 which contained independent samples from 50 patients with either early-stage LUAD or MIA/AIS (n = 17 PML and n = 33 tumor samples), and GSE193725 which contained a mixture of paired normal/PML, normal/tumor, and independent samples from 42 total patients (n = 23 PML, n = 10 normal samples, and n = 18 tumor samples). C, Heatmap of summed binarized gene adjacency matrices from each of the four PML datasets after filtering to identify highly connected genes (n = 1628 genes). Heatmap shades indicate number of datasets in which the genes are correlated, ranging from 0 (white) to 4 (deep red). Rows/columns are semi-supervised by modules identified using igraph clustering, with color annotations indicating each module. D, Heatmap of GSVA scores computed from each PML-derived gene module. Row labels indicate pathway enrichment of the nine modules, and row colors correspond to modules identified in C. Columns are hierarchically clustered using ward.d2 similarity of the Euclidean distance of samples. E, Boxplots of GSVA scores computed for each PML-derived module across the Lung PCA cohort samples (n = 177 samples) stratified by module and colored by histology. *, histology is significantly different from AAH samples (linear mixed-effects models with histology as a fixed effect and patient as a random effect, P < 0.05).

The HTAN Lung PCA cohort and multiple publicly available datasets were used to identify LUAD PML archetypes and to associate these archetypes with clinical phenotypes of LUAD tumors. Participants in the Lung PCA cohort (n = 40 patients) underwent resection for LUAD. Normal, PML, and tumor tissues from the resection were profiled by bulk RNA-seq (n = 177 samples), with subsets of these samples additionally profiled by whole-exome sequencing (n = 90 samples) and/or mIF (n = 39 samples). The Lung PCA RNA-seq data, together with three additional PML cohorts, were used to define four molecular archetypes. The archetypes were then projected into large-scale LUAD studies to evaluate associations between PML molecular features and clinical and histopathologic features of indolent vs. aggressive LUAD. [Figure was created in BioRender. Mazzilli, S. (2026) https://BioRender.com/y253r09.]

Archetype analysis on PMLs demonstrates correlation of archetypes with molecular and clinical phenotypes. A, Scatterplots of first two PCs computed from GSVA scores for PML-derived modules across PML samples from all four cohorts. Each plot is colored by the distance of each sample from an archetype (vertex of polytope). Each sample has four associated distances—one for each archetype vertex. More negative numbers reflect samples further away from the vertex. B, Scatterplot of first two PCs computed from GSVA scores for PML-derived modules for all PML samples. Samples in the top 0.125 proportion of distances from each archetype were assigned archetypes, whereas other samples were labeled as “secondary,” with colors representing samples closest to each respective archetype. Shapes indicate histology of samples. C, Heatmap of t-statistics from linear mixed-effects models of GSVA scores for each PML-derived module (rows) modeled with distance from each archetype (columns) and study as fixed effects and patient as a random effect.*, FDR < 0.05 for the distance term. Rows are hierarchically clustered using ward.d2 similarity of the Euclidean distance. Row annotations represent the median GSVA score for each module in Lung PCA samples of either normal tissue or tumor. D, Visualization of χ2 residuals from χ2 test of association between archetypes and histology. Plot shows contribution of each comparison in the matrix to the test statistic, where circle size is proportional to contribution. Positive (red) values specify a positive association between the corresponding row and column variables (χ2 test, P = 0.18, n = 37 AAH and n = 19 AIS/MIA).

PML-derived module co-expression (A) Density plot of correlation of each gene’s expression values to module GSVA score stratified by genes within the module and genes outside the module for all PML data (left) and PML data from each cohort (B-D) Boxplots of GSVA scores computed for each PML-derived module across (B) GSE102511 samples (n = 47 samples) stratified by module and colored by histology. Asterisk (*) indicates histology is significantly different from AAH samples (linear mixed effects model with histology as fixed effect and patient as random effect, p < 0.05) (C) GSE193725 samples (n = 51 samples) stratified by module and colored by histology. Asterisk indicates histology is significantly different from AIS/MIA samples (linear mixed effects model with histology as fixed effect and patient as random effect, p < 0.05) (D) GSE166720 samples (n = 50 samples) stratified by module and colored by histology. Asterisk indicates histology is significantly different from AIS/MIA samples (linear model, p < 0.05).

Association of archetypes with LUAD features in GSE68465 Visualization of χ2 residuals from χ2 test of association of archetypes with tumor features in GSE68465 study. Plot shows contribution of each cell to the test statistic, where circle size is proportional to cell contribution. Positive (red) values specify a positive association between the corresponding row and column variables (A) Tumor grade in GSE68465 cohort (p < 0.0001) (B) Lymph node invasion (N0 versus N1 or above) in GSE68465 cohort (p = 0.0038).

PML archetypes in LUAD cohorts (A-H) Scatterplot of first two principal components computed from GSVA scores for PML-derived modules in LUAD datasets, with annotated projection of PML archetypes (A1-A4) into principal component space. Samples in the top 0.125 proportion of distances from each archetype were assigned archetypes, while other samples were labeled as “secondary”, with colors representing samples in the bin closest to archetype (A) Lung PCA tumor samples (B) Yoo and colleagues tumor samples (C) TRACERx tumor samples (D) GSE68465 tumor samples (E) GSE31210 tumor samples (F) GSE30219 tumor samples (G) TCGA tumor samples (H) Chen and colleagues PML samples.

Genes in each PML-derived module.

Supplementary File 1 from Archetype Analysis of Lung Adenocarcinoma Premalignancy Links Heterogeneity in Premalignant Lesions to Diverging Features of Invasive Disease