Multi-omics integration identifies three molecular subtypes of non-small-cell lung cancer with divergent immunotherapy response patterns

Anti-PD-1/PD-L1 checkpoint immunotherapy has transformed first- and second-line NSCLC management, but durable response rates remain below 30% in unselected populations. PD-L1 tumour proportion score (TPS) is the established response predictor but has limited discriminatory power: many TPS-high tumours do not respond, and a non-trivial fraction of TPS-low tumours do. A more biologically-grounded stratification would improve patient selection.

Three subtypes emerged from unsupervised consensus clustering across the transcriptome / methylome / proteome integrated representation. Subtype A (38% of tumours) — interferon-gamma-rewired, high HLA class I/II expression, prominent tertiary lymphoid structure signatures. Median checkpoint-response rate 54% (95% CI 49-59%); median OS 28.4 months. Subtype B (44%) — mesenchymal-shifted, low antigen presentation, neutrophil-dominant immune infiltrate. Median checkpoint-response rate 12% (95% CI 9-15%); median OS 11.2 months. Subtype C (18%) — KEAP1/NFE2L2 pathway-driven oxidative-stress phenotype, moderate immune infiltrate. Median checkpoint-response rate 27% (95% CI 22-32%); median OS 16.8 months. Subtype assignment predicted checkpoint response better than PD-L1 TPS in both discovery (AUROC 0.78 vs 0.64) and external validation cohorts (AUROC 0.74 vs 0.62).

If the subtype assignment can be operationalised into a clinical-grade assay, it would substantially improve immunotherapy patient selection — particularly for the Subtype B group, where the 12% response rate would argue against first-line single-agent checkpoint therapy in favour of alternative regimens (combination chemotherapy + checkpoint, or trial enrolment).

Tumours from TCGA-LUAD/LUSC (n=1,026), the Lung Cancer Mutation Consortium (LCMC, n=892), and two academic medical centres (UCLA and MSKCC, n=929) with matched RNA-seq, EPIC methylation array, and DIA mass-spectrometry proteome data. Multi-omics integration via MOFA+ factor analysis with 30 latent factors. Subtypes derived from k-means clustering on factor weights, k chosen via consensus clustering stability across 1,000 bootstrap iterations. Immunotherapy response assessed via RECIST 1.1 in patients receiving any PD-1 or PD-L1-targeting agent (n=1,420 across cohorts). External validation in two independent cohorts (POPLAR, OAK) using transcriptome-only subtype prediction models derived from the discovery integration. Survival analyses via Cox proportional hazards adjusted for age, sex, stage, smoking history, and prior treatment line.

The integration depends on bulk-tissue omics, which obscures tumour-immune spatial relationships that single-cell or spatial-omics data would resolve. Subtype C is the smallest group (18%) and its response-rate estimate carries wider uncertainty. Validation cohorts (POPLAR, OAK) used transcriptome-only subtype prediction; whether full multi-omics typing would improve prediction further is unknown. The cohorts skew non-Hispanic white (78%) and population generalisability beyond that demographic is limited.

Whether subtype assignment is stable across primary vs. metastatic tumour tissue is unresolved. The Subtype B group's poor response demands investigation of mechanism — whether the mesenchymal-low-antigen-presentation phenotype is targetable pharmacologically (e.g., epigenetic re-rewiring with HDAC or DNMT inhibitors) would be high-impact. Single-cell follow-up to characterise the immune-infiltrate composition per subtype is the next natural step.