Multimodal AI Identifies a High-Risk, Immune-Cold NF1-mutant sqNSCLC Patient Subtype Enriched by a 31-Gene Prognostic Signature

IASLC WCLC 2026

Aug 19, 2026
Oncology
Presentation

S. Kaushik, M. Carty, A. Singhania, J. Guinney, R.M. Johnson, J. Patel

Introduction
Squamous Non small Cell Lung Cancer (SqNSCLC) is a heterogeneous disease characterized by poor clinical outcomes. Previous efforts to characterize the molecular landscape of SqNSCLC have used small sample sizes and emphasized single data modalities. To enable precise therapeutic strategies, we employed a multimodal AI framework to integrate gene expression, copy number alterations (CNA), and somatic mutation data from the Tempus real-world database to categorize SqNSCLC into distinct molecular subtypes. Analysis of these subtypes yielded clinically actionable vulnerabilities and patient stratification strategies critical for advancing the next-generation of lung cancer treatments.

 

Methods
We analyzed de-identified clinico-genomic records from SqNSCLC patients profiled with Tempus DNA (xT; 648 genes) and RNA (xR; whole transcriptome) assays. We developed a multimodal autoencoder for molecular subtyping that integrates gene expression, CNA, and mutation profiles from 4,973 SqNSCLC biopsies. Modality-specific encoders were trained and optimized to align latent spaces across encoders. Clustering of patients using the learned joint embeddings were used to identify patient subtypes. Elastic net penalized Cox proportional hazard regression model was used to develop a prognostic signature. Tempus spatial transcriptomics data was used to characterize the tumor microenvironment.

 

Results
Multimodal AI identified seven distinct subtypes of SqNSCLC associated with significant differences in real-world overall survival (rwOS; p=0.02). These subtypes were defined by distinct mutational landscapes and immune microenvironments: the C2 (11%) and C5 (14%) subtypes are immune-rich and characterized by RB1 alterations, while the C3 (28%) and C7 (29%) subtypes represent immune-cold tumors driven by NFE2L2 mutations and high proliferation rate. Notably, C1 (15%) and C6 (12%) subtypes shared enrichment of NF1 mutations and high stromal content. Despite both sharing upregulated EMT and TGF-β signaling, they exhibit differences in their immune profiles where C1 is an immune-cold subtype associated with the worst prognosis (11.7 months; CI 9.4-16.3), and C6 is immune-rich (17.7 months; CI 10.78-23.6). To stratify high risk patients, a prognostic signature of 31 genes was identified for the C1 subtype. Multivariate CoxPH confirmed this signature as an independent predictor of rwOS and rwPFS (Figure 1A). Spatial transcriptomics links high C1 signature expression with tumor infiltration of cancer associated fibroblasts (CAFs), while low C1 signature showed compartmentalized topology and stromal lymphocytes (Figure 1B).

 

Conclusions
Our study identified a high risk, immune-cold, NF1-mutant subtype spatially enriched with infiltrative CAFs and may inform future therapeutic strategies targeting TGF-β and EMT pathways in stroma-rich tumors.