Integrative multi-omics characterization and AI-driven biomarker discovery for NSCLC stage III outcomes

Background: The PACIFIC trial established consolidation durvalumab as standard of care for unresectable stage III non-small cell lung cancer (NSCLC) following concurrent chemoradiotherapy (CRT). Real-world studies such as PACIFIC-R confirmed these benefits; however, comprehensive characterization of molecular features in this setting remains limited. Advances in AI enable integration of heterogeneous genomic and transcriptomic data, opening new opportunities for biomarker discovery. Our study applies a multimodal AI framework combining contrastive learning and foundational-model representations to identify distinct molecular features associated with survival outcomes from pre-treatment tumor data. Methods: Stage III NSCLC EGFR / ALK -negative patients with unresectable disease who received CRT with subsequent durvalumab (CRT+D, N=281) or CRT alone (N=72) were identified from the Tempus AI multimodal real-world database. Molecular characterization included assessment of radiation and immunotherapy (IO)-related biomarkers to evaluate the impact on real-world progression-free survival (rwPFS). To identify potential DNA and RNA biomarkers, a multimodal AI framework was developed via integration of two recently published models: 1) COMPASS, a foundation model encoding RNA profiles into 43 interpretable immune-related concept scores, and 2) Predictive Biomarker Mapping Framework, a deep learning model mapping molecular features to treatment outcomes in CRT+D when compared to CRT-alone. Results: In the CRT+D group, patients with high levels of biomarkers linked to IO response, such as PD-L1 ≥50% (p-value = 0.01) and a high tumor mutational burden (p-value = 0.02), showed improved survival rates. Patients exhibiting low STK11 gene expression signature had shortened rwPFS (p-value=0.03). Tumors classified as radioresistant using the radiosensitivity index also showed diminished rwPFS benefit (p-value=0.02). Our multimodal AI framework revealed additional markers associated with improved rwPFS to CRT+D therapy, including tumors with adenocarcinoma histology, mutations in LRP1B , NF1 , KRAS , and CDKN2A , and enrichment in expression of immune activation pathways (cytotoxic T-cell, IFN-gamma, tertiary lymphoid structure, and immune-checkpoint signatures). In contrast, tumors with mutations in TP53 , RB1 , NOTCH1 , CUX1 , and STK11 as well as those with elevated expression of regulatory and stromal signals (Tregs, exhaustion, stroma) were linked to worsened outcomes in patients treated with CRT+D. Cross-validation confirmed model reproducibility and feature stability. Conclusions: Our study demonstrates the utility of a multi-omics approach to characterize molecular landscape of tumors treated with consolidation durvalumab and to drive biomarker discovery using an AI-driven, foundational-model-based framework.