Machine-learning driven strategies for adapting immunotherapy in metastatic NSCLC

Area of Science:

• Oncology
• Computational Biology
• Genomics

Background:

• Immune checkpoint inhibitors (ICIs) improve survival in advanced non-small cell lung cancer (NSCLC).
• Current guidance for selecting between ICI monotherapy and combination chemotherapy is limited.
• Single biomarkers like PD-L1 are insufficient for predicting treatment response.

Purpose of the Study:

• To develop a machine learning model predicting individual benefit from adding chemotherapy to ICIs in advanced NSCLC.
• To estimate heterogeneous treatment effects using clinicogenomic data.
• To provide prospective guidance for optimizing NSCLC immunotherapy selection.

Main Methods:

• Developed an integrated machine learning model, A-STEP (Attention-based Scoring for Treatment Effect Prediction).
• Utilized clinicogenomic data from four large cohorts (totaling 2,220 patients).
• Calculated benefit scores using 28 genomic and 6 clinical features to predict treatment effect.

Main Results:

• A-STEP achieved the largest reduction in 3-month progression risk, improving weighted risk reduction by 13-23% compared to stand-alone models.
• The model recommended treatment changes for over 50% of patients, predominantly favoring ICI-Chemo.
• Simulations on an external cohort showed improved 2-year progression-free survival with A-STEP guided treatment (HR=0.60 for ICI-Mono, HR=0.58 for ICI-Chemo).
• Key predictive features included FBXW7, APC, and PD-L1.

Conclusions:

• Machine learning can address critical gaps in NSCLC immunotherapy selection by modeling treatment heterogeneity.
• A-STEP enables precision medicine by moving beyond conventional biomarker limitations.
• This approach optimizes treatment decisions for advanced NSCLC patients, improving outcomes.