A Deep Learning Framework Integrating Tumor Microenvironmental Features Accurately Predicts Multiple Driver Gene...

Area of Science:

• Oncology
• Computational Pathology
• Genomics

Background:

• Deep learning (DL) shows promise for predicting gene mutations from lung cancer histopathology slides.
• Current DL methods often lack precision in identifying driver mutation subtypes or exonic variants, limiting clinical translation for targeted therapy.

Purpose of the Study:

• To develop an advanced DL framework for precise prediction of driver mutations and tumor mutational burden (TMB) in lung cancer using histopathology images.
• To integrate tumor microenvironment (TME) features for enhanced mutation prediction accuracy.

Main Methods:

• Assembled a large multicenter dataset of 2573 lung cancer patients with paired pathology images and next-generation sequencing data.
• Developed NAVF-Bio, an adaptive multi-view feature fusion framework using multiple-instance learning to integrate TME features from whole-slide images (WSIs).
• Benchmarked NAVF-Bio against 11 state-of-the-art DL methods for predicting driver mutations (TP53, EGFR, KRAS, ALK) and TMB status.

Main Results:

• NAVF-Bio significantly outperformed existing DL models in predicting driver mutations and TMB status, demonstrating clinically relevant performance in external validation.
• The framework accurately predicted mutated driver gene exons across different centers.
• Interpretability analyses confirmed NAVF-Bio's ability to identify pathologically relevant tumor features.

Conclusions:

• NAVF-Bio offers a robust approach for predicting lung cancer driver gene mutations and TMB from WSIs, mimicking pathologist workflows.
• This tool can facilitate driver gene mutation screening, potentially guiding targeted therapy selection for lung cancer patients.
• The developed multi-gene mutation prediction platform aids in personalized treatment strategies.