Genomic Characterization of Lung Cancer in Never-Smokers Using Deep Learning

Area of Science:

• Computational pathology and bioinformatics
• Artificial intelligence in oncology
• Molecular diagnostics

Background:

• Deep learning shows potential for inferring genetic features from whole-slide images (WSIs).
• Limited research exists on applying these methods to never-smoker lung adenocarcinoma (NS-LUAD), a distinct cancer subset.
• Existing models often focus on smoker populations with restricted molecular scope and variable performance.

Purpose of the Study:

• To develop and validate a customized deep convolutional neural network for multilabel classification of NS-LUAD.
• To enable simultaneous prediction of 16 molecular alterations directly from H&E-stained WSIs.
• To optimize the model for reduced computational complexity while maintaining predictive accuracy.

Main Methods:

• A customized ResNet50-based deep convolutional neural network was designed with architectural modifications for multilabel classification.
• The model incorporated a simplified residual block, selective shortcut connections, and a sigmoid-based classification head.
• Training and evaluation were performed on 495 WSIs from the Sherlock-Lung study using 10-fold cross-validation and a held-out validation set.

Main Results:

• The model achieved high AUROC values (0.84-0.93) for 11 molecular features, including EGFR mutations, KRAS mutations, TP53 mutations, RBM10 mutations, MDM2 amplification, kataegis, CDKN2A deletion, ALK fusion, and whole-genome doubling.
• Performance for tumor mutational burden (AUROC=0.67), APOBEC signature (AUROC=0.57), and specific KRAS hotspot mutations (p.G12C: AUROC=0.74, p.G12V: AUROC=0.55, p.G12D: AUROC=0.43) was moderate to low.
• The proposed model demonstrated significantly improved performance compared to established architectures like Inception-v3 for most features.

Conclusions:

• The developed deep learning model effectively predicts multiple molecular alterations in NS-LUAD from WSIs.
• This approach holds potential for improving the triaging of patients for molecular testing.
• Further optimization could lead to enhanced precision treatment strategies for NS-LUAD patients.