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

Area of Science:

• Computational pathology
• Oncology
• Bioinformatics

Background:

• Deep learning shows promise for inferring genetic features from whole-slide images (WSIs).
• Lung adenocarcinomas in never-smokers (NS-LUAD) are molecularly distinct but understudied by current deep learning models.
• Existing models often focus on smoker populations with limited molecular scope and variable performance.

Purpose of the Study:

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

Main Methods:

• A ResNet50-based deep convolutional neural network with architectural modifications (simplified residual blocks, selective shortcuts, sigmoid classification head) was developed.
• The model was trained and evaluated on 495 WSIs from the Sherlock-Lung study using a 70% training, 10% internal test, and 30% held-out validation set split.
• Performance was assessed using area under the receiver operating characteristic curve (AUROC) for predicting 16 molecular features.

Main Results:

• The model achieved high AUROC values (0.84-0.93) for 11 features, including EGFR, KRAS, TP53 mutations, and ALK fusions.
• Performance was moderate for tumor mutational burden (AUROC=0.67) and KRAS hotspot mutations (e.g., p.G12C: AUROC=0.74).
• The customized model outperformed established architectures like Inception-v3 for most features on the same dataset.

Conclusions:

• The developed deep learning model demonstrates significant potential for predicting molecular alterations in NS-LUAD directly from WSIs.
• This approach could facilitate triaging for molecular testing and inform precision treatment strategies for never-smoker lung cancer patients.
• Further optimization may enhance its utility in clinical decision-making for NS-LUAD.