A Distinct Molecular Disease. Lung cancer in never-smokers is increasingly recognized as a biologically distinct entity from smoking-associated lung cancer. Never-smokers account for approximately 10-25% of all lung cancer cases globally, and their tumors harbor different driver mutations -- most commonly EGFR mutations and ALK fusions -- compared to the KRAS mutations and high tumor mutational burden typical of smoking-associated cancers.
The Cost and Accessibility of Molecular Testing. Comprehensive genomic profiling of lung adenocarcinoma requires next-generation sequencing (NGS) of tumor tissue, which is expensive, time-consuming, and dependent on having sufficient tissue from biopsy. In many clinical settings, particularly lower-resource environments, full genomic characterization is not routinely accessible for all patients who might benefit from targeted therapy.
Histology as a Proxy for Genomics. Hematoxylin and eosin (H&E)-stained slides are produced routinely for every resected tumor as part of standard pathological diagnosis. Deep learning models applied to whole-slide images (WSIs) have demonstrated the ability to predict molecular alterations directly from tissue morphology -- essentially using the spatial patterns that histology encodes about underlying tumor biology to infer genomic features without additional molecular testing.
Study Purpose. The Sherlock-Lung study assembled an international multi-institutional cohort of never-smoker lung adenocarcinoma cases to characterize the genomic landscape of this underrepresented population. This work extended that effort by developing a multilabel deep learning model to simultaneously predict 16 molecular features from H&E WSIs, creating a potential non-molecular tool for comprehensive genomic profiling.
Sherlock-Lung Cohort. The study used 495 whole-slide images from 410 never-smoker lung adenocarcinoma patients drawn from the Sherlock-Lung international multi-institutional cohort. This cohort was specifically assembled to capture the genomic diversity of never-smoker lung cancer across contributing centers, providing pathologically confirmed cases with comprehensive molecular characterization from NGS data as ground truth labels.
Multilabel CNN Architecture. Researchers developed a customized multilabel convolutional neural network based on the ResNet50 backbone architecture. Unlike standard binary or multiclass classifiers trained on single outcomes, this multilabel design was specifically optimized to simultaneously predict 16 different molecular alterations from a single WSI in one forward pass. The architecture was modified to handle the correlated, co-occurring nature of molecular alterations that characterize real tumor biology.
Targets: 16 Molecular Alterations. The model was trained to predict a comprehensive set of features including driver mutations (EGFR, KRAS, TP53, RBM10), specific EGFR hotspot mutations (p.L858R, p.E746_A750del), structural alterations (ALK fusion, CDKN2A deletion, MDM2 amplification), whole-genome doubling (WGD), and mutational signatures (Kataegis, APOBEC), as well as tumor mutational burden (TMB). This breadth distinguishes the approach from prior models that typically targeted one or two alterations at a time.
Benchmark Comparisons. The customized multilabel model was compared against standard Inception-V3 and unmodified ResNet50 architectures to evaluate whether the architectural modifications specifically optimized for multilabel prediction conferred performance advantages. This comparison allowed assessment of both the choice of backbone and the value of the multilabel design relative to running multiple independent binary classifiers.
Patch-Based Processing. Whole-slide images are far too large to process directly as single inputs to convolutional neural networks. The model used a patch-based approach, dividing each WSI into smaller image tiles for feature extraction. Tile-level predictions were then aggregated to produce slide-level predictions for each of the 16 molecular alterations, balancing computational feasibility with preservation of spatial tissue context.
Handling Class Imbalance. Molecular alterations vary substantially in prevalence within the cohort -- EGFR mutations are common in never-smoker adenocarcinoma while APOBEC signature mutations are less frequent. Training strategies to address class imbalance were essential to prevent the model from defaulting to predicting the majority class and were implemented for each of the 16 prediction targets independently within the multilabel framework.
Grad-CAM Interpretability. To understand which histological regions drove the model's predictions, gradient-weighted class activation mapping (Grad-CAM) was applied to generate heatmaps overlaid on the original WSI tiles. These heatmaps highlighted spatial regions most influential for each molecular prediction, enabling pathologists to assess whether the model's attention aligned with biologically plausible tissue structures such as tumor infiltrating lymphocytes, nuclear morphology, or stromal patterns.
Interactive Web Viewer. To make the model's predictions and interpretability maps accessible to researchers and clinicians, an interactive web viewer was developed and made publicly available through the project's GitHub repository. This viewer allows users to explore slide-level predictions and Grad-CAM heatmaps for each of the 16 molecular features, supporting both research use and educational applications in computational pathology.
Strong Performance for 11 Features. The customized multilabel CNN achieved AUROC values of 0.84 to 0.93 across 11 of the 16 molecular targets. The highest performance was observed for MDM2 amplification (AUROC 0.93), EGFR mutation (0.93), EGFR p.L858R hotspot (0.92), KRAS mutation (0.92), TP53 mutation (0.91), and Kataegis mutational signature (0.91). These AUROC values indicate strong discriminative ability between positive and negative cases for these alterations.
Intermediate Performance for Structural Features. Features involving structural genomic alterations showed slightly lower but still clinically meaningful performance: RBM10 mutation (AUROC 0.90), CDKN2A deletion (0.89), ALK fusion (0.86), whole-genome doubling (0.84), and EGFR p.E746_A750del hotspot (0.86). These results demonstrate that morphological correlates of chromosomal instability and structural rearrangements are partially but not fully encoded in H&E tissue patterns.
Lower Performance for Mutational Signatures. Tumor mutational burden (AUROC 0.67) and APOBEC mutational signature (0.57) showed substantially lower discriminative ability. KRAS hotspot mutations also underperformed relative to overall KRAS status. The APOBEC result near random chance is biologically plausible: APOBEC-driven mutations affect DNA sequence context without necessarily creating distinctive histological patterns that are visually distinguishable in H&E images.
Architecture Comparison. The customized multilabel CNN outperformed Inception-V3 on the majority of the 16 prediction targets and showed comparable or superior performance relative to standard ResNet50 across most features. This confirms that the multilabel architecture modifications -- designed to capture correlations between co-occurring genomic alterations -- provided a genuine performance advantage over running equivalent single-target models independently.
First DL Model for Never-Smoker Adenocarcinoma. This study represents the first deep learning model specifically designed and validated for the never-smoker lung adenocarcinoma population. Prior computational pathology work on lung cancer has predominantly used smoking-associated cohorts where KRAS mutation and high TMB are common, meaning the histological patterns learned may not transfer to never-smoker tumors with their distinct driver mutation landscape dominated by EGFR and ALK alterations.
Clinical Value for Triage and Access. In resource-limited settings where comprehensive NGS is unavailable or prohibitively expensive, a model achieving AUROC above 0.90 for EGFR status and ALK fusions could serve as a preliminary triage tool -- identifying patients most likely to harbor actionable alterations for targeted therapy. Patients flagged as high-probability EGFR-positive could be prioritized for molecular testing, optimizing laboratory resource allocation.
Morphological Correlates of Molecular Alterations. Grad-CAM analyses suggested that the model attended to biologically plausible tissue regions for different alterations, including nuclear features, architectural growth patterns, and tumor-stroma interfaces. The spatial attention patterns provide insight into the histomorphological correlates of specific driver mutations, with potential implications for understanding how genomic alterations shape tumor tissue architecture.
Multilabel Design Captures Co-Occurrence. The multilabel architecture explicitly models the joint distribution of co-occurring molecular alterations, which is biologically more realistic than treating each alteration as independent. EGFR mutations rarely co-occur with KRAS mutations, but co-occur with specific copy number alterations at non-random rates. Capturing these correlations during training may help the model learn representations that reflect the actual molecular landscape of these tumors rather than each alteration in isolation.
Cohort Size and Generalizability. The study used 495 WSIs from 410 patients, which while among the larger never-smoker cohorts assembled for any computational pathology study, remains relatively small for training and validating a 16-output multilabel model. External validation in independent never-smoker cohorts from diverse geographic and institutional sources is needed to confirm that performance generalizes beyond the Sherlock-Lung institutions.
Lower Performance Features Require Caution. The near-chance AUROC for APOBEC signature and moderate TMB performance mean that these outputs should not be used clinically without substantial additional validation and improvement. The model's differential performance across features implies that some molecular alterations may not be reliably inferable from H&E morphology with current architectures and cohort sizes, regardless of modeling approach.
Expanding the Application. Future work could incorporate additional tissue modalities such as immunohistochemistry alongside H&E to improve performance for features with low AUROC. Integration with clinical metadata including CT imaging features and patient demographics could further refine predictions. Prospective validation studies in clinical workflows -- where the model's output influences actual testing decisions -- are the necessary next step toward clinical translation.
A Platform for Never-Smoker Research. Beyond the immediate clinical application, this model and the associated interactive viewer provide a research platform for exploring the histomolecular relationships unique to never-smoker lung cancer. As larger never-smoker cohorts are assembled through international collaborations, retraining on expanded datasets will likely yield substantially improved performance for the lower-performing molecular features and broader generalizability.