Sehhoon Park, Jongchan Park, Minuk Ma, Hyun-Ae Jung, Jong-Mu Sun, Yoon-La Choi, Jin Seok Ahn, Myung-Ju Ahn, Sanghoon Song, Gahee Park, Sukjun Kim, Huijeong Kim, Seunghwan Shin, Chan-Young Ock, Se-Hoon Lee<\/p>\n
AACR, 2023<\/strong><\/p>\n Background:<\/strong> As the KRAS G12C mutation became targetable in non-small cell lung cancer (NSCLC), tissue based KRAS mutation test is now an essential practice for the treatment decision. Recently, predicting KRAS mutations using deep-learning models with H&E images to potentially increase the pre-test probability has been reported with modest performance. Herein, we conducted a novel approach to improve the performance of KRAS G12C prediction based on an ensemble model trained not solely on H&E images, but also with multi-layered semantic content produced by a pre-trained artificial-intelligence (AI) analyzer, Lunit SCOPE IO.<\/p>\n Methods:<\/strong> The Cancer Genome Atlas LUAD and LUSC (TCGA-Lung) samples were used for model development. A self-supervised vision transformer was used to extract deep features from raw H&E images; and an AI-based pathology profiling analyzer extracted semantic contents such as the spatial information of tumor cells, lymphocytes, cancer epithelium, and cancer stroma. A set of classifiers was trained based on the two features, and the ensemble of these features was used to improve robustness. The final model was evaluated through cross-validation and assessed on independent NSCLC samples from Samsung Medical Center (SMC) who tested KRAS mutation by various methods including whole exome sequencing or target sequencing.<\/p>\n Results:<\/strong> TCGA-Lung dataset (n = 930) includes 150 (16.1%) KRAS driver mutations, and 62 (6.7%) KRAS G12C. The best cross-validation performances of the models predicting KRAS G12C, measured by mean area-under-the-curve (AUC), were 0.768 trained by only H&E images (HE-only), 0.714 by AI semantic content (AISC) with MLP classifier (AI-MLP), and 0.697 by AISC with random forest (AI-RF), respectively. An ensemble of the three models showed an increased AUC of 0.787 in TCGA-Lung by cross-validation. These models were applied to an independent SMC dataset (n = 363), including 54 (14.9%) KRAS driver mutations, and 22 (6.1%) KRAS G12C. The mean AUC to predict KRAS G12C by HE-only, AI-MLP, and AI-RF models were 0.599, 0.644, and 0.678, respectively, implying limited robustness. However, the AUC of an ensemble of the 3 models was 0.745 in the SMC dataset, showing 71.0% sensitivity and 72.7% specificity. Similar results were observed regardless of the KRAS testing method (TruSight Oncology 500 panel; n = 249; AUC 0.787, other tests; n = 114; AUC 0.720), the tissue size (surgical resection; n = 138; AUC 0.724, biopsy n = 225; AUC 0.763), and histology (excluding squamous cell carcinoma, n = 286; AUC 0.697).<\/p>\n