OWL: an optimized and independently validated machine learning prediction model for lung cancer screening based on the UK Biobank, PLCO, and NLST populations.

BACKGROUND: A reliable risk prediction model is critically important for identifying individuals with high risk of developing lung cancer as candidates for low-dose chest computed tomography (LDCT) screening. Leveraging a cutting-edge machine learning technique that accommodates a wide list of questionnaire-based predictors, we sought to optimize and validate a lung cancer prediction model. METHODS: We developed an Optimized early Warning model for Lung cancer risk (OWL) using the XGBoost algorithm with 323,344 participants from the England area in UK Biobank (training set), and independently validated it with 93,227 participants from UKB Scotland and Wales area (validation set 1), as well as 70,605 and 66,231 participants in the Prostate, Lung, Colorectal, and Ovarian cancer screening trial (PLCO) control and intervention subpopulations, respectively (validation sets 2 & 3) and 23,138 and 18,669 participants in the United States National Lung Screening Trial (NLST) control and intervention subpopulations, respectively (validation sets 4 & 5). By comparing with three competitive prediction models, i.e., PLCO modified 2012 (PLCOm2012), PLCO modified 2014 (PLCOall2014), and the Liverpool Lung cancer Project risk model version 3 (LLPv3), we assessed the discrimination of OWL by the area under receiver operating characteristic curve (AUC) at the designed time point. We further evaluated the calibration using relative improvement in the ratio of expected to observed lung cancer cases (RIEO), and illustrated the clinical utility by the decision curve analysis. FINDINGS: For general population, with validation set 1, OWL (AUC = 0.855, 95% CI: 0.829-0.880) presented a better discriminative capability than PLCOall2014 (AUC = 0.821, 95% CI: 0.794-0.848) (p < 0.001); with validation sets 2 & 3, AUC of OWL was comparable to PLCOall2014 (AUCPLCOall2014-AUCOWL < 1%). For ever-smokers, OWL outperformed PLCOm2012 and PLCOall2014 among ever-smokers in validation set 1 (AUCOWL = 0.842, 95% CI: 0.814-0.871; AUCPLCOm2012 = 0.792, 95% CI: 0.760-0.823; AUCPLCOall2014 = 0.791, 95% CI: 0.760-0.822, all p < 0.001). OWL remained comparable to PLCOm2012 and PLCOall2014 in discrimination (AUC difference from -0.014 to 0.008) among the ever-smokers in validation sets 2 to 5. In all the validation sets, OWL outperformed LLPv3 among the general population and the ever-smokers. Of note, OWL showed significantly better calibration than PLCOm2012, PLCOall2014 (RIEO from 43.1% to 92.3%, all p < 0.001), and LLPv3 (RIEO from 41.4% to 98.7%, all p < 0.001) in most cases. For clinical utility, OWL exhibited significant improvement in average net benefits (NB) over PLCOall2014 in validation set 1 (NB improvement: 32, p < 0.001); among ever smokers of validation set 1, OWL (average NB = 289) retained significant improvement over PLCOm2012 (average NB = 213) (p < 0.001). OWL had equivalent NBs with PLCOm2012 and PLCOall2014 in PLCO and NLST populations, while outperforming LLPv3 in the three populations. INTERPRETATION: OWL, with a high degree of predictive accuracy and robustness, is a general framework with scientific justifications and clinical utility that can aid in screening individuals with high risks of lung cancer. FUNDING: National Natural Science Foundation of China, the US NIH.

A Large-Scale Genome-Wide Gene-Gene Interaction Study of Lung Cancer Susceptibility in Europeans With a Trans-Ethnic Validation in Asians.

INTRODUCTION: Although genome-wide association studies have been conducted to investigate genetic variation of lung tumorigenesis, little is known about gene-gene (G × G) interactions that may influence the risk of non-small cell lung cancer (NSCLC). METHODS: Leveraging a total of 445,221 European-descent participants from the International Lung Cancer Consortium OncoArray project, Transdisciplinary Research in Cancer of the Lung and UK Biobank, we performed a large-scale genome-wide G × G interaction study on European NSCLC risk by a series of analyses. First, we used BiForce to evaluate and rank more than 58 billion G × G interactions from 340,958 single-nucleotide polymorphisms (SNPs). Then, the top interactions were further tested by demographically adjusted logistic regression models. Finally, we used the selected interactions to build lung cancer screening models of NSCLC, separately, for never and ever smokers. RESULTS: With the Bonferroni correction, we identified eight statistically significant pairs of SNPs, which predominantly appeared in the 6p21.32 and 5p15.33 regions (e.g., rs521828C6orf10 and rs204999PRRT1, ORinteraction = 1.17, p = 6.57 × 10-13; rs3135369BTNL2 and rs2858859HLA-DQA1, ORinteraction = 1.17, p = 2.43 × 10-13; rs2858859HLA-DQA1 and rs9275572HLA-DQA2, ORinteraction = 1.15, p = 2.84 × 10-13; rs2853668TERT and rs62329694CLPTM1L, ORinteraction = 0.73, p = 2.70 × 10-13). Notably, even with much genetic heterogeneity across ethnicities, three pairs of SNPs in the 6p21.32 region identified from the European-ancestry population remained significant among an Asian population from the Nanjing Medical University Global Screening Array project (rs521828C6orf10 and rs204999PRRT1, ORinteraction = 1.13, p = 0.008; rs3135369BTNL2 and rs2858859HLA-DQA1, ORinteraction = 1.11, p = 5.23 × 10-4; rs3135369BTNL2 and rs9271300HLA-DQA1, ORinteraction = 0.89, p = 0.006). The interaction-empowered polygenetic risk score that integrated classical polygenetic risk score and G × G information score was remarkable in lung cancer risk stratification. CONCLUSIONS: Important G × G interactions were identified and enriched in the 5p15.33 and 6p21.32 regions, which may enhance lung cancer screening models.