Abstract 2750: Leveraging machine-learning approaches to dissect drivers of clinical metastatic dynamics in lung adenocarcinoma

Abstract Background: Non-small cell lung cancer (NSCLC) survival is closely linked to metastatic progression. 5-year survival rates drop from >65% when localized to ~9% when distant metastases occur. Despite this impact on mortality, drivers of overall metastasis and site-specific organotropism are not well understood in NSCLC. To investigate these questions, we focused on Lung adenocarcinoma (LUAD), the most common subtype of NSCLC. Although previous efforts have been made to examine LUAD organotropism, we sought to expand upon longitudinal modeling approaches and explore associations in an independent large clinical and genomic annotated cohort. Methods: Using the artificial intelligence model from (Kehl et al. 2021), we annotated a clinico-genomic cohort of 2777 patients with lung adenocarcinoma, resulting in 59,177 annotated imaging reports. This resulted in time to metastatic site annotations for each patient, with median of 2 years follow up. Metastatic sites included brain, bone, adrenal, liver, lymph node, and mesentery. All patients were evaluated at the Dana-Farber Cancer Institute, and each patient had at least one tumor biopsy sequenced with targeted panel sequencing of 400+ cancer related genes. To infer genomic and clinical correlates of site-specific metastasis, we performed cause-specific and Fine-Gray hazard modeling in this cohort. We used the recurrent event Andersen-Gill model to infer predictors of longitudinal metastasis rates. Results: Site-specific analyses yielded genes whose mutant status were significantly associated with a change in risk and/or rate of certain metastatic sites. We found that NOTCH1 is significantly associated with an increased rate and risk of brain metastases (HR=2.46, P<0.001). SETD2 is associated with decreased incidence of brain (HR=0.42, P<0.001). SMARCA4 is significantly associated with increased incidence and rate of adrenal metastasis (HR=1.64, P<0.01), while NF1 is significantly associated with decreased rate of adrenal metastasis (HR=0.72, P<0.01). In a multivariable model, Female patients and younger patients had a lower rate of new metastatic sites. As previously seen, TP53 and SMARCA4 mutations associated with higher rates of new metastatic sites (HR=1.17, P<0.001; HR=1.36, P<0.001 respectively), and SETD2 is associated with a lower rate of new metastatic sites (HR=0.71, P<0.001). Conclusions: By using time-to-event analysis with longitudinal metastatic annotations, we identify potential drivers of overall metastasis and site-specific organotropism in LUAD patients. Identification of new associations and concordance of our results with previous studies also supports the utility of artificial-intelligence-annotated datasets, allowing for larger cohorts without the need for manual annotation. Citation Format: Tyler Aprati, Michael Manos, Giuseppe Tarantino, Marc Glettig, Maryclare Griffin, Alexander Gusev, Kenneth Kehl, David Liu. Leveraging machine-learning approaches to dissect drivers of clinical metastatic dynamics in lung adenocarcinoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 2750.