Utilizing phenotypic characteristics of metastatic brain tumors to predict the probability of circulating tumor DNA detection from cerebrospinal fluid.

Abstract

2507 Background: Brain metastases occur in approximately 20% of tumor patients and is often associated with terminal events and poor prognosis. Cerebrospinal fluid (CSF) can be a promising source for detecting circulating tumor DNA (ctDNA) specific to the central nervous system (CNS) instead of peripheral blood due to the blood-brain barrier. However, CSF’s suboptimal ctDNA detection rate might limit its clinical application. Precise screening of suitable patients is needed to maximize clinical benefit. Methods: We sequenced 425 cancer-relevant genes in CSF and matched extracranial tissue or blood samples obtained from 67 lung cancer patients with brain metastases. The impact of clinical factors, including age, gender, tumor size, number of lesions, and distance of lesions to the ventricle on CSF ctDNA detection was then evaluated by univariate logistic regression. To predict the probability of successful CSF ctDNA detection, best subsets regression was employed for feature selection and cross validation was used for performance assessment to determine the final model. Results: We detected somatic alterations in 39/67 (58%) CSF ctDNA, 57/66 (86%) plasma ctDNA and 45/49 (92%) tissue samples. Mutation detection rate of CSF ctDNA was significantly lower than that from extracranial tissue and plasma (P < 0.001). Univariate analysis revealed significant association (P < 0.05) of high CSF ctDNA detection rate with the following features: (1) intracranial lesion size ( T), (2) shortest distance between the largest lesion and the ventricle ( Dtop), and (3) shortest distance between all intracranial lesion and the ventricle ( Dall). We also revealed a trend of higher detection rate in patients with CNS symptoms ( SCNS). Subsequent best subsets analysis and cross validation suggested best prediction power with lesion size and largest lesion-ventriclar distance (area under curve [AUC], 0.76 [95% CI, 0.71 to 0.85]; accuracy, 0.75 [95% CI, 0.70 to 0.81]). Final probability can then be derived from Logit P = 0.11×T−0.16×Dall (AUC, 0.82; sensitivity, 0.91; specificity, 0.74). The detection of CSF ctDNA was significantly improved from 58% to 83% (P = 0.03) based on the model. Conclusions: This study established a regression model to predict the probability of CSF ctDNA that can be useful to facilitate clinical decisions and avoid excessive practice when monitoring tumor evolution in the brain.