Purpose: In metastatic colorectal cancer, the detection of RAS mutations by circulating tumor DNA (ctDNA) has emerged as a valid and noninvasive alternative approach to determining RAS status. However, some RAS mutations may be missed, that is, false negatives can occur, possibly compromising important treatment decisions. We propose a statistical model to assess the probability of false negatives when performing ctDNA testing for RAS.
Methods: Cohorts of 172 subjects with tissue and multipanel ctDNA testing from MD Anderson Cancer Center and 146 subjects from Massachusetts General Hospital were collected. We developed a Bayesian model that uses observed frequencies of reference mutations (the maximum of APC and TP53) to provide information about the probability of KRAS false negatives. The model was alternatively trained on one cohort and tested on the other. All data were collected on Guardant assays.
Results: The model suggests that negative KRAS findings are believable when the maximum of APC and TP53 frequencies is at least 8% (corresponding posterior probability of false negative <5%). Validation studies demonstrated the ability of our tool to discriminate between false-negative and true-negative subjects. Simulations further confirmed the utility of the proposed approach.
Conclusion: We suggest clinicians use the tool to more precisely quantify KRAS false-negative ctDNA results when at least one of the reference mutations (APC, TP53) is observed; usage may be especially important for subjects with a maximum reference frequency of <8%. Extension of the methodology to predict false negatives of other genes is possible. Additional reference genes can also be considered. Use of personal training data sets is supported. An open-source R Shiny application is available for public use.