br Conflicts of interest The authors disclose no conflicts

Conflicts of interest The authors disclose no conflicts.
Funding
This work was supported by the MD Anderson Moonshot Program and the Khalifa Bin Zayed Al-Nahyan Foundation (no grant numbers apply); the National Institutes of Health (U01CA196403 and U01CA200468 to Anirban Maitra); the Cancer Prevention Research Institute of Texas (RP160517 to Anirban Maitra; Vincent Bernard and Nabiollah Kamyabi were funded through a fellowship from Cancer Prevention Research Institute of Texas Research Training Program RP170067; and the German Research Foundation (SE-2616/2-1 to Alexander Semaan). We are grateful the Precision Medicine Research Associates/Fox Family Foundation for partially supporting the studies described in this manuscript.
Supplementary Methods
ddPCR Analysis
ddPCR (QX200; BioRad) was used for highly sensitive detection of genetic mutations with a multiplex KRAS assay containing G12V, G12D, G12R, G12C, G12S, G12A, and G13D mutant codons. Estimation of false positive rate was first determined across multiple wells containing KRAS wild-type DNA from a healthy individual and a nontemplate control. A cutoff of more than 2 droplets in the mutant channel was determined to be optimal for providing no false positive rate and classifying a sample as having mutant molecules. A lower limit of detection was then determined as previously described at 0.01% MAF.1 Subsequently, for each experi-ment done on clinical samples, wells containing a positive control and 2 negative controls were included to determine the absence of contamination and PCR efficiency of the ddPCR probes in each plate. Positive controls consisted of 1 sample of either pancreatic cell line (Pa04C or Panc1), and the negative controls included a wild-type well of DNA from a healthy individual and a well with just water as a non-template control. Interpretation and analysis of results was done in accordance with BioRad “Rare Diphenylterazine Detection Best Practice Guidelines for Droplet Digital PCR.” Data were processed using QuantaSoft, version 1.6 (BioRad).
Statistical Analysis
Our 2 primary goals were to test the predictive abilities of exoDNA MAF in determining progression and its prog-nostic capability with regard to both progression and outcome. To find our cutoffs, we used receiver operating characteristic analysis to estimate optimal thresholds by maximizing the sensitivity and specificity on each graph through the Youden index. A consensus value was drawn from OS and PFS before being tested against clinical out-comes. P values for our predictive subpopulation were ob-tained through the Fisher test. In addition, posterior probability was evaluated through Bayesian inference. Priors were estimated from frequency of progression in our  Liquid Biopsies in Pancreatic Cancer 118.e1
prognostic population. To assess prognosis, we measured 2 time-to-events, PFS and OS, for our endpoints. PFS was defined as the time from the start of therapy, or date of pathologic diagnosis if the patient does not undergo ther-apy, to the first event (progression according to RECIST 1.1 criteria or death due to any cause). Patients tubulins did not progress were censored at their date of last radiologic follow-up. OS was defined as the time from pathologic diagnosis to event (death due to any cause). After con-firming the proportional hazards assumption using Schoenfeld residuals, the Cox proportional hazards model was used to obtain hazard ratios (e^ybeta) and P values in both univariate and multivariate models for a binary threshold detection of exoDNA and covariates. The Wald and log-rank tests were used to estimate significance, and additional Kaplan–Meier analyses were done on our pri-mary variables of interest. Patients who did not h therapy were not included in the multivariate analysis.
To address distributions of covariates, we used Wil-coxon, Kruskal–Wallis, and linear regressions where appropriate. Continuous variables were transformed with a log10(x þ 1) scale to reduce variance. All statistical ana-lyses were done through R version 3.4.22 and Prism 6 (GraphPad Software), by using the ggplot2 package3 in both R and Prism 6 for graphical display. The survival package4 in R was used to create the Cox proportional hazards models.
References
1. Allenson K, Castillo J, San Lucas FA, et al. High preva-lence of mutant KRAS in circulating exosome-derived DNA from early-stage pancreatic cancer patients. Ann Oncol 2017;28:741–747.
2. R Core Team. R: A language and environment for sta-tistical computing. https://www.R-project.org/; 2017.
3. Wickham H. ggplot2: elegant graphics for data analysis.
4. Therneau TM, Grambsch PM. Modeling survival data: extending the Cox model. New York: Springer, 2000.