Abstract 401: Comparison of molecular response calculations for prediction of patient outcome

Abstract

Background: Molecular response (MR) estimated as a change in circulating tumor (ctDNA) load between an early on treatment sample (usually 2-9 weeks post treatment start) and pre-treatment baseline has been shown to predict patient response and outcomes across solid tumors and therapy types in many retrospective studies. There is no consensus, however, regarding the best method for assessing molecular response. Therefore, we aimed to assess several molecular response calculations and determine the optimal method for predicting outcomes in individual advanced cancer patients. Method: Aggregate results of > 4,000 patient sample pairs (3-10 weeks apart), > 1000 patient sample technical replicates, > 100 contrived sample dilutions, and in silico simulations were analyzed with Guardant360TM or GuardantOMNITM. Baseline and on-treatment paired patient samples were collected from advanced cancer patients with over 12 tumor types, including lung, colon, and breast. MR calculations included variant allele fractions (VAFs) of somatic SNVs, indels and fusions. Methods including Ratio treatment/baseline (R) of Maximum VAF (maxVAF), Ratio of Mean VAF (mVAF), and Mean of VAF Ratios were compared, with consideration of VAF precision. Analytical accuracy, reproducibility and limit of detection (LoD) were assessed. Results: Comparison of methods for calculating net change in ctDNA load on > 1500 sample pairs showed high correlation (ρ ranged from 0.93 to 0.98) and categorical agreement split by the median (93%). Therefore selecting an optimal method based on outcome prediction would require prohibitively large patient cohorts. Analytical evaluation and in silico simulations can predict the behavior of each method. Simulations of changes in tumor fraction of real pre-treatment samples found that RmVAF or RmaxVAF are more accurate than mVAFR, which can be skewed by low VAF ratios. Almost 25% of sample pairs have a tumor driver or resistance mutation that is not the maxVAF, suggesting tumor dynamics are better captured by mVAF than maxVAF. Newly-detected on-treatment variants can be an important signal of rising ctDNA levels, impacting MR in approximately 2% of sample pairs. Importantly, MR accuracy for all methods decreases as maxVAF approaches or falls below the variant LoD, due to both stochastic detection and higher CV of variants at low VAF. Thus the assay variant LoD is a key determinant of the fraction of patients who can receive MR evaluation. Technical replicates identified the variant criteria at which a 50% change in tumor fraction differs significantly from technical variation, and could define analytical reporting limits. Conclusions: Comparison of MR methods in a large set of patient samples and simulations supports RmVAF with inclusion of newly-detected mutations. Clinical validation of these methods will support patient-specific MR prediction of outcomes. Citation Format: Allysia J. Mak, Katie J. Quinn, Carin Espenschied, Kyle Chang, Han-Yu Chuang, Elena Helman, Darya Chudova, Justin Odegaard, Becky Nagy, AmirAli Talasaz. Comparison of molecular response calculations for prediction of patient outcome [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 401.