Issue	Findings	Outlook & recommendations
Mutation frequency	Mutations present above ~0.5% VAF were detected with high sensitivity and reproducibility by all participating ctDNA assays but performances were generally suboptimal below this level and variable between assays (Fig. 4).	A key challenge in ongoing development of ctDNA sequencing assays is to improve detection sensitivity for low-frequency mutations (< 0.5% VAF).
Coverage depth & heterogeneity	Fragment-depth was a critical variable in ctDNA assays, with high coverage essential for sensitive detection of low-frequency mutations (Fig. 3,4). In addition to depth, even coverage across target regions was important to ensure high sensitivity and reproducibility (Fig. S3).	Improvements to the efficiency/stability of capture enrichment, NGS library conversion and amplification may yield increased coverage depth and decreased heterogeneity, leading to improved performance and robustness.
DNA input quantity	Increasing DNA input quantity generally improved fragment-depth, sensitivity and reproducibility (Fig. 5).	Limited availability of cell-free DNA is a challenge for clinical translation of ctDNA assays. Input material may be increased via improvements to efficiency of plasma-DNA extractions, increasing the volume of patient blood draws (when feasible) or obtaining cell-free DNA from other body fluids (e.g., urine, stool, CSF, etc.) for relevant cancers.
UMIs	Unique molecular identifiers (UMIs) enabled effective consensus error correction, minimizing the detection of false-positives (Table 2).	Wherever possible, UMIs should be employed for consensus error correction in ctDNA sequencing assays.
Inter-laboratory variation	Participating assays were robust to technical variables between test labs – from plasma extraction to sequencing workflow stages – and were impacted largely by random, rather than systematic variation (Fig. 4, Fig. S6).	Robustness to technical variables is essential for clinical implementation of ctDNA sequencing assays.
Random sampling	The detection of low-frequency mutations (VAF < 0.5%) by random sampling poses an inherent statistical challenge, even when high fragment-depth is available (Fig. 1).	Novel strategies for the enrichment of ctDNA fragments over non-cancerous cell-free DNA (e.g., by fragment size selection) and alternative signals, such as ctDNA methylation profiles, may help overcome limitations of random sampling.
Targeted enrichment method	Performance was broadly comparable between participating amplicon and hybrid-capture assays, with sensitivity and robustness largely determined by the fragment-depth achieved, not the method of enrichment (Fig. 6).	Amplicon methods can enable sensitive, cost effective detection of ctDNA mutations in single genes or mutation hotspots but small panel sizes limit their suitability for unbiased surveillance (e.g., for tumor evolution profiling).
Exon edge-effect	In hybrid-capture sequencing, mutations in exon edge regions were detected with lower sensitivity than central regions, due to lower coverage (Fig. 1,2).	Increasing the size of captured flanking regions around exons during panel design may alleviate this exon edge-effect.
Sequence context	Mutations in challenging genome sequence contexts, such as high/low GC-content, low sequence complexity or suboptimal alignability, were detected with lower sensitivity (Fig. 1,2).	Some of these effects may be alleviated by increasing capture-probe density in challenging regions or via improvements to NGS library preparations.
Reference standards	Reproducibility measurements provided a useful but imperfect proxy for analytical performance that is not dependent on the availability of a reference sample and annotation (Fig. 4).	Well-characterized reference standards can directly measure analytic performance characteristics in absence of confounding biological variables and are a useful tool for comparing ctDNA assays.