a, Tumor and cfDNA samples were collected from patients with MBC, NSCLC and CRPC. Tumor and matched normal samples were sequenced using the MSK-IMPACT assay, while plasma and buffy coat samples from patients with cancer and from controls without cancer from the San Diego Blood Bank underwent sequencing followed by de novo assembly and mutation detection using the high-intensity targeted cfDNA assay by GRAIL, based on a bespoke joint-variant-calling pipeline. Tumor and cfDNA somatic variant detection results were unblinded for concordance analyses. UMI, unique molecular identifier. b, Analytical performance of the targeted DNA assay. The detection probability is shown as a function of increasing VAF in HD753 cell line DNA titrations. The curves correspond to the mean target coverage of 2,430× from 30 ng cell line DNA input and the mean target coverage of 4,577× obtained from simulated FASTQs. Shading represents 95% confidence intervals. c, Estimated variant calling specificity using samples from controls without cancer and corresponding variant calling sensitivity using the methods described in the Supplementary Methods (joint variant analysis using the machine learning (ML) error model). Controls without cancer were not used to train the model here. AF, allele fraction. d, Comparison of allele fractions of variants detected using either of the two targeted DNA assay protocols in five patients. e, One patient with hypermutation and MBC was excluded from this analysis to avoid biased regression. Concordant mutation detection between the two replicates (triangles indicate biopsy-matched variants; circles indicate biopsy-unmatched variants) is enriched in allele fraction above the limit of detection. f, Comparison of VAFs measured using the targeted DNA assay (y axis) and ddPCR (x axis). cfDNA extracted from five patients with cancer with canonical hotspot mutations was subjected to ddPCR. An aliquot of the same cfDNA sample was employed for the targeted DNA assays using two versions of the protocol (version 1 and version 2). One sample lacking canonical hotspot mutation in the ddPCR measurements was excluded.