Clinical and analytical validation of MI Cancer Seek^®, a companion diagnostic whole exome and whole transcriptome sequencing-based comprehensive molecular profiling assay

Clinical validation for FDA companion diagnostic (CDx) and tumor profiling claims

Non-inferiority studies for CDx claims

MI Cancer Seek has eight CDx claims representing high clinical burden in the oncology patient population (Table 1, Supplementary Table 1). As a follow-on CDx (FCD) device, clinical validity was demonstrated through non-inferiority to an appropriate FDA-approved test (Supplementary Table 2), as described by Li et al. [28]. One replicate tested by MI Cancer Seek (FCD replicate) was compared to two replicates tested by each comparator companion diagnostic (CCD) test for each CDx claim, as shown in Figure 2. The results for FCD comparison to concordant CCD replicates are summarized in Table 2, and full concordance results are shown in Supplementary Tables 3–5. Regarding PIK3CA alterations in breast cancer, KRAS/NRAS wild-type status in advanced CRC, BRAF^V600E mutations in CRC, BRAF^V600E or BRAF^V600K mutations in melanoma, EGFR exon19 deletions and L858R mutations in NSCLC, and MSI status in endometrial carcinoma and solid tumors, positive percent agreement (PPA) and negative percent agreement (NPA) ranged from 97–100%, and all upper limits of the 95% confidence intervals met pre-specified non-inferiority margins, as described in ‘Methods’.

Figure 2: Schematic of MI Cancer Seek validation for CDx claims.

Remnant samples were enrolled meeting study criteria. Samples were de-identified and lab personnel blinded to their biomarker status. Unstained slides (USS) were used to complete two tests with the comparator companion diagnostic (CCD) and one test with MI Cancer Seek as the follow-on companion diagnostic (FCD). Concordance was evaluated by determining the agreement between the two CCD replicates to the agreements of FCD to CCD1 and FCD to CCD2. Concordance with FCD was also determined using the concordance between both CCD replicates as “truth”. Non-inferiority was the primary endpoint for analysis.

Table 2: Summary of PPA and NPA for MI Cancer Seek CDx clinical concordance studies

Biomarker and indication	FDA-approved Comparator Method	Sample Size, Na	PPA (95% CI)	NPA (95% CI)
MSI Status in Solid Tumor Types	Ventana MMR RxDx Panel	401	97.5% (94, 99.1)	98.5% (95.4, 99.7)
MSI Status in EC	Ventana MMR RxDx Panel	251	98.4% (93.8, 99.9)	97.6% (92.8, 99.5)
BRAFV600E/K in Melanoma	bioMérieux THxID BRAF Kit	330	98.9% (95.6, 99.9)	99.3% (95.9, 100.2)
BRAFV600E in CRC	therascreen BRAF V600E RGQ PCR Kit	352	99.4% (96.5, 100.2)	100% (97.3, 100.4)
KRAS and NRAS wild-type in CRC	Praxis Extended RAS Panel	262	100.0% (96.1, 100.6)	97.2% (92.7, 99.1)
EGFR exon 19 deletions or L858R mutations in NSCLC	Roche cobas EGFR Mutation Test V2	315	98.1% (91.4, 99.6)	99.4% (96.1, 100.2)
PIK3CA alterations in BC	therascreen PIK3CA RGQ PCR Kit	343	99.4% (96.4, 100.2)	100.0% (97.2, 100.4)

^aIncludes samples with complete records for analysis. There were invalid results on either CCD or FCD for BRAF^V600E/K in melanoma (n = 4), BRAF^V600E in CRC (n = 5), KRAS and NRAS wild-type in CRC (n = 24), EGFR exon 19 deletions or L858R mutations in NSCLC (n = 1), and PIK3CA alterations in BC (n = 5). Invalids were excluded for overall concordance results calculated using CCD1 and CCD2 consensus calls as a reference, and additional sensitivity analyses were performed to evaluate non-inferiority, including sensitivity analyses for prevalence and invalids where applicable (data not shown). Abbreviations: BC: breast carcinoma; CRC: colorectal adenocarcinoma; EC: endometrial carcinoma; MSI: microsatellite instability; NPA: negative percent agreement; NSCLC: non-small cell lung cancer; PPA: positive percent agreement.

Tumor profiling validation

MI Cancer Seek provides tumor profiling for use by qualified health professionals in accordance with professional oncology guidelines. Reportable genes with boosted bait coverage are listed in Supplementary Table 6. In order to validate the detection of single nucleotide variants (SNVs) and insertions/deletions (INDELs), MI Cancer Seek results were compared to PGDx elio™ tissue complete, an FDA-cleared targeted gene panel assay [29, 30]. Among 454 eligible specimens, PPA was ≥95% and NPA was 100% for all variant types (Figure 3A, 3B). These metrics demonstrated strong concordance of MI Cancer Seek with the comparator assay. Where discordance was noted, this was determined primarily to be due to inherent differences between the two assays with regards to percent variant frequency (VF) cutoff and variant calling rules, and not due to true discordance between MI Cancer Seek and PGDx elio tissue complete. The ability of MI Cancer Seek to detect ERBB2 (HER2) copy number amplifications (CNAs) in breast cancer was also analyzed. MI Cancer Seek results for ERBB2 CNAs were compared to results of the PathVysion HER2 DNA Probe Kit, yielding 84.7% PPA and 99.4% NPA (Figure 3B). Due to the lower PPA for this comparison, additional clinical investigation to confirm the presence of ERBB2 CNAs with another FDA-approved or cleared test is strongly recommended.

Figure 3: Tumor profiling validations.

(A) Validation of variant detection by MI Cancer Seek compared to PGDx elio tissue complete and ERBB2 copy number amplification compared to fluorescent in situ hybridization (FISH). Numbers of true positive (TP), true negative (TN), false positive (FP), and false negative (FN) are shown. One intermediate sample was counted as positive for ERBB2. One sample that was invalid by MCS and five samples that were invalid by FISH were excluded for ERBB2. (B) Positive percent agreement (PPA) and negative percent agreement (NPA) between MI Cancer Seek and the comparator assay are shown for each variant type. (C) Deming regression analysis of MI Cancer Seek (MCS) TMB values versus the orthogonal test values. Abbreviations: INDEL: insertion/deletion; MNV: multi nucleotide variant; SNV: single nucleotide variant; TMB: tumor mutational burden.

To validate MI Cancer Seek detection of tumor mutational burden (TMB), results from 497 eligible specimens were compared to those from a WES assay (Personalis, Inc., Menlo Park, CA, USA). Deming regression analysis demonstrated a strong correlation between MI Cancer Seek and the Personalis, Inc. WES assay (Pearson r = 0.990; slope = 0.856; intercept = −0.859) (Figure 3C). When comparing TMB using only a subset of exome regions where both assays had ≥50× depth, the concordance between MCS and the orthogonal assay improved even further (slope = 0.927; intercept = −0.68).

Analytical sensitivity of MI Cancer Seek

Analytical sensitivity of MI Cancer Seek was evaluated through limit of detection (LoD) studies to identify the lowest VF (SNVs, INDELs) and lowest tumor content at which at least 95% of the test replicates produced a positive result. CDx biomarkers (including BRAF, EGFR, PIK3CA, and RAS mutations) and other tumor profiling biomarkers (including clinically relevant alterations and TMB) were represented in marker positive samples diluted with lineage matched marker negative samples. Observed VF was concordant with the target VF for each biomarker and reproducible across replicates (Supplementary Figure 1). The results of LoD studies support the recommended sensitivity of variant detection at >5% VF for MI Cancer Seek, illustrated by reduced detection rates at 5% VF for certain mutations (Figure 4A). LoD for ERBB2 CNAs was determined to be 8.3 copies, which was the lowest level of copies yielding ≥95% detection (Figure 4B). Furthermore, tumor content LoD studies for MSI and TMB found reliable detection of these biomarkers at 20% cutoff (Figure 4C, 4D), which exceeded the tumor content LoD for other alterations (15% for ERBB2 CNAs, 10.5% for all SNVs and INDELs, and 12.0% for CDx variants) (Supplementary Table 7). Thus, the 20% tumor content requirement for MI Cancer Seek is considered sufficient to enable reliable detection of all variants/biomarkers.

Figure 4: Limit of detection (LoD).

(A) LoD for variant frequency (VF) of SNVs and INDELs (including CDx biomarkers). Caris cutoff is 5% VF for reporting. (B) Copies LoD for ERBB2 CNA. (C, D) LoD for % tumor content (TC) including MSI (C), and TMB (D). Hit rate (%) at each given VF (A), nReads (B) or tumor content (C) is shown by the heatmap color, with the quantitative hit rate (%) shown in each box. The coefficient of variation (% CV) is shown for TMB in (D). The red boxes highlight the LoD cutoff employed by MI Cancer Seek. In (C, D), the quantitative values for MSI (C) and TMB (D) are indicated in parentheses. Abbreviations: BC: breast carcinoma; CNA: copy number amplification; CRC: colorectal carcinoma; EC: endometrial carcinoma; INDEL: insertion/deletion; Mel: melanoma; MSI: microsatellite instability; NSCLC: non-small cell lung cancer; OSEC: ovarian surface epithelial carcinoma; PC: prostate carcinoma; TC: tumor content; SNV: single nucleotide variant; VF: variant frequency.

A limit of blank (LoB) study established the false positive rate using non-tumor wild-type (WT) samples across multiple tissue types. The results demonstrate a false positive rate of 0% for most DNA capabilities (INDELs, MSI, TMB) and <0.01% for SNVs (Supplementary Table 8). Together, there were a total of three false positive calls for TERTc.-124C>T, which were found to have VF near the variant calling threshold for MI Cancer Seek.

Finally, we performed a DNA input study to investigate the performance of MI Cancer Seek at minimum (50 ng) and optimal/maximum (220 ng) DNA input, as well as suboptimal and overloading levels (25 ng, 37.5 ng, 275 ng, and 330 ng). The PPA and NPA for SNVs, INDELs, MSI, and ERBB2 CNAs were 100% between minimum and optimal DNA input (Table 3). At other input levels, PPA and NPA were also 100% aside from one false negative call for MSI at 37.5 ng. In addition, mean TMB was consistent between 50 ng and 220 ng: the upper bound of the 95% confidence interval (CI) for absolute percent difference in mean TMB values between control and challenge input ranged from 0.0–10.5% (Table 3 and Supplementary Figure 2).

Analytical specificity of MI Cancer Seek

We next evaluated analytical specificity through several studies including interfering substances (exogenous and endogenous), carry-over, and cross-contamination. Exogenous substances included paraffin, xylene, Proteinase K, and 80% ethanol, spiked at the appropriate step to mimic carryover of the substance being tested. Endogenous substances included hemoglobin, colloid, calcium/calcium phosphate, mucin, and conjugated bile acids. The impact of necrotic tissue, melanin, and fat cells on assay performance will be evaluated in a post-market study, although some of these are mitigated whenever possible through microdissection of tumor tissue. There was minimal or no effect detected for exogenous or endogenous substances on assay performance in intended tumor types (Supplementary Table 9). Likewise, carry-over (run-to-run) and cross-contamination (within run) were evaluated at the highest DNA input of 220 ng, which demonstrated 100% PPA and NPA for global mutations, SNVs, INDELs, and MSI. Further specificity studies showed that for any sample, there was a <0.004% chance of index cross-contamination (misassignment of reads to a sample due to both sample indexes being incorrectly switched to another sample). An in silico analysis demonstrated low off-target depth of coverage (mean 1.5×), which confirmed capture bait specificity for reportable gene regions.

Precision of MI Cancer Seek

Next, we performed a precision analysis at low (1–1.5×) and high (2–3×) LoD levels to evaluate the total within-lab variability expected in the MI Cancer Seek assay by evaluating both the individual and combined impact of reagent lots, instruments, operators, and non-consecutive run days across a panel of samples containing representative types of genetic alterations detected by the test. For SNVs, multi-nucleotide variants (MNVs), and INDELs, precision increased according to variant buckets (≥5%, ≥8%, ≥10%, and ≥15% VF), but exceeded 96% even at the lowest VF bucket (Figure 5A). Sample level precision analysis also revealed overall high positive and negative call rates at high and low LoD, although some samples had reduced positive call rates at low LoD (Figure 5B). All CDx markers achieved 100% agreement across replicates except for three cases that had borderline MSI status resulting in agreements of 97.2%, 91.7%, and 94.4%. Coefficient of variation (CV), which is the ratio of the standard deviation to the mean and is a measure of assay precision, was determined for MSS/MSI-high (measured by quantitative loci) and TMB (measured across a range of values from 2.3–30.2), incorporating operator/instrument, reagent lot, within-run, and between-run variability (Figure 5C–5F). Together, the results of these analyses support the precision of MI Cancer Seek for tumor profiling and CDx biomarker detection at LoD thresholds.

Figure 5: Precision study.

There were 49 panel members in the precision study. Each panel member was tested a total of 36 times, including three operator teams, three instrument sets, three reagent lots, and was minimally executed across three non-consecutive days over a 20-day span. (A) Precision per variant frequency (VF) bucket for INDELs, MNVs, and SNVs. Variants included passenger variants that were not the basis of panel member selection. (B) Sample-level positive and negative calls rates for tracked variants in all 49 panel members in precision study. There were 1–19 variants per sample. 37 panel members were tested at high (2–3×) and low (1–1.5×) LoD. One additional panel member was tested only at high LoD and one additional panel member only at low LoD. The remaining panel members did not have variant frequency LoD adjusted (two panel members were MSS, four panel members were MSI-high, two panel members were TMB-high, and two panel members had an ERBB2 CNA). (C) Mean MSI loci (n = 6) is shown with total within-lab standard deviation, including operator/instrument, reagent lot, within-run, and between-run variability. (D) %CV for MSI loci is shown. (E) Mean TMB (n = 49) is shown with total within-lab standard deviation, including operator/instrument, reagent lot, within-run, and between-run variability. TMB was measured in all 86 samples (including panel members tested at high and low LoD). (F) %CV for TMB is shown. For (D) and (F), the red dotted line indicates acceptable %CV. Abbreviations: CNA: copy number amplification; CV: coefficient of variation; INDEL: insertion/deletion; LoD: limit of detection; MNV: multi-nucleotide variant; MSI-H: microsatellite instability high; MSS: microsatellite stable; SNV: single nucleotide variant; TMB: tumor mutational burden.

Laboratory developed test (LDT) validation

MI Cancer Seek is the FDA-approved version of its LDT predecessor, MI Tumor Seek Hybrid, which provides additional information relevant to indicated tumor types. While findings other than those listed in Table 1 for MI Cancer Seek are not prescriptive or conclusive for labeled use of any specific therapeutic product, these capabilities have been validated internally and meet CAP/CLIA requirements. Therefore, they may be considered at the physician’s discretion when evaluating possible treatments or clinical trials, or for research purposes. These LDT capabilities include the following: copy number variants (CNVs; including 1p19q co-deletion and 7+/10- co-occurrence), karyotype images, structural variants (SVs)/fusions, splice site variants (SSVs; including MET exon 14 skipping, ARv7, EGFRvIII), human leukocyte antigen (HLA) typing, loss of heterozygosity (LoH), homologous recombination deficiency (HRD) for predicting PARP inhibitor response in ovarian cancer, FOLFIRSTai signature for predicting first-line chemotherapy response in metastatic CRC, Genomic Probability Score (GPSai) for predicting tissue of origin in cancers of unknown primary (CUP), cancer-associated virus detection (HPV, EBV, MCPyV), and gene expression. Several of these capabilities have been or will be described in detail elsewhere, including HRD, FOLFIRSTai, and GPSai. The validation of HRD in ovarian cancer has been fully described [31]. Given the evolving landscape and potential clinical applications of this marker, HRD will also be reported for research-use-only on all tumor types, with further validations forthcoming. Development of the FOLFIRSTai 67-gene signature and the original GPSai model are described by Abraham et al. [32–34]. The GPSai model has since been updated and clinically validated, which has been reported separately. When compared to validated internal or external comparator assays (Supplementary Table 10), MI Tumor Seek Hybrid fulfilled acceptance criteria with regards to accuracy and precision for all LDT capabilities (Table 4).

Separate validations were performed in 3,968 samples for ERBB2 amplification in breast cancer as a complement to the MI Cancer Seek tumor profiling validation, and in 9,213 samples for ALK fusions in NSCLC. For this LDT validation, ERBB2 CNA results were compared to results of Ventana’s PATHWAY anti-HER2/neu (4B5) IHC test. These large-scale validations focusing on these clinically relevant alterations [35–38] in their respective tumor types yielded PPAs of 92.1% and 94.3% and NPAs of 99.6% and 99.8%, respectively (Table 4). An additional real-world analysis of MSI status in a cohort of 46,976 solid tumors and 3,488 endometrial tumors also demonstrated high concordance with the Ventana MMR RxDx Panel (Supplementary Table 11), further supporting our CDx claim.

Finally, MI Tumor Seek Hybrid provides gene expression data for panels of genes that are determined on a lineage basis for matching to clinical trials. Transcripts per million (TPM) and the accompanying percentile in the Caris cohort of the tumor type profiled are included in the report. Validation of MI Tumor Seek Hybrid gene expression data required a different approach due to the continuous nature of the data as well as technical differences from the comparator WTS assay, MI Transcriptome. We first determined area under the receiver operating characteristic curve (AUROC) for key genes that could be validated against immunohistochemistry (IHC). MI Tumor Seek Hybrid and MI Transcriptome demonstrated comparable AUROC values relative to pathologist-signed IHC results for androgen receptor (AR), HER2, estrogen receptor (ER), progesterone receptor (PR), and Claudin-18 (AUROC >0.89 for all) (Supplementary Figure 3). We then compared TPM between MI Transcriptome and MI Tumor Seek Hybrid numerically by applying a percentile transformation method, which is described in Supplementary Figure 4. Using this method, there was a strong correlation between MI Transcriptome and MI Tumor Seek Hybrid for a set of frequently reported genes (n = 214) (Supplementary Figure 5). Quantitative gene expression reporting by MI Tumor Seek Hybrid was also determined to be sufficiently precise, especially at TPM ≥25 (Supplementary Figure 6). Together, the gene expression validation studies support this LDT capability and culminate the comprehensive molecular reporting of MI Cancer Seek/MI Tumor Seek Hybrid.

General methodology for validation

All samples, kits, reagents, and supplies were pre-specified and acquired in advance of the validation plan execution. All operators were pre-identified and trained on the workflow. All operators, samples, kit lot numbers, reagent lot numbers, expiration dates, instrument serial numbers, run dates, and validation results were documented. For any failures, inconsistencies, and/or questions, the technical supervisor (a board-certified molecular geneticist), and/or a board-certified pathologist was consulted. The studies were executed over multiple runs, by multiple operators and with different lots of reagents to ensure that testing performance was reproducible across multiple variables. The majority of cases for accuracy studies used an optimal input of DNA (220 ng), unless otherwise indicated. Non-optimal inputs (25, 37.5, 50, 275, and 330 ng) were tested for a DNA input and guard banding study. A summary of the study designs and parameters for all clinical and analytical studies performed is shown in Supplementary Table 13. Samples for the accuracy analysis underwent orthogonal testing using other Caris in-house assays or were tested using commercial assays by qualified laboratories (Supplementary Tables 2, 10).

Samples

A total of 2,063 unique samples, representative of 48 tumor lineages, were enrolled in the clinical validation studies. For clinical validation studies, biomarker positive samples were enrolled consecutively starting from the most recent case moving backward, and an equal number of date-matched biomarker negative samples were enrolled. Samples for the KRAS/NRAS study were enrolled consecutively regardless of biomarker status, since this study did not require biomarker enrichment. Samples were excluded from testing if they did not meet minimum requirements for MI Cancer Seek or the comparator assay. Samples with invalid results on MI Cancer Seek or the comparator assays were excluded from the primary analysis but were analyzed in best- and worst-case sensitivity analyses. For clinical validation studies, blinding was performed so that no internal or external operator was aware of the biomarker status of the tested sample. The demographics of the cohorts (age and sex) were comparable to the cohorts from the pivotal clinical trials investigating each biomarker. Samples for analytical validation studies shown here (e.g., DNA input, LoD, precision, interfering substances, LoB, cross-contamination, carry-over) were selected based on sample biomarker status and where applicable were diluted with normal tissue samples or other cancer samples as indicated to achieve designated DNA input levels or tumor purity, VF, etc. according to prespecified study designs. Additional samples were used for the tumor profiling validation, analytical validation, and LDT validation studies (Supplementary Table 13). Positive and negative controls specific to each capability of the assay were included in this study and are included for each MI Cancer Seek run.

Total nucleic acid (TNA) extraction

FFPE slides underwent review by a board-certified pathologist to measure tumor content. For accuracy analyses, a minimum of 20% tumor content across 25 mm² area was required, with manual microdissection performed to meet requirements if necessary. Microdissection was also performed as needed to exclude interfering substances as much as possible (melanin, adipose, necrosis). TNA was auto-extracted using a MagMax FFPE DNA/RNA Ultra extraction kit (Thermo Fisher Scientific, Waltham, MA, USA) on a Biomek i7 Automated Workstation (Beckman Coulter, Brea, CA, USA) and KingFisher Flex (Thermo Fisher Scientific). DNA quantification was performed using a Quant-iT 1X dsDNA High Sensitivity Assay Kit (Thermo Fisher Scientific) and read on a GloMax^® Microplate Reader (Promega, Madison, WI, USA). Samples were prepared for quantification, normalized, and plated for downstream processes using a Biomek i7 Automated Workstation. RNA quantification was not required, since starting DNA and RNA molecules were converted into libraries for NGS in the same hybrid workflow. The RNA/DNA ratio in TNA from tissue was always >1; thus, the input into library preparation was controlled by limiting material (DNA).

MI Cancer Seek library preparation and next-generation sequencing

Library preparation was automated on the Bravo B Liquid Handler (Agilent, Santa Clara, CA, USA) using KAPA NGS Library preparation reagents (Roche Diagnostics, Indianapolis, IN, USA) and custom cDNA primers (IDT, Coralville, IA, USA; GeneLink, Elmsford, NY, USA). NGS library preparation is a fully customized workflow, referred to as a ‘hybrid’ workflow, where starting TNA molecules proceed through the process together. DNA and RNA underwent fragmentation, followed by cDNA synthesis from RNA, end repair and A-tailing of cDNA and original DNA, ligation of universal adapters, pre-capture PCR, and addition of P5/P7 indexes. Genomic targets were enriched through hybridization with a custom bait panel and captured using 120 nucleotides long, double-stranded, biotinylated baits. The bait panel was customized to add additional coverage for clinically relevant and reportable genes (Supplementary Table 6). Capture and purification of enriched DNA targets was performed following hybridization of dsDNA probes (baits) to adapter-ligated DNA. Post-capture PCR was performed on enriched DNA targets, and the resulting library was quantified using a Quant-iT 1X dsDNA High Sensitivity Assay Kit. Finally, normalization and pooling were performed, with quantification of the final library pool by a Qubit 1X dsDNA High Sensitivity Assay Kit (Thermo Fisher Scientific). Sequencing was performed on a NovaSeq 6000 instrument (Illumina, San Diego, CA, USA), using recommended reagents. A Veriti Thermocycler (Thermo Fisher Scientific) was used for all PCR steps. Quality control elements were checked at applicable steps. RNA was chemically labeled during library preparation, which enables separating RNA-originated data during bioinformatics data analysis. The MI Cancer Seek workflow is shown in Figure 1.

MI Cancer Seek bioinformatic pipeline

Sequencing data were extracted into split FASTQ files (RNA and DNA) for further processing through Caris’ proprietary bioinformatics pipeline. For RNA, Spliced Transcripts Alignment to a Reference (STAR) software was used for alignment, and STAR-Fusion was employed for fusion detection [59]. Transcripts per million (TPM) molecules were generated using the Salmon expression pipeline [60]. DNA variants detected were mapped to reference genome hg38, and well-established bioinformatics tools such as BWA, Samtools, Pindel, and snpEff were incorporated to perform variant calling functions; germline variants were filtered with various germline databases, including dbSNP. CNVs were detected using CNVkit [61]. Genetic variants identified were interpreted by board-certified molecular geneticists and annotated as ‘pathogenic,’ ‘likely pathogenic,’ ‘variant of unknown significance,’ ‘likely benign,’ or ‘benign,’ according to ACMG standards.

Tumor signatures

The threshold to determine MSI-high status was 39 or more loci with frameshift mutations out of 5,721 loci examined. TMB was measured by counting all non-synonymous missense, nonsense, in-frame insertion/deletion, and frameshift mutations found per tumor that had not been previously described as germline alterations in dbSNP151, Genome Aggregation Database (gnomAD, AC>0) databases, or benign variants identified by Caris geneticists. A cutoff point of ≥10 mutations per MB was used based on the KEYNOTE-158 pembrolizumab trial [62].

LoH was determined by calculating the LoH of single nucleotide polymorphisms (SNPs) within 552 chromosome segments of 2–6 Mb in length. 250K SNPs were analyzed, with a minimum of 17 SNPs per Mb of genome sequence. The final call of genomic LoH was based on the percentage of all 552 segments with observed LoH (High ≥16%; Equivocal ≥11% and ≤15%; Low <11%; Indeterminate ≤3,000 SNPs read or sample depth <200×). A normal epithelial ovarian genome (NA12878) that has no non-polymorphic variants, gene fusions or other cancer hallmarks, was used as a negative control. LoH was also used to determine HRD. The presence of LoH and large-scale state transitions (LST), which are the loss of chromosomal segments larger than 10 Mb, contributed to the genomic scar score (GSS). HRD was called when the sample was positive for BRCA1 or BRCA2 mutation or was genomic scar score (GSS)-high (GSS ≥46), and negative when it did not meet those criteria. HRD was restricted to ovarian cancer cases.

Statistical analysis

Sample sizes for CDx studies were targeted to achieve at least 80% power to rule out non-inferiority, accounting for projected invalid rates based on historical data. The agreement between two replicates of the comparator companion diagnostic (CCD1 and CCD2) was compared to the agreements of the follow-on companion diagnostic (FCD) to CCD1 and FCD to CCD2, according to Li et al. where a reference standard is not available and the concordance study sample is not a random sample from the MI Cancer Seek intended use population (with the exception of KRAS/NRAS, since no enrichment was required) [28]. Concordance results were used to calculate agreement values (NPA and PPA) and differences (𝜁) in those values. Differences in agreements were calculated according to the following formulas: 𝜁PPA1 = PPA _C1C2 − PPA_C1F; 𝜁PPA2 = PPA _C2C1 − PPA_C2F; 𝜁NPA1 = NPA _C1C2 − NPA_C1F; and 𝜁NPA2 = NPA _C2C1 − NPA_C2F. NPA/PPA_C1C2 is the proportion of CCD1 negative/positive results in which CCD2 is negative/positive. NPA/PPA_C1F is the proportion of CCD1 negative/positive results in which FCD is negative/positive. NPA/PPA_C2C1 is the proportion of CCD2 negative/positive results in which CCD1 is negative/positive. NPA/PPA_C2F is the proportion of CCD2 negative/positive results in which FCD is negative/positive. Prevalence-adjusted agreement values (i.e. adjusted for enrichment of the sample population) were calculated using Bayes’ Theorem and the Law of Total Probability.

FCD was considered non-inferior to CCD by a margin of ε, where ε is the maximum upper limit of the 95% CIs for each 𝜁, and δ is the target non-inferiority margin. ε1 is the true non-inferiority margin for FCD as a predictive test, ε0 is the true non-inferiority margin for FCD as a selective test, δ1 is the target non-inferiority margin for FCD as a predictive test, and δ0 is the target non-inferiority margin for FCD as a selective test. If ε <δ, FCD achieved the targeted non-inferiority level. Two-sided 95% confidence intervals (CIs) for agreement differences (𝜁) were calculated using the percentile bootstrap method with 1000 bootstrap samples for all studies except KRAS/NRAS, which used the continuity-corrected Wilson interval for paired data, as described in method 10 of Newcombe et al. [63]. Sensitivity analyses were performed to evaluate the impact of prevalence using an upper and lower bound based on values reported in the literature. Standard approaches in JMP Version 17 or higher, SAS Version 9.4 or higher, or R version 4.3.0 or higher were used for data analysis.