Circulating tumour DNA and its clinical utility in predicting treatment response or survival in patients with metastatic colorectal cancer: a systematic review and meta-analysis

Abstract

Background

We investigate the current knowledge on circulating tumour DNA (ctDNA) and its clinical utility in predicting outcomes in patients with metastatic colorectal cancer (mCRC).

Methods

PubMed, Embase, Cochrane Database of Systematic Reviews and Cochrane Central Register of Controlled Trials were searched. Last search 16/12/2020. We included studies on patients with mCRC reporting the predictive or prognostic value of ctDNA. We performed separate random-effects meta-analyses to investigate if baseline ctDNA and early changes in ctDNA levels during treatment were associated with survival. The risk of bias was assessed according to the Quality in Prognosis Studies tool.

Results

Seventy-one studies were included with 6930 patients. Twenty-four studies were included in meta-analyses. High baseline ctDNA level was associated with short progression-free survival (PFS) (HR = 2.2; 95% CI 1.8–2.8; n = 509) and overall survival (OS) (HR = 2.4; 95% CI 1.9–3.1; n = 1336). A small or no early decrease in ctDNA levels during treatment was associated with short PFS (HR = 3.0; 95% CI 2.2–4.2; n = 479) and OS (HR = 2.8; 95% CI 2.1–3.9; n = 583). Results on clonal evolution and lead-time were inconsistent. A majority of included studies (n = 50/71) had high risk of bias in at least one domain.

Conclusions

Plasma ctDNA is a strong prognostic biomarker in mCRC. However, true clinical utility is lacking.

Background

Colorectal cancer is the third most common cancer worldwide, with nearly 1.4 million new cases [1] and 829,000 deaths every year [2]. Quality of life and survival in patients with metastatic colorectal cancer (mCRC) have improved considerably during the last 30 years, due to improved surgical and oncological treatment strategies. Patient care pathways have evolved in parallel, and are intended to provide patients and physicians with a framework for treatment and follow-up. Despite this, there is a lack of biological markers in clinical use that can benefit patients and support physicians in deciding when and which treatment to provide, and evaluating if a given treatment has an effect.

Over the last decade, there has been an increasing interest in liquid biopsies as a conceptual approach to meet this need. In theory, liquid biopsies reflect tumour-specific genomic aberrations and may capture intra- and inter-tumour heterogeneity. Both the type of aberration detected and the levels may provide information that can predict treatment response or prognosis upfront. Furthermore, the non-invasive nature of sampling gives an opportunity for repeated measurements throughout the course of treatment, which is usually not feasible with standard biopsy techniques. This introduces the possibility to capture genomic changes over time, information that may further support clinical decisions and improve outcomes.

Circulating tumour DNA (ctDNA) in blood seems to be the most promising liquid biopsy approach in mCRC. The discovery of circulating cell-free DNA (cfDNA) is old [3], but has had a renewed interest in the post-genomic era thanks to improvements in analytical technology and personalised treatment strategies. In the European Society of Medical Oncology (ESMO) 2016 consensus guidelines on mCRC, liquid biopsies were described as an emerging new tool that was predicted to be used therapeutically “in the near future” [4]. However, both the American Society of Clinical Oncology (ASCO) guidelines from 2017 and updated ESMO guidelines from 2018 conclude that this emerging technology cannot yet be recommended in routine practice [5, 6].

The primary objective of this study was to present a systematic review and meta-analysis of ctDNA as a biomarker and its clinical utility in predicting response to treatment or survival in patients with mCRC. This review summarises the current knowledge and provides future perspectives on research in this field.

Methods

Pre-specified eligibility criteria

Eligibility criteria include: (1) human studies on patients with metastatic colorectal adenocarcinoma, (2) studies reporting the value of circulating tumour-specific DNA in plasma or serum and (3) studies involving at least one of three pre-defined clinical settings; (a) patients receiving systemic anti-cancer treatment where ctDNA is associated with a survival endpoint, (b) patients receiving systemic anti-cancer treatment where ctDNA is associated with a response endpoint or (c) patients receiving locoregional treatment for metastases where ctDNA is associated with a survival or response endpoint. The review was registered with PROSPERO (CRD42019125630) and is reported in accordance with the PRISMA statement [7].

Information sources and systematic search strategy

We searched PubMed, Embase, Cochrane Database of Systematic Reviews and Cochrane Central Register of Controlled Trials as of 16/11/2018 to identify relevant studies. Updated searches were performed from 01/01/2018 to 13/11/2019 and from 01/01/2019 to 16/12/2020 to include the most recently published studies. We did not limit our searches by language, year of publication or type of publication. Specific search strategies and strings are specified in Supplementary Methods. Studies were managed using Endnote (Clarivate Analytics, Philadelphia, United States).

Study evaluation and selection

At least two researchers independently evaluated the studies through each phase of the review (i.e., screening, eligibility and inclusion) (LC, JH, AB). Study screening for eligibility and data collection of relevant studies was performed using a pre-defined data extraction form modified from Cochrane (London, England). All studies were initially screened by title and abstract, followed by a full-text assessment. The consensus was reached by discussion. At least two researchers (LC, JH, AB) tabulated core data items from included studies. We collected data on the study (author, year, design), the number of patients, treatment (line, type), ctDNA marker/analytical method, evaluated cut-off(s), treatment response, hazard ratio (HR) for progression-free survival (PFS) and overall survival (OS) including 95% confidence interval (CI) and p-value, and lead-time.

Risk of bias

The risk of bias was assessed independently by two researchers (LC, JH, AB) according to the Quality in Prognosis Studies (QUIPS) tool [8]. Specifically, this included biases in the domains of study participation; study attrition; biomarker measurement; outcome measurement; study confounding; statistical analysis and reporting. Each domain was scored according to the categories of low, moderate or high risk of bias. The consensus was reached by discussion (with a third researcher when necessary). The overall risk-of-bias judgement across domains was not done. The workflow is illustrated in Supplementary Fig. 1.

Summary and synthesis of results

To illustrate the different aspects of the clinical utility of ctDNA as a biomarker in patients with mCRC this issue was approached from different clinically relevant angles as described below (Fig. 1).

Fig. 1. Schematic overview of clinical applications of liquid biopsies in patients with metastatic colorectal cancer during systemic anti-cancer treatment.

PFS progression-free survival, OS overall survival, ctDNA circulating tumour DNA, RECIST Response Evaluation Criteria in Solid Tumors PR partial response, PD progressive disease.

Baseline ctDNA level and early ctDNA level dynamics during treatment (a and b—Fig. 1)

Baseline ctDNA level is defined as ctDNA level prior to study treatment. Early dynamics is a change in ctDNA level from baseline to a time point during the first courses of treatment (Fig. 1). The predictive and prognostic values of these parameters were evaluated. Data were summarised in tables. Meta-analyses were performed and data were summarised in forest plots. Studies were ordered according to the publication year.

If multiple results for one outcome (PFS, OS or treatment response) were available from one included study (i.e., different cut-offs, ctDNA marker of interest), all results were included if each result originated from separate study cohorts. When originating from the same study cohort, the results with the most complete data (HR, 95% CI, p-value, n) were included. If the completeness was even, we included the results that applied to most patients. This also applied when selecting results for the meta-analyses.

Lead-time (c—Fig. 1)

Lead-time is defined as the time from molecular biological progression (an increase in ctDNA level or the emergence of new mutations in ctDNA) during systemic treatment to a radiologically verified progression. Lead-time was evaluated and summarised in narrative data synthesis.

Clonal evolution (d—Fig. 1)

Clonal evolution is defined as the emergence of ctDNA mutations during systemic treatment. Clonal evolution was evaluated and summarised in narrative data synthesis.

Metastasectomy

The value of ctDNA in relation to metastasectomy was described briefly in supplementary results.

Statistical analysis

HRs with corresponding 95% CIs reported in the studies were used to estimate the strength of the relationship between the ctDNA marker of interest and survival (PFS, OS). Only studies providing HR based on univariate analysis and corresponding 95% CI were included in the meta-analyses. Due to inter-study heterogeneity, study-specific results were pooled using a random-effects model. Funnel plots were generated to visually assess publication bias.

Heterogeneity was quantified by Chi-squared tests and inconsistency index (I²) tests statistics. A Chi-squared test of p < 0.10 or I² > 50% indicated heterogeneity among studies. All analyses were performed with Stata software version 17.0 (Stata Corporation, College Station, TX). A two-sided p-value below 0.05 was considered statistically significant.

Results

Eligibility assessment

A total of 2363 unique study entries were identified and screened in abstract form according to the pre-defined inclusion and exclusion criteria. Only studies published in full-text were considered for inclusion. Reasons for exclusion by abstract included studies involving multiple cancer types, multiple stages and fewer than five patients in the analysis. Seventy-one articles were included in the final review, fulfilling all inclusion criteria (Supplementary Table 1). The eligibility assessment was summarised in a flow diagram (Fig. 2a). A steady increase in the number of eligible publications was seen from 2010 to 2020 (Fig. 2b). The studies included 6930 patients and reported an association between ctDNA and outcomes relevant to this systematic review in at least 2875 patients.

Fig. 2. PRISMA diagram and eligible studies by year of publication.

PRISMA flow diagram of the study selection process (a). Number of included studies by year of publication (b).

Treatments

Most patients received life-prolonging systemic therapy in the form of 5-fluorouracil, oxaliplatin and/or irinotecan-based chemotherapy with or without an epidermal growth factor receptor inhibitor (anti-EGFR) or vascular endothelial growth factor inhibitor (anti-VEGF). A few patients received anti-EGFR monotherapy, trifluridin/tipiracil, regorafenib, a BRAF proto-oncogene inhibitor, mitogen-activated protein kinase inhibitor, tyrosine kinase inhibitor, human epidermal growth factor receptor 2 inhibitor, or immune checkpoint inhibitors in various combinations (with or without chemotherapy).

Risk of bias

The risk of bias assessment revealed that no studies had a low risk of bias in all six domains. A majority of studies (50/71) had a high risk of bias in at least one domain. A high risk of bias was most common in the domain ‘Statistical analysis and reporting’. Low risk of bias was most common in ‘Outcome measurement’ and ‘Study Attrition’ (Fig. 3, Supplementary Table 2). There was a tendency for a lower risk of bias in studies included in the meta-analyses (Supplementary Fig. 2).

Fig. 3. Assessment of risk of bias using QUIPS (Quality in Prognosis Studies) tool.

The authors’ judgements regarding each risk-of-bias item presented as pecentages across all included studies.

Baseline ctDNA levels and treatment response/survival (a—Fig. 1)

Thirty-four studies reported on the correlation between baseline ctDNA level and survival and/or treatment response (Table 1) [9–42]. The studies were heterogeneous with regard to study design, therapy, treatment line, analytical method, ctDNA marker/cut-off and the number of included patients. Only five [14, 21, 23, 26, 40] of the 34 publications claimed to follow the REMARK guidelines [43].

Table 1.

Studies reporting on correlation between baseline ctDNA level and treatment response and/or survival.

First author	Year	Line of treatment	Study design	Analytical method	ctDNA marker	Cut-off	n	ctDNA marker associated with			Ref.
					Genetic mutation unless otherwise specified			Treatment response	PFSk	OSk
Lefebure	2010	Various/unknown	Prospective biomarker study	qPCR (serum)	KRAS, methylation of RASSF2A	Detection	25	Yes	Yes	Yes	[9]
Spindler	2012	Third-line	Prospective biomarker study	qPCR	KRAS	75th percentile ctDNA level	35	Yes	Yes	Yes	[10]
Spindler	2013	Second-line	Prospective biomarker study	qPCR	KRAS	Detection	96	Yes	Yes	Yes	[21]
Sefrioui	2015	Various/unknown	Prospective biomarker study	dPCR	KRAS	Detection	16	NA	NA	Yes	[32]
Wong	2015	Late-line	Prospective open-label non-RCT	BEAMing	KRAS	Detection	14	NA	Yes	NA	[37]
Shitara	2016	Second-line	Propective phase II RCT	Targeted Sequencing (serum)	KRAS, NRAS, BRAF	Detection	109	D	NA	NA	[38]
Yamada	2016	Various/unknown	Retrospective cohort	Peptide nucleic acid clamping	KRAS	Detection	24	D	NA	NA	[39]
El Messaoudi	2016	Various/unknown	Retrospective cohort	qPCR	KRAS, BRAF	Detectionc	92	NA	NA	No	[40]
Herbsta	2017	First-line	Prospective phase III RCT	qPCR	Methylation of HPP1	Detection	467	NA	NA	Yes	[41]
Toledo	2017	First-line	Retrospective case study	BEAMing	KRAS, NRAS, BRAF, PIK3CA	Detection	ND	D	NA	NA	[42]
Vidal	2017	Various/unknown	Retrospective cohort	BEAMing	KRAS, NRAS	MAF = 1.0%	22	D	Yes	Yes	[11]
Garlan	2017	Various/unknown	Prospective biomarker study	ddPCR	Tumour-informed (KRAS/ BRAF/TP53) or hypermethylation WIF1/ NPY	>10 ng/ml vs ≤0.1 ng/mld	ND	NA	NA	Yes	[12]
Yao	2018	First-line	Retrospective cohort	Targeted Sequencing	KRAS, NRAS, HRAS, BRAF	Detection	27	NA	Yes	NA	[13]
Thomsena	2018	First-line	Prospective biomarker study	ddPCR	KRAS, NRAS, BRAF	Median ctDNA level	77	NA	Yes	NA	[14]
Kimb	2018	Late-line	Prospective phase III non-RCT	Targeted Sequencing	KRAS, NRAS	Detection	238	D	NA	Yes	[15]
Khan	2018 2	Late-line	Prospective phase II non-RCT	ddPCR	RAS pathway genes	Detection	22	No	Yes	Yes	[16]
Takayama	2018	Various/unknown	Prospective biomarker study	ddPCR	KRAS	Detection	39	NA	Yes	NA	[17]
Maurel	2019	First-line	Prospective biomarker study	qPCR	KRAS, NRAS, BRAF	Detectione	178	NA	Yes	Yes	[18]
Cremolini	2019	Third-line	Propective phase II non-RCT	ddPCR + Targeted Sequencing	KRAS, NRAS, BRAF	Detection	25	Noi	Yes	No	[19]
Elez	2019	Various/unknown	Retrospective cohort	BEAMing	KRAS, NRAS	MAF = 5.8%	69	Yesj	Yes/nol	Yes	[20]
Lyskjaer	2019	First-line	Propective biomarker study	ddPCR	Tumour-informed	75th percentile ctDNA levelf	21	NA	Yes	NA	[36]
Amatu	2019	Late-line	Prospective biomarker study	BEAMing	Methylation of EYA4, GRIA4, ITGA4, MAP3K14-AS1, MSC	Median meth-ctDNA levelg	57	NA	Yes	Yes	[22]
Jensen	2019	Late-line	Prospective phase II non-RCT	ddPCR	Methylation of NPY	Median meth-ctDNA level	82	NA	Nom	Yes	[23]
Siravegna	2019	Various/unknown	Prospective phase II non-RCT	Targeted Sequencing	ApCN HER2	ApCN = 25.82	28	D	Yes	NA	[24]
Holm	2020	First-line	Retrospective case study	ddPCR	KRAS	–	10	D	NA	D	[25]
Thomsena	2020	First-line	Prospective biomarker study	ddPCR	Methylation of NPY	Median meth-ctDNA level	123	NA	No	No	[26]
Ma	2020	First-line	Prospective phase II RCT	Targeted Sequencing	Tumour-informed (77 genes)	Median ctDNA levelh	54	NA	No	NA	[27]
Lueonga	2020	First-line	Prospective phase III RCT	ddPCR	KRAS	32 mutated copies per ml	151	No	Yes	Yes	[28]
Bouchahda	2020	Late-line	Prospective biomarker study	MassArray	KRAS, NRAS, BRAF, PIK3CA, EGFR	Detection	16	D	D	D	[29]
Yamada	2020	First-line	Prospective biomarker study	dPCR	KRAS, NRAS, BRAF, EGFR	Detection	30	Yes	Yes	NA	[30]
Yu	2020	Various/unknown	Prospective biomarker study	BEAMing	KRAS, NRAS	Detection	16	D	Yes	Yes	[31]
Jacobs	2020	Late-line	Prospective 3 + 3 phase 1 b study	Targeted Sequencing	74 genes	–	11	D	NA	NA	[33]
Unseld	2020	Late-line	Prospective phase II non-RCT	sWGS	IchorCNA tumour fraction (iTF)	ichorCNA tumour fraction = 5%	30	NA	No	Yes	[34]
Kang	2020	First-line	Retrospective cohort	Targeted Sequencing	KRAS	Detection	51	NA	NA	Yes	[35]

Studies included in at least one meta-analysis are indicated in bold type. Unless otherwise specified under "Analytical method", cfDNA was extracted from plasma.

RCT Randomised Controlled Trial, qPCR quantitative polymerase chain reaction, ddPCR droplet digital polymerase chain reaction, sWGS shallow whole-genome sequencing, ApCN adjusted plasma copy number, MAF mutant allele fraction, D descriptively with no statistical test, NA not assesed, ND not described, Ref. reference, n number of patients with reported association between baseline ctDNA level and at least one outcome. For some outcomes n may be lower.

^aThomsen 2018 & Thomsen 2020, Lueong 2020 & Herbst 2017: Same cohort but different methods.

^bAlso reported in another study (ref. [75]).

^cPatients with detectable KRAS mutation compared to patients without both KRAS and BRAF mutations.

^dAnalysis of ctDNA was based either on detection of: (1) a mutation previously identified in tumour tissue or, (2) hypermethylatiom of WIF1 or NPY when no dPCR assay was available for the identfied mutation(s) or when no mutation was detected in tumour tissue. In case of several detectable mutations the analysis of KRAS mutation was prioritised. detectable mutations the analysis of KRAS mutation was prioritised.

^eThree groups were compared: Patients harbouring BRAF mutations in ctDNA vs. patients harbouring RAS mutations in ctDNA vs. patients without BRAF and RAS mutations in ctDNA.

^fOne patient-specific mutationwas chosen for each patient.

^gMeth-ctDNA level was estimated as an average of the level of all positive markers.

^hMean Allele Fraction (mAF) for each plasma sample was calculated across somatic variants pre-defined by the matched tissue sample.

ⁱThe fraction of patients with RAS mutations in the group of patients who achieved partial response was compared with the fraction of patients with RAS mutation in the group of patients who did not achieve partial response.

^jMAF was significantly higher in patients whose outcome was progressive disease compared to partial response. Response rates were not stratified according to dichotomous MAF.

^kUnivariate analysis.

^lTwo cohorts with different results.

^mPFS at 2 months.

Patients received treatment in first-line (n = 13) [13, 14, 17, 18, 25–28, 30, 35, 36, 41, 42], second-/third-line (n = 4) [10, 19, 21, 38] or late-line (n = 8) [15, 16, 22, 23, 29, 33, 34, 37]. The treatment line was various or unknown in a substantial number of studies (n = 9) [9, 11, 12, 20, 24, 31, 32, 39, 40].

The most commonly used study design was prospective biomarker study (n = 14) [9, 10, 12, 14, 17, 18, 21, 22, 26, 29–32, 36] followed by retrospective cohort (n = 6) [11, 13, 20, 35, 39, 40]. In the prospective studies, blood samples were collected prospectively but most frequently analysed retrospectively.

The majority of studies used polymerase chain reaction (PCR) based analytical methods (i.e., ddPCR, qPCR, BEAMing) (n = 25) [9–12, 14, 16–23, 25, 26, 28, 30–32, 36, 37, 39–42], one in combination with targeted sequencing [19]. The remaining studies used targeted sequencing (n = 7) [13, 15, 24, 27, 33, 35, 38], shallow whole-genome sequencing [34] and mass spectrometry [29].

RAS/RAF mutations were the ctDNA markers of interest in 24 studies [9–11, 13–21, 25, 28–32, 35, 37–40, 42], followed by methylation of one or more genes (i.e., NPY, HPP1, RASSF2A) in five studies [9, 22, 23, 26, 41].

The limit of detection was the chosen cut-off in 19 of the studies [9, 13, 15–19, 21, 29–32, 35, 37–42]. In seven studies the 50th or 75th ctDNA level percentile was used as cut-off [10, 14, 22, 23, 26, 27, 36], and in another six studies the choice of cut-off was based on data from healthy individuals [28], other statistics [20, 24] or not justified [11, 12, 34]. In two studies cut-off was not described and the association was reported without statistical tests [25, 33].

Five studies described a statistically significant association between ctDNA and treatment response [9, 10, 20, 21, 30]. Four studies reported a significantly higher disease control rate [9, 10, 21] or response rate [30] in patients with ctDNA below the cut-off compared to patients with ctDNA levels above the cut-off. Elez et al. found that mutant allele fraction (MAF) was significantly higher in patients whose outcome was progressive disease (PD) compared to partial response (PR), but did not stratify response rates according to dichotomous MAF [20]. Three studies failed to show any statistically significant association between ctDNA and treatment response [16, 19, 28]. Ten studies described a positive association between ctDNA and treatment response without performing statistical tests [11, 15, 24, 25, 29, 31, 33, 38, 39, 42], whereas the remaining studies did not assess the relation between ctDNA and treatment response [12–14, 17, 18, 22, 23, 26, 27, 32, 34–37, 40, 41].

Baseline ctDNA level above the cut-off was associated with shorter PFS in 18 studies [9–11, 13, 14, 16–22, 24, 28, 30, 31, 36, 37], whereas four studies did not find a statistically significant association [23, 26, 27, 34]. The remaining studies did not test the association between ctDNA and PFS [12, 15, 25, 29, 32, 33, 35, 38–42].

Twenty studies evaluated the association between baseline ctDNA and OS [9–12, 15, 16, 18–23, 26, 28, 31, 32, 34, 35, 40, 41]. Seventeen studies found a statistically significant shorter OS in patients with baseline ctDNA level above the cut-off [9–12, 15, 16, 18, 20–23, 28, 31, 32, 34, 35, 41]. Three studies evaluated the association without significant results [19, 26, 40]. The remaining studies did not test the association between ctDNA and OS [13, 14, 17, 24, 25, 27, 29, 30, 33, 36–39, 42].

Meta-analysis of the association between baseline ctDNA and PFS included 11 studies [10, 13, 16, 19–22, 27, 28, 31, 36]. Minimal heterogeneity was found among studies (chi-squared = 8.2, d.f. = 11, p = 0.7, I² = 0.0%). A baseline ctDNA level above the cut-off was associated with a shorter PFS (pooled HR = 2.2; 95% CI 1.8–2.8) (Fig. 4a).

Fig. 4. Forest plots of the association between ctDNA levels and survival.

Forest plots of the association between unfavorable baseline ctDNA levels and progression-free survival (a) and overall survival (b); between unfavorable early dynamics in ctDNA levels and progression-free survival (c) and overall survival (d); all under the random-effects model. Dotted lines indicate first-line studies. Studies are ordered according to publication year. PFS progression-free survival, OS overall survival, n number of patients included in the analysis, HR hazard ratio, CI confidence interval.

Thirteen studies were included in the meta-analysis of baseline ctDNA and OS [10, 12, 15, 16, 19–22, 28, 31, 34, 40, 41]. Heterogeneity was found among studies (chi-squared = 22.7, d.f. = 13, p = 0.05, I² = 45.8%). A baseline ctDNA level above the cut-off was associated with a shorter OS (pooled HR = 2.4; 95% CI 1.9–3.1) (Fig. 4b).

Funnel plots indicate no publication bias (Supplementary Fig. 3).

Despite the inter-study heterogeneity, the outcome of the studies was uniform and indicates that the baseline ctDNA level had a prognostic value.

Early dynamics of ctDNA levels and treatment response/survival (b—Fig. 1)

Twenty-two studies [11, 12, 22, 26, 28, 36, 37, 41, 44–57] reported correlation between early dynamics in ctDNA and treatment response and/or survival (Table 2). The studies were heterogeneous. One [56] of the 22 publications claimed to follow the REMARK guidelines [43].

Table 2.

Studies reporting on correlation between early dynamics in ctDNA level and treatment response and/or survival.

First author	Year	Line of treatment	Study design	Analytical method	ctDNA marker	Second Sample	Groups compared	n	ctDNA marker associated with			Ref
					Genetic mutation unless otherwise specified				Treatment response	PFSq	OSq
Tie	2015	First-line	Prospective biomarker study	Targeted Sequencing	Tumour-informedb	After cycle 1	<10-fold reduction change in ctDNA	52	Yes	No	No	[56]
							≥10-fold reduction change in ctDNA
Wong	2015	Late-line	Prospective open-label non-RCT	BEAMing	KRAS	Day 8 of treatment	–	14	NA	D	NA	[37]
Hong	2016	Various/unknown	Prospective phase I 3+3 non-RCT	ddPCR + Targeted Sequencing	BRAF	After one dose of treatment	–	12	Yesn	NA	NA	[57]
Herbsta	2017	First-line	Prospective phase III RCT	qPCR	Methylation of HPP1	After cycle 1	Detection at 2nd sample	300	NA	NA	Yes	[41]
							Undetectable at BL and at 2nd sample
Vidal	2017	Various/unknown	Retrospective cohort	BEAMing	KRAS, NRAS	At 8–12 weeks of treatment	–	21	Yeso	NA	NA	[11]
Garlan	2017	Various/unknown	Prospective biomarker study	ddPCR	Tumour-informed (KRAS/BRAF/TP53) or hypermethylation WIF1/NPY	Before cycle 2 and/or 3	Bad ctDNA respondersf	108	Yes	Yes	Yes	[12]
							Good ctDNA respondersf
Khan	2018 1	Late-line	Prospective Phase II non-RCT	ddPCR	RAS	At 2 months of treatment	No decrease in RAS FA	21	NA	Yes	No	[44]
							Decrease in RAS FA
Boeckx	2018	First-line	Prospective biomarker study	ddPCR	Methylation of NPY and tumour-informed	Before cycle 2 and at 1st radiographic evaluation	–	23	D	NA	NA	[45]
Hsu	2018	First-line	Retrospective cohort	Targeted Sequencing	12 oncogenes	Next available plasma sampling	ctDNA variant frequency reduction of ≤80%g	15	Yes	Yes	NA	[46]
							ctDNA variant frequency reduction of >80%g
Vandeputte	2018	Late-line	Prospective biomarker study	ddPCR + Targeted Sequencing	47 genes	Day 14 of first cycle	Absolute increase in mutated copies/mlh	19	NA	Yes	Yes	[47]
							Absolute decrease in mutated copies/mlh
Corcoran	2018	Various/unknown	Prospective phase I RCT	BEAMing	BRAF	At 4 weeks of treatment	–	71	Yesn	NA	NA	[48]
Barault	2018	Late-line	Retrospective cohort	BEAMing	Methylation of EYA4, GRIA4, ITGA4, MAP3K14-AS1, MSC	Lowest measured ASM during treatmente	Best ASM change <0i	25	Yes	Yes	NA	[49]
							Best ASM change ≥0i
Lyskjaer	2019	First-line	Propective biomarker study	ddPCR	Tumour-informed	Day 7, 14, 21 and 60 of treatment	Two temporary increases in MAFj	21	NA	Yes	No	[36]
							Less than two temporary increases in MAFj
Jia	2019	First-line	Prospective biomarker study	Targeted Sequencing	50 genes	After cycle 1	ctDNA log2 (C1/C0) > −0.126k	41	Yes	Yes	NA	[50]
							ctDNA log2 (C1/C0) ≤ −0.126k
Osumi	2019	Second-line	Prospective biomarker study	Targeted Sequencing	14 genes	At 8 weeks of treatment	>50% after/before ratio in ctDNA levelsl	26	Yes	Yes	Yes	[51]
							≤50% after/before ratio in ctDNA levelsl
Amatu	2019	Late-line	Prospective biomarker study	BEAMing	Methylation of EYA4, GRIA4, ITGA4, MAP3K14-AS1, MSCc	1st blood sample after treatment start	cmDNA increase	52	NA	Yes	NA	[22]
							cmDNA decrease
Thomsena	2020	First-line	Prospective biomarker study	ddPCR	Methylation of NPY	After cycle 1	Others	123	No	Yes	Yes	[26]
							Level = 0 at BL or 0 in 95% CI at 2nd sample
Lueonga	2020	First-line	Prospective phase III RCT	ddPCR	KRAS	15–22 days after start of treatment	Detectable at BL and at 2nd sample	113	NA	Yes	Yes	[28]
							Detectable at BL and undetectable at 2nd sample
							Undetectable at BL and detectable at 2nd sample	38	NA	NA	Yes
							Undetectable at BL and at 2nd sample
Moser	2020	Various/unknown	Prospective biomarker study	Targeted Sequencing	Tumour-informedd	52 h after treatment start	–	11	Yesp	NA	NA	[52]
Klein-Scory	2020	First-line	Prospective biomarker study	ddPCR	RAS	Various (4–24 weeks)	–	12	D	NA	NA	[53]
Wang	2020	Late-line	Retrospective cohort	Targeted Sequencing	74 genes	At 4 weeks of treatment	Increase in ctDNA or emergence of new clonesDecrease in ctDNA	13	D	NA	NA	[55]
							Decrease in ctDNA
Parikh	2020	Various/unknown	Prospective biomarker study	ddPCR	Tumour-informed	At 4 weeks of treatment	<30% decrease in MAFm	55	Yes	Yes	NA	[54]
							≥30% decrease in MAFm

Studies included in at least one meta-analysis are indicated in bold type. In all studies included in Table 2, cfDNA was extracted from plasma.

RCT Randomised Controlled Trial, qPCR quantitative polymerase chain reaction, ddPCR droplet digital polymerase chain reaction, FA fractional abundance, ASM average of selected markers, MAF mutant allele fraction, cmDNA circulating methylated DNA, BL baseline, NA not assesed, D descriptively with no statistical test, Ref. references, n number of patients with reported association between early dynamics of the ctDNA level during treatment and at least one outcome. For some outcomes n may be lower.

^aThomsen 2018 (baseline) & Thomsen 2020, Lueong 2020 & Herbst 2017: Same cohort but different methods.

^bOne of the mutations identified in the tumour tissue was assessed.

^cMeth-ctDNA level is estimated as an average of the level of all positive markers.

^dIn the tumour of one patient, no mutation was identified. Instead, a mutation identified in plasma by targeted sequencing was tracked.

^eASM: calculated based exclusively on the loci which displayed positive methylation at baseline.

^fA combined marker integrating baseline ctDNA level (below/above 0.1 ng/mL) and if the drop in ctDNA during the first 4 weeks was smaller or greater than 80%.

^gctDNA frequency reduction was defined as the difference between the frequency of the ctDNA variant with the highest variant frequency at treatment initiation and the frequency of the ctDNA variant with the highest variant frequency at follow-up.

^hOn average 3 mutations (range 1–4) were selected per patient based on the highest VAFs

ⁱASM change: ASM at a longitudinal time point was substracted from the ASM at baseline, and the best change (lowest) over the course of treatment was selected for correlative assesment.

^jctDNA increase relative to previous sample. One patient-specific mutation was chosen for each patient.

^kFor each patient, the mutation of the maximal frequency in the pretreatment plasma ctDNA sample was selected as the candidate target for analysis.

^lctDNA level was defined as the highest allele frequency of the detected mutant alleles at each time point in each patient when two or more mutations were detected.

^mFor patients with multiple assesable mutations, the percent change in MAF of up to three mutations in ctDNA was averaged.

ⁿPatients with partial response were more likely to demonstrate a deeper reduction in ctDNA fraction relative to those with stable disease or progressive disease at the time of first restaging (P < 0.01).

^oMAF percentage of change was significantly lower in patients with progressive disease compared with patients with partial response or stable disease at the time of first restaging (P = 0.027).

^pCompared to patients with progressive disease patients with stable disease/partial response/complete response ctDNA mAFs remained at decreased levels at the time of last blood collection, i.e. T9 (between T1 and T5: p = 0.03906; between T1 and T9: p = 0.01563).

^qUnivariate analysis.

Patients received treatment in first-line (n = 9) [26, 28, 36, 41, 45, 46, 50, 53, 56], second-line (n = 1) [51] or late-line (n = 6) [22, 37, 44, 47, 49, 55]. The treatment line was various or unknown in six studies [11, 12, 48, 52, 54, 57].

The most common study design was a prospective biomarker study with retrospective ctDNA analysis (n = 12) [12, 22, 26, 36, 45, 47, 50–54, 56], followed by a retrospective cohort design (n = 4) [11, 46, 49, 55].

Fourteen studies used PCR-based analytical methods (ddPCR, BEAMing and qPCR) [11, 12, 22, 26, 28, 36, 37, 41, 44, 45, 48, 49, 53, 54], six used targeted sequencing [46, 50–52, 55, 56] and two studies both approaches [47, 57].

RAS/RAF mutations were the ctDNA markers of interest in seven studies [11, 28, 37, 44, 48, 53, 57], followed by methylation of one or more genes (i.e., HPP1, NPY, WIF1) in six studies [12, 22, 26, 41, 45, 49].

Most commonly, ctDNA from a baseline sample was compared to ctDNA in a blood sample taken at a pre-defined time point or time period during treatment (second sample) [11, 12, 26, 28, 37, 41, 44, 47, 48, 50–52, 54–57]. The time point for the second sample varied from 52 hours [52] to 12 weeks after treatment started [11]. Two studies compared the baseline sample to “the next available sample” without defining the time point for sampling [22, 46]. Another two studies did consecutive sampling and compared the dynamics with the outcome [36, 45]. Barault et al. compared the baseline sample with the lowest measured average of selected markers [49].

Fifteen studies dichotomised the study population based on early dynamics in ctDNA for further analysis of relation to outcome. Four studies reported an increase versus decrease in the ctDNA level, where the ctDNA level was measured as a fraction of total cell-free DNA [44, 55] or copies per ml plasma [22, 47]. Two studies described changes in detectability when dichotomising the study population [28, 41]; one study defined detectability based on data from healthy controls [28] while the other did not define the limit of detection [41]. Five studies defined a decrease in ctDNA level (i.e., 10-fold decrease, 30% decrease), and divided the study population into two based on whether this decrease was present or not [46, 50, 51, 54, 56]; the decrease was based on ROC curves [46, 50, 56], Kaplan–Meier curves [51], or based on a reason not described [54]. The remaining four studies dichotomised the study populations in various other ways [12, 26, 36, 49]. Seven studies did not dichotomise the study population [11, 37, 45, 48, 52, 53, 57].

Twelve studies investigated the statistical association between early dynamics of ctDNA level and treatment response [11, 12, 26, 46, 48–52, 54, 56, 57], and all but one [26] found a statistically significant association. Eight studies investigated whether a pre-defined early change in ctDNA level was correlated to treatment response [12, 26, 46, 49–51, 54, 56]. Conversely, four studies analysed if the treatment response correlated to early change in ctDNA level [11, 48, 52, 57]. Three studies described the association between early dynamics in ctDNA level and treatment response, without doing statistical tests [45, 53, 55]. The remaining studies did not assess the relation between early dynamics in ctDNA level and treatment response [22, 28, 36, 37, 41, 44, 47].

A small or no early decrease in ctDNA levels during treatment was associated with short PFS in 12 studies [12, 22, 26, 28, 36, 44, 46, 47, 49–51, 54], whereas only one study did not find a significant association [56]. In the remaining studies an association between early dynamics in ctDNA level and PFS was not tested [11, 37, 41, 45, 48, 52, 53, 55, 57].

A small or no early decrease in ctDNA levels during treatment was associated with short OS in six studies [12, 26, 28, 41, 47, 51], whereas three studies did not find a significant association [36, 44, 56]. Thirteen studies did not test the association [11, 22, 37, 45, 46, 48–50, 52–55, 57].

In addition, four studies reported on the association between ctDNA level on a single early pre-defined time point during treatment and treatment response and/or survival [14, 25, 28, 58]. See Supplementary Table 3.

Meta-analysis of the association between early dynamics in ctDNA level and PFS included eleven studies [12, 22, 26, 36, 44, 46, 47, 49, 51, 54, 56]. Minimal heterogeneity was found among studies (chi-squared = 13.6, d.f. = 10, p = 0.19, I² = 35.2%), and as shown in Fig. 4c, the largest decrease in ctDNA level was associated with a longer PFS (Pooled HR = 3.0; 95% CI 2.2–4.2).

Seven studies were included in the meta-analysis of early dynamics in ctDNA level and OS [12, 26, 36, 41, 44, 47, 51]. Minimal heterogeneity was found among studies (chi-squared = 7.3, d.f. = 6, p = 0.30, I² = 27.3%), and the largest decrease in ctDNA level was also associated with a longer OS (Pooled HR = 2.8; 95% CI 2.1–3.9) (Fig. 4d).

Funnel plots indicate no publication bias (Supplementary Fig. 3).

Despite the inter-study heterogeneity, the outcome of the studies was uniform and indicates that early ctDNA dynamics had a prognostic value.

Lead-time (c—Fig. 1)

Ten studies reported on lead-time (Supplementary Fig. 4) [14, 16, 23, 30, 39, 45, 59–62]. Median lead-time was reported in eight studies and ranged from 0 to 3.6 months [14, 16, 23, 45, 59–62]. The remaining two studies did not report median lead-time [30, 39]. Generally, lead-time was only reported for a subset of patients, and in some studies confined to cases with a positive lead-time [16, 23, 39, 45, 59].

Frequency of blood sampling and radiological evaluation, and the coordination of these, varied from study to study. The majority of the studies did not report on the interval between radiological evaluations [16, 39, 45, 59, 62]. In four studies, blood samples were more frequent than radiological evaluations [14, 23, 30, 60]. Hence, the reported lead-time could be influenced by the intervals between blood sampling and radiological evaluation. Liu et al. collected blood samples at each radiological evaluation [61] and they reported a lead-time range of 0 to 7 months with a median of 0 months.

The definition of molecular biological progression differed from study to study, but can be grouped into two categories; detection of previously undetectable mutations in cfDNA [16, 30, 39, 62] (primarily RAS mutations) or an increase in ctDNA level [14, 23, 45, 59–61].

Zou et al. defined an increase as a rise of ≥5 mutation copies per ml plasma [59]. Thomsen et al defined an increase as a rise in ctDNA level where the 95% CI does not overlap with the immediately preceding measurement [14]. The remaining studies did not report a definition of an increase [23, 45, 60, 61].

The data presented here indicates that molecular biological progression could precede radiological progression.

Clonal evolution (d—Fig. 1)

Of the eligible studies, 22 presented exploratory analyses of potential clonal evolution during therapy [11, 15, 16, 18, 30, 31, 36, 39, 42, 48, 49, 57, 58, 62–70]. However, inter-study variation was substantial i.e., in regard to mutations evaluated, analytical methods, treatment and purpose of the analysis.

In total, 674 study participants had blood samples drawn at radiologically confirmed disease progression, and the emergence of mutations was found in 235 patients.

Most studies investigated acquired resistance to anti-EGFR treatment [11, 15, 16, 18, 30, 31, 39, 42, 48, 49, 62–65, 68, 70]. Results were most often confined to a description of frequency and type of mutation [11, 16, 30, 39, 42, 48, 49, 57, 63–66, 68, 70]. Some  studies also investigated the association with survival; two studies found a significantly shorter OS in patients with emerging RAS/BRAF mutations compared to patients without emerging mutations during systemic therapy, including anti-EGFR [18, 69], whereas four studies did not find a significant association [15, 31, 62, 67].

Interestingly, a few studies observed that RAS mutations emerged during anti-EGFR treatment and became undetectable when treatment was interrupted, supporting the rationale for anti-EGFR re-challenge [16, 49, 64]. Of note, prospective investigations of re-challenge have been reported by Cremolini et al. and Yu et al. [19, 31]. Both studies reported that patients with RAS mutations at baseline had a significantly shorter PFS when re-challenged with anti-EGFR compared to patients without RAS mutations [19, 31]. In addition, Yu et also found a significantly shorter OS in patients with RAS mutations at baseline [31].

Finally, Choi et al. [63] evaluated the presence of potentially actionable targets in 78 patients using targeted sequencing analysis of ctDNA at the time of progression, demonstrating that 76% of the patients had potentially actionable targets. In this study it was unknown if mutations were present in ctDNA at baseline or were a result of clonal evolution.

The search strategy for this systematic review was not aimed at presenting a systematic overview of clonal evolution, and further review of this topic is beyond the scope of this publication.

Discussion

Summary of main results

We conducted a systematic review and identified four major clinical applications of liquid biopsies, with potential relevance to this patient group (Fig. 1). Seventy-one studies were included in the systematic review, 24 of these were eligible for a meta-analysis investigating the prognostic impact. We demonstrated a statistically significant prognostic value of both the baseline ctDNA level and the early dynamics in ctDNA levels during treatment.

None of the included studies were designed to investigate if ctDNA is a predictive biomarker [71]. However, several studies investigated a possible predictive value of the selected biomarker by evaluating the relation to treatment response. There is a fairly consistent association between early dynamics in ctDNA and treatment response. However, the association between baseline ctDNA level and treatment response was less clear. None of the studies reporting on lead-time were designed to investigate this specifically. Hence, it is not possible to conclude on lead-time from these results.

The studies reporting on clonal evolution focus on mechanisms of resistance to anti-EGFR and few studies explore the possibilities of re-challenge. Only one of the eligible studies evaluated the presence of potentially actionable targets.

Overall completeness and applicability of evidence

Most studies investigated if ctDNA could predict prognosis and treatment response prior to/during systemic treatment. We identified fewer articles thoroughly investigating lead-time or clonal evolution. The patient populations included were heterogeneous in terms of disease characteristics, and in most studies, patients were included in first- or late-line. Interestingly, a fair number of studies did not specify a treatment line, which questions the generalisability and applicability of the results.

There is now consensus that when dealing with blood, cfDNA should be extracted from plasma [72]. Most of the eligible studies were published prior to this consensus. Despite this, only three studies extracted cfDNA from serum. Although most studies utilised plasma as the source of ctDNA, there was extensive variation in pre-analytical and analytical procedures when comparing the different studies. The processing steps varied (e.g., blood draw, centrifugation, storage) as well as the total plasma/serum volume of which ctDNA was extracted. Variations in processing may affect the analytical result, for instance by introducing contamination of DNA from leukocytes [73]. A low plasma volume may lead to subsampling issues and limit test sensitivity. Both are relevant examples of variations that can affect the associations identified between ctDNA and clinical outcome.

The choice of analytical method seemed closely related to the choice of ctDNA marker, and we believe the latter should be guided by the research question at hand. When investigating prognosis or treatment response, ctDNA markers were often tumour-informed (e.g., individualised based on somatic tumour tissue mutations) using one or a few genetic loci. Different analytical methods could be applied, including; qPCR, ddPCR and sequencing approaches. The potential limitations include failure of identifying suitable somatic tissue mutations, and the risk of choosing markers representing sub-clones that are not relevant for the research question at hand. Some studies try to overcome this challenge by choosing several loci, making sure that they include variants with high variant allele frequency in the tissue.

Obvious pitfalls include choosing a tumour-agnostic (i.e., uninformed) ctDNA marker in an unselected population of mCRC patients; which can falsely score patients with a ctDNA negative result. Some studies have tried to identify a combination of methylation markers that can identify ctDNA in mCRC irrespective of somatic mutations. This approach needs to be investigated further, but could prove useful as a tumour-agnostic ctDNA marker for mCRC patients.

When investigating clonal evolution, ctDNA markers were often tumour-agnostic (i.e., uninformed) using multiple genetic loci. Sequencing approaches were more commonly used for this purpose, and it is crucial that the gene panels are clinically relevant (e.g., identifying resistant clones and/or relevant drug targets). An obvious advantage of this approach is the potential of gaining novel information about clonal evolution along the care pathway without multiple biopsies (spatially and temporally). Theoretically, this could also guide treatment. However, none of the identified studies investigated this prospectively, and further studies will be needed to show its cost-effectiveness in improving survival and quality of life.

Test sensitivity is a major analytical challenge when working with ctDNA, and it can be limited both by low total cfDNA input and by the chosen analytical method. Most studies utilise DNA isolated from a fixed plasma volume. Knowing that cfDNA levels vary between and within individuals, the actual test sensitivity may vary to a great extent unless one defines a “minimum cfDNA input” and a threshold test sensitivity. Although some studies describe sensitivity/detection limit as a function of total cfDNA input, it is our impression that this is lacking too often. In this setting, a negative result cannot be fully interpreted. More importantly, a clinical cut-off using detection vs. non-detection has no generalisability unless test sensitivity is acknowledged.

Due to the great heterogeneity among studies, it was not possible to investigate whether discrepancies in results were attributed to differences in pre-analytical and/or analytical parameters. Concerning different analytical methods used, it has previously been shown that there is a strong correlation between next-generation sequencing and ddPCR [74]. In accordance with this, we did not find systematic differences in outcome based on whether a PCR-based method or targeted sequencing were used, when visually assessing forest plots of the meta-analyses.

Quality of the evidence

The body of evidence strongly suggests that elevated baseline ctDNA levels confer poor prognosis, but most patients were included in first- or late-line, and there is no consensus as to what ctDNA marker is better or what is the optimal cut-off level. It is still unclear to what degree elevated ctDNA levels can complement or replace existing prognostic markers in mCRC.

There may be an association between early ctDNA dynamics and radiological response to treatment and prognosis; but there is no consensus as to which ctDNA marker is better, the optimal timing of sampling during treatment or the most clinically relevant ctDNA changes. The length of time between molecular biological and radiological progression (lead-time) is still under investigation, and it is unknown if clinical outcomes will improve if treatment is guided by ctDNA rather than/in addition to radiological assessment.

The body of evidence indicates that ctDNA in some instances can be used as a tool to investigate clonal evolution and potentially guide treatment; but up until now, most studies have investigated this concept in subgroups or even in single cases.

Close to 70% of the included studies had a high risk of bias in at least one domain, and only 10% explicitly state that they followed the REMARK guidelines when conducting the study.

Potential biases in the review process

We aimed to describe the potential clinical utility of ctDNA across different parts of the care pathway. This resulted in an extensive literature search and broad inclusion criteria. We acknowledge that our review may have several potential biases. Firstly, our initial focus was the predictive/prognostic value of ctDNA during treatment of mCRC; but during the literature screening process, we chose to include topics including lead-time, clonal evolution and drug targets. Although we decided to include literature covering all topics, we cannot rule out that there has been an inclusion bias in this regard. Secondly, due to substantial heterogeneity between the studies included we chose to summarise results both quantitatively and qualitatively, depending on the data that was available to us. A substantial number of studies did not have sufficient data for inclusion in meta-analyses, which is an obvious limitation. Thirdly, the quality and risk of bias associated with studies in the meta-analysis varied considerably, and hence we cannot rule out overstatements of the strengths and precision of our combined effect measures. Finally, we did not source unpublished material for inclusion. This could introduce a publication bias, and subsequently lead to overestimating the true effects in the meta-analyses. However, we did not see any major tendencies of publication bias when considering the associated funnel plots visually.

Future perspective

Although ctDNA holds prognostic value in patients with mCRC, prospective clinical trials with standardised methodologies are needed to show an added clinical value for this patient group. For instance, it can be hypothesised that a risk-stratified or risk-adapted treatment strategy using ctDNA could be beneficial. However, this strategy has not yet proven to prolong survival or improve quality of life compared with standard of care and established prognostic markers.

In addition to harbouring prognostic information, ctDNA has the advantage of carrying tumour-specific information of potential relevance for guiding treatment. This potential has only partially been unravelled in the care pathway of patients with mCRC, and should be evaluated in larger prospective trials designed to investigate the predictive value of ctDNA. Potential clinical applications include identification of patients eligible for re-challenge with anti-EGFR therapy, BRAF targeted therapy, HER2-related therapy or checkpoint inhibition (i.e., MSI-H or high tumour mutational burden).

Numerous explorative and descriptive studies during the last ten years have laid the foundation for further exploring the clinical utility of ctDNA in mCRC. We believe that the current treatment landscape for mCRC is awaiting prospective randomised trials with companion diagnostics, in order to bring ctDNA guidance closer to daily patient care.

Author contributions

Study concept and design: JH, AB, KLS, LC, data collection and collation: JH, AB, LC, statistical analysis: JH, LC, writing—original draft: JH, LC, writing—review and editing: JH, AB, NP, TKG, EHK, KLS, LC, Interpretation of the data, critical revision of the paper for important intellectual content and approval of the final paper for submission: JH, AB, NP, TKG, EHK, KLS, LC. The corresponding author proves that all listed authors meet the authorship criteria, and that no other eligible authors have been omitted.

Funding

LC was supported by DCCC ctDNA Research Center—The Danish Research Center for Circulating Tumor DNA Guided Cancer Management, Danish Cancer Society (grant no. R257-A14700) and Danish Comprehensive Cancer Centers. KLS was supported by Health Research Foundation of Central Denmark Region (grant no. A1602). Funding has been provided by Danish Cancer Society, Health Research Foundation of Central Denmark Region. Registered PROSPERO(CRD42019125630). The remaining authors received no specific funding for this work.

Data availability

The full text of all included studies were retrieved from the online databases PubMed, Embase, Cochrane Database of Systematic Reviews and Cochrane Central Register of Controlled Trials. The data of this systematic review and meta-analyses are all public and available from PubMed, Embase, Cochrane Database of Systematic Reviews and Cochrane Central Register of Controlled Trials. Template data collection forms, data extracted from included studies, data used for analyses and analytic code used and/or analysed during the current study are available from the corresponding author on reasonable request.

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Supplementary information

The online version contains supplementary material available at 10.1038/s41416-022-01816-4.