Liquid biopsy in non-small cell lung cancer: a meta-analysis of state-of-the-art and future perspectives

Introduction: To date, tissue biopsy represents the gold standard for characterizing non-small-cell lung cancer (NSCLC), however, the complex architecture of the disease has introduced the need for new investigative approaches, such as liquid biopsy. Indeed, DNA analyzed in liquid biopsy is much more representative of tumour heterogeneity. Materials and methods: We performed a meta-analysis of 17 selected papers, to attest to the diagnostic performance of liquid biopsy in identifying EGFR mutations in NSCLC. Results: In the overall studies, we found a sensitivity of 0.59, specificity of 0.96 and diagnostic odds ratio of 24,69. Since we noticed a high heterogeneity among different papers, we also performed the meta-analysis in separate subsets of papers, divided by 1) stage of disease, 2) experimental design and 3) method of mutation detection. Liquid biopsy has the highest sensitivity/specificity in high-stage tumours, and prospective studies are more reliable than retrospective ones in terms of sensitivity and specificity, both NGS and PCR-based techniques can be used to detect tumour DNA in liquid biopsy. Discussion: Overall, liquid biopsy has the potential to help the management of NSCLC, but at present the non-homogeneous literature data, lack of optimal detection methods, together with relatively high costs make its applicability in routine diagnostics still challenging.

Introduction: To date, tissue biopsy represents the gold standard for characterizing non-small-cell lung cancer (NSCLC), however, the complex architecture of the disease has introduced the need for new investigative approaches, such as liquid biopsy.Indeed, DNA analyzed in liquid biopsy is much more representative of tumour heterogeneity.

Materials and methods:
We performed a meta-analysis of 17 selected papers, to attest to the diagnostic performance of liquid biopsy in identifying EGFR mutations in NSCLC.

Results:
In the overall studies, we found a sensitivity of 0.59, specificity of 0.96 and diagnostic odds ratio of 24,69.Since we noticed a high heterogeneity among different papers, we also performed the meta-analysis in separate subsets of papers, divided by 1) stage of disease, 2) experimental design and 3) method of mutation detection.Liquid biopsy has the highest sensitivity/specificity in highstage tumours, and prospective studies are more reliable than retrospective ones in terms of sensitivity and specificity, both NGS and PCR-based techniques can be used to detect tumour DNA in liquid biopsy.

Introduction
Lung cancer is the main cause of cancer-related death, particularly regarding the broader group of non-small-cell lung cancer (NSCLC), with its three histologic variants: squamous cell carcinoma (SCC), adenocarcinoma (ADC) and large-cell carcinoma (Pujol et al., 2022; https://www.cap.org/).ADC is the most common subtype (Bray et al., 2018), accounting for 50% of all lung cancer diagnoses and showing an increase in occurrence in the latter decades (Barta et al., 2019).
NSCLC is often asymptomatic in its early phases and thus many patients are diagnosed only at an advanced stage, resulting in a poor prognosis, with a limited survival rate (approximately 18% at 5 years) (Wu et al., 2019;Abbasian et al., 2022).

EGFR mutational status and target therapy response
Among the most predictive alterations, are the mutations in epidermal growth factor receptor (EGFR).EGFR is a transmembrane receptor identified as an NSCLC oncogenic driver.In normal cells, it is activated by the binding of epidermal growth factor, which triggers different pathways involved in cell cycle progression, growth and angiogenesis (Casula et al., 2023).Mutations in EGFR lead to its constitutive activation, protein over-expression, and tumour progression (Bethune et al., 2010).EGFR mutation status is currently investigated to characterize NSCLC patients and to guide pharmacological treatment.Indeed, specific EGFR mutations confer sensitivity to selective EGFR-TKI inhibitors, thus allowing a targeted therapy, based on the molecular profile of the tumour, with the potential of improving the patient's overall and progression-free survival, compared to standard chemotherapy (Arbour and Riely, 2019;Cheema et al., 2020).In detail, in NSCLC, EGFR presents with recurrent hot-spot alterations (single nucleotide missense variant as well as small insertion/deletions) at exons 18 to 21, codifying for the tyrosine kinase domain.The highest proportion of gene alterations (80%-90%) is represented by deletions within exon 19 and the point mutation c.2573T>G, p.L858R, at exon 21.Notably, patients harbouring exon 19 deletions have a better outcome, compared to patients with p.L858R, when treated with TKIs (Lino et al., 2023).The remaining 10%-20% of pathogenic EGFR variants are defined as "uncommon mutations" (Attili et al., 2022).Finally, EGFR can also have mutations that confer resistance to TKI inhibitors (e.g., p.T790M).The resistance generally occurs as a consequence of the treatment with TKI inhibitors in patients who showed a previous sensitizing EGFR mutation (Wu and Shih, 2018).

Tissue biopsy vs. liquid biopsy: state-ofthe-art
At present, tissue biopsy (TB) is considered the gold standard for tumour diagnosis and molecular investigation of predictive biomarkers.Nevertheless, it is invasive for patients and has several limitations, mainly related to intra-tumour heterogeneity (i.e., different regions of the same tumour can bring different molecular alterations), as well as inter-tumour heterogeneity (i.e., different molecular profiles between the primary tumour and local or distant metastases of the same patient) (Gerlinger et al., 2012), both making tissue biopsy unrepresentative of the complete genetic makeup of the neoplasia.In addition, tumours can dynamically change over time, with the emergence of treatmentresistant subclones not detectable in the biopsy of the primary tumour (Bedard et al., 2013).Moreover, limiting factors in biomarker testing from tissue biopsy include the adequate quality of nucleic acids (DNA and RNA quality sub-optimal in formalin-fixed tissue, which is the routinary source of tumour tissue in molecular pathology diagnostics) as well as the availability of a sufficient amount of tumour tissue (e.g., tumour cellularity and size of the specimen) due to small tissue-samples delivered per patient respect to the increasing number of molecular markers that need to be investigated.
To overcome these limitations, liquid biopsy (LB), consisting of the analysis of tumour-released nucleic acids circulating in body fluids, such as blood, could provide a non-invasive and welltolerated approach for tumour investigation.It allows the detection of circulating tumour DNA (ctDNA) carrying molecular tumour markers, which are more representative of the entire tumour and enable to follow disease evolution and dynamic changes in the molecular profile (Pérez-Callejo et al., 2016;Rijavec E et al., 2020;Bonanno et al., 2022;Pesta et al., 2022).In addition, in advanced/metastatic NSCLC, liquid biopsy has the potential to drive target therapy, by monitoring the response to treatment and identifying the possible molecular mechanisms of therapy resistance.Recently, Gristina et al. reported on the clinical potential of cfDNA in monitoring outcomes of NSCLC following first-line treatment.cfDNA has shown to be a reliable marker in helping clinicians in the decision-making process.Indeed, dynamic changes in cfDNA correlated with response to therapy with TKI and IO-based therapies.(Gristina et al., 2022).Finally, recent data have shown a significant ability of liquid biopsy in detecting minimal residual disease in early-stage lung cancer, underlying the potential application of LB in the adjuvant setting, in early detection of recurrence, and also for screening (Nigro et al., 2023).
Despite the evident practical advantages of liquid over tissue biopsy, LB is not yet widely adopted in clinical practice (Esagian et al., 2020) and standardized methods of LB investigation are currently lacking.Current Guidelines of ESMO (European Society for Medical Oncology) indicate liquid biopsy as complementary or alternative to tissue for biomarker evaluation of treatment-naïve NSCLC and recommend ctDNA evaluation only when a significant diagnostic delay is expected in obtaining tumour tissue for genotyping, when invasive procedures may be risky or notindicated, or when bone would be the only site that could be biopsied (Pascual et al., 2022).
In the present manuscript, we performed a systematic review with meta-analysis to assess the state-of-the-art diagnostic potential of liquid biopsy in revealing EGFR predictive mutations in NSCLC patients.When different EGFR mutations were distinct in the text, we focused on exon 19 deletions.
The ethical approval was not applicable, because we performed a meta-analysis of the literature without involving human subjects.

Inclusion and exclusion criteria
PubMed database was independently screened by two Authors for articles of interest, according to the inclusion and exclusion criteria listed below; a double cross-check was performed and, in case of discrepancies, a third, independent supervisor was asked to review the collection.Articles were selected, included, and excluded following preferred reporting items for systematic reviews and meta-analysis (PRISMA) guidelines (PRISMA (prisma-statement.org).The inclusion criteria were: 1) human-based studies; 2) studies including at least 20 patients; 3) the absolute number of true positive (TP), true negative (TN), false positive (FP) and false negative (FN) presented in a 2X2 contingency table or easily deducible from the results section.The exclusion criteria comprised: 1) animal studies; 2) not sufficient data to construct a 2X2 contingency table; 3) reviews, meta-analyses, comments, and case reports.

Statistical analysis
A 2X2 table was generated including the absolute number of TP, TN, FP and FN, coupled with sensitivity and specificity data, to assess the diagnostic power of liquid biopsy in comparison with tissue biopsy.
We performed the bivariate Reitsma model (Reitsma et al., 2005) to explore the correlation between the logit of True Positive Rate (TPR) and logit of False Positive Rate (FPR); the confidence interval for correlation was estimated by the semi-parametric bootstrap percentile method.Separate meta-analyses of TPR and FPR were performed using the random-effects frequentist meta-analysis.Sensitivity (SE) and Specificity (SP) were pooled by generalized linear mixed models (GLMM) with logit transformation (Lin and Chu, 2020) by using the maximum likelihood to estimate the between-study variance.Clopper-Pearson confidence intervals were computed for an individual study.Diagnostic odd ratio (DOR), positive (PLR) and negative (NLR) likelihood ratios were pooled using the inverse-variance weighted random-effects frequentist metaanalysis with DerSimonian-Laird estimator for between-study variance (DerSimonian and Laird, 1986).Statistical heterogeneity was evaluated by the I 2 index: a value ≤25% was defined as low heterogeneity, a value between 50% and 75% as moderate heterogeneity, and 75% or larger as high heterogeneity (Higgins and Thompson, 2002).
The 95% confidence intervals (95% CI) for pooled effect estimates were based on standard normal quantile.The prediction interval for the treatment effect of a new study was calculated according to Borenstein et al. (Borenstein et al., 2009).The one-leave-out sensitivity analysis was also performed.The continuity correction of 0.5 in studies with zero cell frequencies was used.The estimation of projected predictive values was based on a prevalence range and pooled (meta-analytical) sensitivities and specificities.All the confidence intervals were computed at a confidence level equal to 95%.

Subgroups stratification
Since we noticed a high variability in the results among different papers, mainly to the heterogeneity of the included studies, we also conducted statistical analyses by dividing patients into subgroups according to the following criteria: 1) tumour stage: high-stage (IIIB and IV) versus low-stage/locally advanced (I, II, IIIA); 2) experimental design: LB performed only in samples EGFR-positive at TB (retrospective studies) versus LB consecutively performed in both EGFR-positive and negative samples (prospective studies); 3) method of EGFR mutation detection: Next-Generation-Sequencing (NGS) versus PCR-based methods (see Table 1).
Table 1 Details of the characteristics of the included studies: first author, year of publication, the country where the study was carried on, number of analyzed patients, and investigated molecular markers.We also indicated the tumour stage, the experimental design (P = prospective, R = retrospective, as specified in the Materials and Methods section) and the method of EGFR mutation investigation by liquid biopsy.Finally, we also reported  *) In the studies where different EGFR mutations were investigated, we only considered deletion in exon 19 (del ex 19).EGFR: epidermal growth factor; KRAS: Kirsten rat sarcoma virus; P: prospective studies in which patients EGFR +/− at tissue biopsy were subjected to liquid biopsy; R: retrospective studies in which only patients EGFR + at tissue biopsy were subjected to liquid biopsy.NGS: Next-Generation Sequencing; ddPCR: digital droplet polymerase chain reaction; CAST-PCR: competitive allele-specific TaqMan PCR; PNA-LNA PCR: peptide nucleic acidlocked nucleic acid PCR; ARMS-PCR: amplification-refractory mutation system; TP: true positive (positive to tissue and liquid biopsy); FP false positive (positive to liquid biopsy and negative to tissue biopsy); TN: negative to tissue and liquid biopsy); FN: false negative (negative to liquid biopsy and positive to tissue positive).
These overall results indicate that, though specificity was high and stable, sensitivity of LB was generally low and highly variable among studies, as shown by the heterogeneity index (I 2 = 92%).This in turn negatively influenced Diagnostic Odd's Ratio, which was also variable (I 2 = 81%).
Overall, the sensitivity and diagnostic odds ratio were higher in the high-stage than in the low-stage tumours.While specificity maintained comparable values between the two groups (0.98 vs. 0.95 in high-and low-stage tumours respectively).
Overall, we observed that liquid biopsy showed slightly higher sensitivity in prospective than in retrospective studies (0.62 vs. 0.55).In contrast, specificity and diagnostic odds ratios were higher in retrospective than in prospective studies (0.99 and 28.40 vs. 0.96 and 26.51).
Though based on a small number of manuscripts, NGS showed slightly higher sensitivity than PCR-based techniques (0.62 vs. 0.56), while specificity was comparable (0.95 vs. 0.97).In contrast, PCRbased methods had a significantly higher diagnostic odds ratio than NGS (36.79 vs. 20.44).

Discussion
In the present meta-analysis, we evaluated data from 17 selected studies with a total of 1711 patients to investigate the diagnostic power of liquid biopsy in identifying EGFR-sensitizing mutations in NSCLC patients.Overall, the obtained results showed that LB has a sensitivity and specificity of 0.59 and 0.96, respectively.These values are in line with literature on NSCLC and other tumours (Zhu et al., 2020;Wang et al., 2021) and highlight an apparent low sensitivity of LB in detecting tumour mutations, while specificity appears high.However, when analyzing the manuscripts in detail, we noticed experimental variability among studies, especially regarding patients' tumour stage (high-stage or low-stage), study design (prospective or retrospective studies), and methods of mutation detection (NGS-or PCR-based methods).These differences were able to modify the diagnostic performances of LB.For this reason, the overall results appeared  highly heterogeneous, as indicated by I 2 > 80% with the random model and by the confidence intervals, that were very wide for sensitivity, specificity and DOR (Figure 2).To overcome the heterogeneity of the results, and obtain more informative data, we separated the 17 articles into different subgroups, based on the above-mentioned variables, and made statistical analyses on grouped studies.
When dividing results based on the tumour stage, we observed in the high-stage subgroup the highest sensitivity (0.75) coupled with the lowest heterogeneity ( I 2 = 0%), and the highest DOR (69.45 -Figures 3A-C).In detail, sensitivity ranged from 0.69 (Ding et al., 2019;Schrock et al., 2019) to 1.0 (Papadopoulou et al., 2019;Batra et al., 2022).Similarly, specificity showed high but variable values (from 0.89 to 1.0).As a consequence, DOR was also variable, ranging from 23.7 to 203.5.On the other hand, early-stage/locally advanced tumours showed the lowest sensitivity (0.27) and DOR (6.46), though specificity remained high (0.95-Figures 3D-F).Indeed, sensitivity was lower than 0.21 in 3 out of 7 cases, due to the high rate of FN samples and specificity was highly variable.Taken together, these findings indicate that the diagnostic performance of LB in NSCLC is influenced by the tumour stage and increases with the increase of tumour aggressiveness.It is conceivable that advanced-stage tumours, characterized by a high rate of apoptosis/necrosis, would release a higher amount of circulating DNA compared to early ones.Accordingly, literature data indicate that the ctDNA fraction varies based on tumour burden and stage, ranging from ≤0.01 to 0.1% in early-stage to ≥5-10% in advanced tumours (Vlataki et al., 2023).
Thus, LB could be properly used to follow the molecular evolution of the tumour over time and to monitor the response to treatment in advanced-stage tumours.Of note, even if in high-stage tumours LB sensitivity is high, it does not reach the same performance as TB.To increase LB sensitivity, blood samples could be repeated at different times or other biological fluids (e.g., saliva, urine, sputum) could be analysed as an alternative or in combination with plasma (Lino et al., 2023;Xin et al., 2023).On the other hand, LB does not seem to be indicated for early-stage tumours, showing a high rate of FN results, possibly due to the scarcity/lack of tDNA shed into circulation, or to the technical limits of current detection methods (Qiu et al., 2023).
The comparison between prospective and retrospective studies evidenced that the different experimental designs could generate discrepancies in the results, that need to be elucidated.In prospective cases, specificity had a mean value of 0.96, with values ranging from 0.86 (Lin et al., 2021) to 1.00 (Ito et al., 2018;Li et al., 2019;Sathapaty et al., 2021).Retrospective studies showed a specificity of 1.00 in all cases (except for Ito et al., 2022, in which the absence of TN cases resulted in a specificity of 0.00); however, specificity in the latter group was not realistic, since it was biased by the absence of false positive results.Sensitivity was higher in prospective than retrospective studies (0.62 vs. 0.55, respectively) probably due to the high proportion of FN in retrospective studies.Notably, both subsets showed high heterogeneity (I 2 = 94% and 91%, respectively).Overall, these results seem indicate that prospective studies are more reliable than retrospective ones, in defining sensitivity and specificity in real-world diagnostic workflow.
Another factor affecting LB diagnostic performance could be the method of mutation detection.At present, NGS and PCR-based techniques are employed for ctDNA detection; however, ctDNA assessment is hampered by its low amount in the bloodstream, requiring the need for even more sensitive techniques for detection and quantification.Comparing NGS with PCR-based methods we found similar sensitivity and specificity values whereas DOR, though very heterogeneous, was higher in the PCR-based subgroup.Of note, in the PCR-based group, the use of different techniques affects the performance.Overall, none of the methods can be considered optimal, since each one of them shows "pros and cons".NGS is widely used for the detection of ctDNA: it allows simultaneous sequencing of different genomic regions in many samples, it is also able to quantitate gene copy number variations, including gene amplification, and to identify chromosomal rearrangements such as oncogenic fusions.Besides, NGS is able to calculate the frequency of the variant allele (Rolfo et al., 2018;Fernandes et al., 2021;Pesta et al., 2022).NGS also allows the comprehension of tumour genetic features and provides crucial information on tumour microenvironment and immune response, that might drive the response to immunotherapy.(Qiu et al., 2023).PCR-based techniques are characterized by a short turnaround time and easy interpretation of results, but allow the detection of only one alteration at a time.Of note, they are very sensitive in the detection of low-frequency alleles; particularly, ddPCR (<0.01 of mutated alleles) and COBAS have proved adequate in the genetic profiling of tumours (Rolfo et al., 2018;Vlataki et al., 2023).
According to current guidelines (Rolfo et al., 2018), both NGS and PCR-based methods can be used by LB: in the presence of EGFR-sensitizing mutations, patients can initiate target therapy.However, given the low sensitivity of LB compared to TB, negative results should be considered non-conclusive and be implemented by other TB tests (Rolfo et al., 2018).To overcome this limitation, recently, Kwon et al. reported on the importance of integrating multi-omics data into machine learning analyses to significantly improve accuracy in cancer diagnosis (Kwon et al., 2023).

Conclusion
In conclusion, though LB has the potential to help the management of NSCLC in advanced-stage patients, at present non-homogeneous literature data, lack of standardized detection methods, together with relatively high costs, make its applicability in routine diagnostics still challenging.acquisition, Writing-review and editing.ST: Conceptualization, Data curation, Supervision, Writing-original draft, Writing-review and editing.

FIGURE 1
FIGURE 1PRISMA flowchart of the literature selection process.

FIGURE 3
FIGURE 3 Forest plot of the diagnostic performance of liquid biopsy in the high-stage (A-C) vs. low-stage/locally advanced (D-F) NSCLC subgroups.(A and D) Sensitivity (Se); (B and E) Specificity (Sp); (C and F) Diagnostic odds ratio (DOR).TP: true positive; FN: false negative; TN: true negative; FP: false positive.

FIGURE 4
FIGURE 4 Forest plot of the diagnostic performance of liquid biopsy in prospective (A-C) vs. retrospective (D-F) NSCLC subgroups.(A and D) Sensitivity (Se); (B and E) Specificity (Sp); (C and F) Diagnostic odds ratio (DOR); TP: true positive; FN: false negative; TN: true negative; FP: false positive.

FIGURE 5
FIGURE 5 Forest plot of the diagnostic performance of liquid biopsy in NGS-based detection (A-C) vs. PCR-based (D-F) method NSCLC subgroups.(A and ) Sensitivity (Se); (B and E) Specificity (Sp); (C and F) Diagnostic odds ratio (DOR); TP: true positive; FN: false negative; TN: true negative; FP: false positive.

TABLE 1
Characteristics of the 17 studies included in the meta-analysis.