Long noncoding RNAs as novel predictors of survival in human cancer: a systematic review and meta-analysis

For meta-analysis eligibility, a study had to also provide the effect size and confidence interval for the association of an individual or group of lncRNAs with any of the above survival outcomes, or report information through which this effect size and confidence interval could be calculated [29, 30]. Wherever the same cohort had published more than one overlapping analysis, we only used the most encompassing data (for example, the classification of glioma would be preferred over glioblastoma multiforme). Two reviewers (S. Serghiou and A. Kyriakopoulou) identified eligible studies, and any contested articles were adjudicated by a third reviewer (J. P. A. Ioannidis).

We systematically searched PubMed (1950 to September, 2015) for studies of any language that analyzed associations between lncRNAs and prognosis in human cancer. Our search strategy was developed in consideration of previous recommendations [30] and used the clinical queries prognosis filter, which has been reported to have an average estimated sensitivity of 92 % for detecting articles related to prognosis [5, 31]. Our search term was: (Prognosis/Broad [filter]) AND ((lncRNA OR “lnc RNA” OR “long noncoding ribonucleic acid” OR “long noncoding RNA” OR “long non-coding ribonucleic acid” OR “long intergenic noncoding RNA” OR “long intergenic non-coding RNA” OR “long non-coding RNA” OR “long ncRNA” OR “lincRNA” OR “linc RNA”) AND (cancer OR carcinoma OR tumor OR neoplas* OR tumour OR malignan* OR metastat* OR metastas* OR leukemia OR leukaemia OR lymphoma OR recurren* OR “lymph node” OR response) AND (Humans[Mesh] AND English[lang])). The search was last updated to include articles published through September 26, 2015.

We used the programming language R [32] to remove duplicate records. Title and abstract were screened to identify relevant articles. The full manuscript of the relevant articles was screened against our eligibility criteria. Any uncertainties were resolved by consensus with JPA. Data were collected by two reviewers (SS, AK) and saved in a pre-designed extraction form on Google Sheets. Where information was ambiguous (such as, for example, mentioning multiple types of lncRNA quantification methods but not clarifying which one of those was used to provide the quantities utilized in the survival analysis), this was labelled as ‘unclear’. An attempt was made to contact the authors when information was clearly logically inconsistent, as in for example quoting a hazard ratio (HR) outside the confidence interval (CI), but none replied. In one paper, the lncRNA expression level [33] was subdivided into low versus medium versus high; for this paper we only extracted the comparison between low versus high expression levels. The following data were extracted for all articles following the CHARMS checklist [34]: title; authors; year of publication; journal of publication; groupings (i.e. whether lncRNAs were studied one by one or in groups); what lncRNAs were studied; whether an agnostic approach to identifying the studied lncRNAs was used (where an agnostic approach would be one assuming no prior knowledge regarding the choice of lncRNA to be studied); cancer site (e.g. brain) and cancer subtype (e.g. glioblastoma multiforme); whether a paper reported clinicopathologic data of its sample and which ones; whether an attempt of associating those clinicopathologic data to lncRNAs was made and for which ones; whether an attempt of associating clinicopathologic data to prognosis was made and for which ones; whether an attempt was made to explain the clinical outcomes using non-clinical studies (in vivo, in vitro); the types of survival analyses used (as above); type of study design (prospective cohort, retrospective cohort, unreported); means of lncRNA quantitative analysis (qRT–PCR, qPCR, in situ hybridization (ISH), other); and whether the paper tried to make any non-clinical associations of the identified lncRNAs to cancer in vitro. For eligible articles we further extracted: country and city of origin of the study cohort, period of sample recruitment, range of sample ages, mean/median age with confidence interval, the population type (general population, non-general population (e.g. veterans), unreported), stage of cancer upon initial patient presentation, sample size, means of tissue preservation (frozen, paraffin-embedded, both, other), any and what preoperative treatment was given, the total number of lncRNAs studied, the type of metric the paper used to characterize their results (hazard ratio, relative risk, odds ratio, p-value), type of analysis (i.e. univariable or multivariable), lncRNA quantity cut-off and its unit (i.e. the threshold based on which lncRNA expression was deemed upregulated or downregulated by the study), the sample size of each comparison group, the minimum and maximum participant follow-up time, the number of censored participants throughout follow-up and whether this was explicitly stated or read off the Kaplan-Meier curves, the HR and its CI (provided or inferred, e.g. from p-values and HR point estimates), the p-value and whether this was statistically significant at p < 0.05 and whether an attempt to validate the reported results was made, and if so, what type of validation method was used (internal or external). For eligibility for meta-analysis, enough information to extract or calculate the natural logarithm of the hazard ratio and its variance must have been provided.

Whenever multiple datasets were combined into a single dataset to study a specific lncRNA, we only extracted the summary HR, rather than extracting the HR respective to each constitutive dataset. If multiple datasets were assessed within the same study without being combined into a single dataset, we extracted the HR respective to each dataset, as they represent separate estimates. Where both the log-rank and Breslow tests were reported, only the log-rank was extracted. No cohort was used more than once and effect sizes describing a broader class of cancer (e.g. glioma) were preferred over subclassifications of that (e.g. glioblastoma multiforme). Three studies reported effect sizes that were excluded from further consideration because the quoted HRs contradicted the text [35] or they were either outside the CI or could not have possibly led to the quoted CI [36, 37]; this led to complete exclusion of two out of these three studies [35, 37]. Our database can be freely accessed here: https://​goo.​gl/​EjCDAp.​

Risk of bias in individual studies was assessed on the basis of the framework of assessing internal validity of articles dealing with prognosis [30, 38] and recommendations regarding reporting of biomarker studies [39, 40].

We meta-analyzed data on lncRNAs for which three or more estimates of their effect on a specific survival outcome were available. Therefore, meta-analyses were only done for OS and RFS. Effect sizes for OS and RFS were meta-analyzed separately. Our principal summary measure was the summary HR. Standard errors were calculated using: ln (upper limit of CI/lower limit of CI)/(2 × 1.96). Estimates were synthesized using a random-effects model and estimated using the restricted maximum-likelihood ratio method. As previously described [27], four meta-analyses were done for each of: (1) multivariable data, (2) univariable data, (3) multivariable data combined with univariable data whenever multivariable data were unavailable (preferentially multivariable) and (4) univariable data combined with multivariable data whenever univariable data were unavailable (preferentially univariable). Given the similarity between the estimates of all four types of meta-analysis and the importance of multivariable modelling in prognostic studies, this report only quotes the estimates of the ‘preferentially multivariable’ meta-analysis; the rest can be found in Additional file 1: Table S2. For each estimate we provide the effect size and 95 % CI. Heterogeneity was analyzed using the Q and I² statistics and the 95 % CI of I² was also calculated [41, 42]. These analyses were done using R and the package metafor 1.9-8 [43]. Data were combined for each type of lncRNA regardless of cancer type. Wherever an lncRNA had been analyzed three or more times for one or more specific cancer type, a post hoc subgroup analysis per cancer type was done for that lncRNA.

Risk of publication bias is a significant concern in prognostic studies [30]. We explored excess significance for factors reported by at least 3 studies [44]. Briefly, for every meta-analyzed risk factor we compare the number of observed significant results (O) at α = 0.05, to the number of expected significant results (E), where E = sum of power of each study within a specific meta-analysis. Power was calculated taking as plausible effect for the risk factor the effect seen in the most precise study (lowest standard error). The difference between O and E was assessed using a two-tailed binomial test, with α = 0.1, as previously suggested [45]. O and E were also summed and compared across all meta-analyses.

The eligible reports studied 18 types of cancer, top three most studied of which were gastric cancer (n = 16 datasets), lung cancer (n = 15) and colorectal cancer (n = 15) (Table 1). Almost half of the reports studied cancer related to the gastrointestinal tract (57/127 datasets). OS was assessed in 92/111 studies (83 %), RFS in 36 (32 %), DSS in 10 (9 %), MFS in 9 (8 %) and PFS in 6 (5 %). The majority of studies did not appear to choose what lncRNAs to study on the basis of agnostic reports (77 %, 85/111). For 98/127 datasets (77 %), there was no information regarding adjuvant treatment; for the 29 studies providing information regarding treatment, only 4 datasets indicated that their patients were treated homogeneously. In addition to survival analysis, 68 % (76/111) of the identified studies attempted to further study their chosen lncRNAs in vitro, to corroborate the results of their survival analyses with mechanistic insights into the function of their chosen lncRNAs. Across 66 studies reporting multivariable analyses, 42 adjusted for stage of cancer (or all three components of the TNM staging) and 27 for grade of cancer; only 19/66 studies adjusted for both. Figure 2 displays a microarray of the covariates that have been studied more than three times within multivariable analyses (Additional file 3: Figure S1 displays the complete data microarray). Out of all 66 studies, 20 (30 %) studies adjusted for the same factors as at least one other paper and the most commonly encountered combination of factors adjusted for was Stage and Lymph Node Metastasis, which was seen in 6/66 studies. The median number of adjustment combinations matching between at least two papers was 1 (IQR, 0).

Out of 92 studies reporting on OS, 87 studies (representing 111/127 analyses, as explained in Additional file 4: Table S1) provided effect estimates, out of which two were completely excluded due to reporting inconsistent effect sizes, as indicated in the Methods [35, 37]. The 85 remaining studies provided effect estimates on 53 lncRNAs and 6 multi-lncRNA risk score scales. The three most frequently studied lncRNAs within OS analyses were HOTAIR (n = 29 effect estimates), MALAT1 (n = 8) and GAS 5, H19 and PVT1 (n = 4 for each). Most individual lncRNAs (42/53) were only studied once (Table 2). Only 7 lncRNAs were studied at least three times in association to OS and for 6 of them more than half of the studies showed statistically significant p-values. These lncRNAs were studied in the context of a median of 4 different types of cancer (IQR, 3). Out of the 52 individual or groups of lncRNAs studied less than three times, 44 were always reported significantly associated to OS. Overall, of the 92 studies reporting on OS (but not necessarily quoting an effect estimate), 88 (96 %) reported at least one statistically significant result for association with prognosis.

A meta-analysis of OS was done for all 7 individual or groups of lncRNAs having been studied three or more times (Fig. 3; Table 3; Additional file 1: Table S2). For p-value < 0.0005, 5 lncRNAs were statistically significantly associated to OS in all of our meta-analyses (HOTAIR, MALAT1, 6 lncRNA risk score, PVT1, SChLAP1) and 6/7 were statistically significant in all of our meta-analyses at p-value < 0.05 (H19; Additional file 1: Table S2). An increase in cellular expression of these lncRNAs was statistically significantly associated to a decrease in overall survival; GAS5 was not statistically significantly associated to OS in our meta-analyses. The funnel plot for HOTAIR (Fig. 4), which is the only lncRNA studied 10 or more times, indicates significant small-study effects (p-value = 0.0006), and this may be suggestive of publication bias. The summary effect size for HOTAIR also displays a moderate amount of between-study heterogeneity (I², 48 %; 95 % CI, 14–78 %). The summary effects for the effect of HOTAIR on OS in cancers for which it was studied 3 or more times were: colorectal cancer (HR, 4.76; 95 % CI, 2.46–9.21), esophageal cancer (HR, 2.29; 95 % CI, 1.68–3.12) and glioma (HR, 1.71; 95 % CI, 1.25–2.34).

Statistically significant heterogeneity was only observed in HOTAIR analyses, but substantial estimates of I² were common. For HOTAIR and OS, a sensitivity analysis excluding the only study reporting an inverse correlation of HOTAIR to cancer survival [33] generated a HR of 2.30 (95 % CI, 1.97-2.70) with I² = 0 % (95 % CI, 0–59 %); for all other meta-analyses, no single study produced a major change in the I².

There was excess significance across the whole field for overall survival and the binomial distribution revealed a two-tailed p-value of 0.0003, with O = 42 statistically significant results and E = 30 expected statistically significant results across all meta-analyses with 3 or more studies each on OS. As far as excess significance within lncRNAs studied 5 or more times is concerned, there was significant excess significance documented for HOTAIR (p-value = 0.002), but not MALAT1 (p-value = 0.46).