We have systematically reviewed PET and SPECT studies of cortical 5-HT2AR binding in MDD. Two earlier reviews [27, 28] summarised differences in regional 5-HT2AR binding between MDD patients and healthy controls by comparing the medians and interquartile ranges of binding values from PET and SPECT studies. Our review extends this work by presenting the first meta-analysis of the PET and SPECT literature, and includes new case-control and within-subjects studies to attempt to separate illness from medication effects. We performed meta-analyses of the group variability of 5-HT2AR binding, which can help to interpret the results of meta-analyses of mean difference, especially if these are negative [29]. We also present the first systematic review and meta-analysis of post-mortem radioligand binding studies of cortical 5-HT2AR binding in MDD and suicide.
This review was registered with PROSPERO (CRD42019137947) in June 2019. Since registration, we decided to recruit additional reviewers to screen studies and extract data; to include studies with groups unmatched for gender; to include studies with groups unmatched for age, provided age was included as a covariate in analyses; and to systematically assess study quality. We follow the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) 2020 statement when reporting this work [30].
We systematically searched PubMed, EMBASE, PsycINFO and Web of Science databases from inception. We searched for all PET or SPECT studies of 5-HTR density in MDD on 8 February 2024 and updated these searches on 30 August 2024. We searched for all post-mortem radioligand binding studies of 5-HTR density in MDD and suicide on 6 September 2024. See Supplementary Figures 1 and 2 for the full search strategy for each database.
The titles of all identified studies, with or without abstracts, were manually screened by the lead author and one co-author (APN for PET and SPECT studies; GT for post-mortem studies), who then reviewed the full-text version of all potentially eligible studies. Any disagreements were resolved by discussion. See Figs. 1 and 2 for the inclusion and exclusion criteria and PRISMA flowchart for PET and SPECT studies. See Supplementary Figures 3-5 for the inclusion and exclusion criteria and PRISMA flowcharts for post-mortem studies. PET and SPECT searches also identified within-patient studies that looked at relationships between 5-HTR binding, antidepressant treatment and clinical outcome.
Mean (sub)regional binding values and standard deviations (SDs) were extracted for patient and control groups. We did not pre-specify regions for analysis; rather, all available cortical binding data were extracted. Binding data extraction was completed independently by the lead author and one co-author (APN for PET and SPECT studies; GT for post-mortem studies), who resolved any discrepancies by discussion. Additional information, such as demographic, medication and imaging details, was extracted by the lead author.
Where studies presented data for medicated and unmedicated subgroups of patients, these data were extracted and recorded separately. Where multiple studies reported binding values from the same subjects, only data from the larger study were included. Data that were only presented graphically were extracted using WebPlotDigitizer (https://automeris.io/WebPlotDigitizer/) -- a semi-automated data extraction tool with high intercoder reliability and validity [31]. We contacted study authors whenever data could not be retrieved from the text or graphs. Binding data were ultimately not retrieved for three eligible PET or SPECT studies [32,33,34] and one eligible post-mortem study [35].
(Sub)regional means and SDs were pooled. We pooled lateralised data from the left and right hemispheres. Then, if a study did not provide an overall binding value for a given region, subregional data were combined as per Supplementary Figure 6. Pooled means were calculated by taking the arithmetic mean of included means. Pooled SDs were calculated using the following formula for combining SDs of paired samples (see [36]):
Where is the pooled SD; and are the standard deviations of the paired samples a and b; and is the coefficient of the correlation between samples a and b, estimated to be 0.8. When we needed to pool three or more SDs, we pooled and to give then entered into the equation with , and so on.
Meta-analyses of standardised mean difference were undertaken. We employed Hedges' g, which is considered superior to other measures when sample sizes are small or uneven [37]. We used an inverse-variance random effects model due to moderate-to-high heterogeneity between some studies, as determined by the I statistic. We analysed in vivo MDD, post-mortem MDD and post-mortem suicide studies separately. A forest plot was constructed for each region for which there were case-control data available from at least three studies. We preferentially included unmedicated patient data (although most post-mortem analyses included medicated, unmedicated and mixed cohorts). If a study presented both medicated and unmedicated patient data, an additional regional forest plot was constructed containing medicated patient data. Exploratory forest plots were constructed for regions for which there were case-control data available from at least two studies in never-medicated MDD cohorts. Forest plots were also constructed to compare regional binding before and after antidepressant treatment in the same patients, wherever there were within-subjects data available from at least three studies. Finally, for post-mortem suicide studies only, we used forest plots to examine any effect of suicide method. If a study separately presented data from violent and non-violent suicides, two regional forest plots were constructed (one containing violent suicide data; one containing non-violent suicide data). If at least three studies presented both violent and non-violent suicide data for the same region, another forest plot was constructed, directly comparing violent with non-violent suicides. The Benjamini-Hochberg procedure can be used to adjust p values to correct for multiple comparisons. We applied this correction wherever the same binding data could have been included in two or more meta-analyses of mean difference.
We also meta-analysed group differences in the variability of 5-HTR binding, by calculating the variance ratio (VR) and coefficient of variation ratio (CVR). In accordance with published methods (see [29]), we first calculated the natural logarithm of VR and CVR -- lnVR and lnCVR -- which were later backtransformed to give VR and CVR. Both ratios compare within-group variance of an outcome, where VR or CVR > 1 indicates that there is greater outcome variability within the patient than control group. The larger the mean value of a biological variable, the greater the variance tends to be [38], to which VR is vulnerable but CVR is not; this is known as mean scaling [29]. We applied the Benjamini-Hochberg procedure wherever the same data could have been included in two or more meta-analyses of variance.
Negative studies are less likely to be published than positive studies, especially if they have smaller sample sizes. This publication bias was assessed using Egger's test [39] and by visual inspection of funnel plots. The risk of bias of case-control studies was assessed using the Newcastle-Ottawa Scale ([40]; see also [41]), which involved rating the selection, comparability and exposure of cases and controls. For comparability, we considered whether studies matched or adjusted for age and gender. If ascertainment of exposure was by clinical interview, investigators need not have been blinded to case or control status for a study to score highly if other processes were robust (e.g., structured interview, independent assessment by two clinicians).
Finally, for patient groups in case-control in vivo studies, we completed meta-regressions to investigate relationships between study-level regional binding effect sizes and depression severity (mean baseline Hamilton Depression Rating Scale (HDRS) score). Data from studies using the 17- or 21-item version of the HDRS were included, as both versions should provide the same mathematical total for the same patient [42]. Simple linear regression using Pearson's coefficient was performed for regions in which a significant difference in mean group binding was found by meta-analysis. We applied the Benjamini-Hochberg procedure wherever the same binding data could have been included in two or more meta-regressions.
All analyses were performed using the metafor package [43] within RStudio (v4.4.1, cran.r-project.org). Results were considered statistically significant if p < 0.05.