Published on 16 September 2026
Schizophrenia: is synaptic density loss on PET independent of the atrophy seen on MRI?
In brief
Twenty-nine patients with schizophrenia, ninety-three healthy controls, and a positron emission tomography scanner able to measure a synaptic vesicle protein, and so to estimate the density of presynaptic terminals in the living brain. Tracer binding is diffusely reduced, with medium to large effect sizes (Cohen’s D from 0.58 to 1.47), and a clear left-sided predominance (D = 1.14; p < 0.001). The result that matters to the clinician lies elsewhere: this reduction does not correlate with the volume loss measured by MRI in the same brains (r = 0.03; p = 0.496), nor with the volume loss maps of four independent cohorts. In other words, what MRI shows and what the synapse undergoes are not the same thing. In the left hemisphere only, the spatial distribution of the loss aligns with that of several neurotransmitter systems, GABAergic, serotonergic and metabotropic glutamatergic, and not with that of the dopamine maps. The sample is small, the study cross-sectional, the illness chronic, and the patients exposed to antipsychotics over the long term.
The context
Two models have coexisted for a long time. The first is dopaminergic, inherited from pharmacology: antipsychotics block D2 receptors, so the illness must have something to do with dopamine. The second is morphological, inherited from structural imaging: a reduction in cortical volume is observed, so there must be atrophy.
Neither says what happens at the scale of the synapse, because no method measured it in the living brain. That is the lock the tracer used here opens. The authors point out, moreover, that MRI-derived morphometric measures index a composite of tissue properties, intracortical myelin, neurite density, glial density, iron, water, vasculature, and not synaptic terminals. The question then becomes testable: is the volume loss seen on MRI a reflection of synaptic loss, or are the two processes distinct?
The mechanism
The target protein, synaptic vesicle protein 2A, sits in the vesicle membrane, monodispersed across all chemical presynaptic boutons. The amount of tracer bound in a region therefore gives an estimate of the density of presynaptic terminals located there. It is a direct measure in the sense that it does not go through a structural intermediary, but it remains an estimate of a population of synapses, not a count.
What the method does not measure is worth stating. It does not distinguish excitatory from inhibitory synapses, it says nothing about the function of the remaining synapses, and it cannot tell whether a low density results from elimination, from a failure of formation or from remodelling. The authors themselves note that a lowered signal could reflect a reduction in protein 2A or in vesicular content without any reduction in the number of terminals. The correspondence between the loss map and the receptor maps is established by comparison with atlases built from other brains, which adds a layer of approximation.
The study at a glance
| Population | |
| Twenty-nine adults meeting DSM-5 criteria for schizophrenia, mean age 43.2 years, 13.8% women, mean duration of illness 15.2 years, mean total PANSS 68.4, mean cumulative antipsychotic exposure of 2693 chlorpromazine equivalents and mean daily dose of 443 chlorpromazine equivalents. Compared with ninety-three healthy controls drawn from other protocols at the same centre, better balanced for sex (39.8% women; p = 0.018). Recruitment at a single university site. | |
| Measurement | |
| Positron emission tomography with [11C]UCB-J, a tracer of synaptic vesicle protein 2A, on a high-resolution scanner, with frame-by-frame motion correction, partial volume correction by the Müller-Gärtner method, and quantification by binding potential using a simplified reference tissue model, with the centrum semiovale as the reference region. T1-weighted structural MRI acquired on a 3 tesla scanner, used for tissue segmentation and for voxel-based morphometry. | |
| Comparator | |
| Healthy controls for the group comparison, then eighteen normative receptor and transporter maps from a public repository, twenty-four maps of cell-type densities deconvolved from transcriptomic data, and a reference structural connectome built from one hundred and fifty healthy participants in the Human Connectome Project. | |
| Outcomes | |
| Between-group difference in binding potential, tested voxel by voxel with non-parametric family-wise correction, then parcellation into 332 regions (300 homotopic cortical, 32 subcortical) for the spatial analyses: hemispheric asymmetry, correlation with the volume loss map, canonical correlation with the receptor and cell-type maps, identification of epicentres by a network diffusion model. | |
| Design | |
| Cross-sectional case-control study, Oxford level of evidence 3b. Group-level analysis, no longitudinal data, no causal inference possible. |
The quality check
| Criterion | Status |
|---|---|
| Nature of the measurement | Direct, in vivo |
| FindingThe tracer estimates the density of presynaptic terminals without going through a structural marker. The authors point out that, in primates, regional levels of the target protein are strongly correlated with synaptophysin, a validated ex vivo marker of presynaptic density. No conventional imaging provides this information, and that is what makes the work worthwhile. | |
| Size of the patient group | Small |
| FindingTwenty-nine patients. The authors justify this sample size from earlier studies with the same tracer, which estimated nine participants per group to be sufficient for a hippocampal difference and sixteen to twenty-one for frontal and anterior cingulate effects, adding that they treat these figures as lower bounds since the object here is a spatial pattern and not a region of interest. An exploratory analysis of the link between synaptic density and PANSS scores left no association surviving correction for multiple comparisons. | |
| Antipsychotic exposure | Major confounder, partly tested |
| FindingThe sample is chronic and heavily exposed. The authors repeated the main analysis adjusting for medication exposure, and found a very similar pattern, significantly correlated with the original one. Statistical adjustment does not replace an unexposed group: they write that they cannot determine whether the findings reflect the illness, its progression, prolonged treatment exposure or a combination of these factors, and that longitudinal work in antipsychotic-naive participants will be needed. The source does not state whether all twenty-nine patients were receiving an antipsychotic at the time of the scan. | |
| Image processing | Careful |
| FindingMotion correction by optical tracking, partial volume correction, non-parametric family-wise correction by permutation. Partial volume correction matters here, since it limits the risk that grey matter loss mechanically produces an apparent drop in signal. The authors show that the main result also holds when this correction is not applied. | |
| Recruitment site | Single site |
| FindingOne centre, one scanner, one team. Thirteen of the twenty-nine patients had already been included in an earlier publication by the same group. No independent replication of the synaptic measurement accompanies the work. On the other hand, the absence of correspondence with volume loss was indeed tested against the maps of four independent MRI cohorts. | |
| Comparison with atlases | Approximation |
| FindingThe eighteen receptor and transporter maps come from healthy populations and from brains other than those of the participants. The authors state this explicitly as a limitation. Correlating a loss map with an external reference map remains a comparison of regional means, not a measurement made in the same people. | |
| Funding and competing interests | Academic |
| FindingPublic and foundation funding, declared author by author. The authors declare no competing interests. Open access publication. | |
The findings
| Outcome | What the data show |
|---|---|
| Extent of the reduction | Cohen’s D from 0.58 to 1.47, left predominance (D = 1.14; p < 0.001) |
| PEB readingThe reduction involves frontal, temporal, cingulate, thalamic, striatal, hippocampal and cerebellar areas. Involvement this diffuse points towards a generalised process rather than a circuit lesion. The authors report no count of significant regions: significance is established at voxel level, and the parcellation into 332 regions serves only the spatial analyses that follow. The left asymmetry is consistent with consortium work on left-predominant cortical thinning in language regions, although this study cannot establish it. | |
| Link with volume measured on MRI | Correlation r = 0.03, p = 0.496 |
| PEB readingThis is the most important result of the paper. A near-zero correlation between measurements made in the same patients is not easily put down to a lack of power, all the more so as the result holds hemisphere by hemisphere, without partial volume correction, and with deformation-based rather than voxel-based morphometry. It also holds against the maps of four independent cohorts covering several stages of illness, whereas those volume maps do correlate with one another (0.33 < r < 0.41). A word of caution on wording, however: the absence of correlation between two measures does not demonstrate that two biological processes are independent, it shows that they do not vary together at the regional scale examined. The authors themselves write that this does not rule out partial convergence in some regions. | |
| Correspondence with neurotransmitter systems | Multivariate association in the left hemisphere, rcca = 0.68; p = 0.022 |
| PEB readingThe association was tested by canonical correlation between the regional synaptic loss map and eighteen normative receptor and transporter maps. Four maps carry reliable loadings after correction: the GABA type A complex and its benzodiazepine site, the serotonin 2A receptor, the metabotropic glutamate receptor type 5 and the serotonin 1B receptor. The value of 0.68 is that of the canonical variate as a whole, not that of any single receptor: receptor-by-receptor correlations are not reported as figures in the text of the article. No significant association is found in the right hemisphere, nor with cell-type densities, in either hemisphere. These are spatial correlations between average maps, not individual receptor measurements. | |
| Correspondence with the dopamine system | No association with dopamine receptors or the dopamine transporter |
| PEB readingThis result needs to be read with caution. The authors themselves offer a technical explanation: their spatial analysis covers the whole brain and is dominated by cortical variation, which overlaps poorly with the essentially striatal distribution of the dopamine maps. They also point out that dopaminergic abnormalities in schizophrenia are thought to reflect presynaptic functional dysregulation rather than regional loss of receptors or synapses. The absence of spatial correspondence therefore says nothing about the functional involvement of dopamine in the illness, nor about the efficacy of antipsychotics. A spatial correspondence and a mechanism of action are two different things. | |
| Network epicentres | Left frontal operculum and anterior insula (rmax = 0.54), left middle temporal gyrus (rmax = 0.53) |
| PEB readingA network diffusion model identifies the regions whose position in the connectivity best accounts for the observed distribution. These two regions are the only ones to survive family-wise correction against two null models, one based on spatial rotation of the maps, the other on rewiring of the connectome. An exploratory variant, in which diffusion is weighted by similarity in receptor composition, again points to the left frontal operculo-insular region. This is a model, not an observation. Nothing indicates that the loss began there, and the authors speak of potential sources. | |
Critical appraisal
| Domain | Judgement |
|---|---|
| Participant selection | Selective |
| FindingTaking part in a positron emission tomography study requires clinical stability and the capacity to cooperate. The mean total PANSS is 68.4, and only 24% of patients are described as acutely psychotic or moderately ill according to the threshold used by the authors. The most severely ill patients are therefore probably under-represented. | |
| Confounding factors | Partly addressed |
| FindingAntipsychotic exposure and substance use were the subject of sensitivity analyses, adjustment for the former, exclusion of seven participants for the latter, with stable results. What remains is the sex imbalance between the groups, 13.8% women among patients against 39.8% among controls, which the authors acknowledge as a limitation despite adjustment for sex, and the chronicity of the illness, fifteen years’ duration on average. A cross-sectional design cannot disentangle illness, duration of illness and treatment. | |
| Statistical analysis | Appropriate |
| FindingGeneral linear models adjusted for age and sex, non-parametric family-wise correction at voxel level over five thousand permutations, null models preserving spatial autocorrelation for all map-to-map comparisons, Bonferroni correction, robust non-parametric canonical correlation with bootstrap estimation of standard errors, two independent null models for the diffusion model. The correction chosen is family-wise, more conservative than false discovery rate control. The statistical method is not the weak point of this work. | |
| Fit between claim and evidence | Measured |
| FindingThe dissociation between synaptic signal and volume is directly measured, and the authors do not present it as proof of two distinct diseases: they write that the two measures index related but non-equivalent processes, dissociable at the whole-brain level. The correlations with atlases are presented for what they are, and the diffusion model is described as a simulation nominating potential sources. | |
| External validity | Limited |
| FindingTwenty-nine patients from a single centre, chronically ill, mostly men, exposed to antipsychotics over the long term. The result calls for independent replication, particularly in first-episode patients not exposed to antipsychotics, which the authors themselves call for. | |
| Immediate clinical relevance | None |
| FindingThis imaging technique belongs to research. The publication uses it only within a research protocol and documents no clinical use, and no clinical decision is made on the basis of it. | |
Level of evidence
Confidence is high that the synaptic signal and cortical volume do not vary together in these patients. The measurement is direct, made in the same people, with a partial volume correction that neutralises the mechanical effect of volume on the signal, and the result withstands several methodological variants as well as comparison with four independent MRI cohorts.
It is moderate for the extent and magnitude of the reduction. Twenty-nine chronically ill patients, at a single site, thirteen of them already published: that makes an estimate that is useful and fragile at the same time.
It is low for the neurochemical interpretation. The association exists only in the left hemisphere, it rests on a single multivariate statistic whose p value is close to the threshold, and it correlates a loss map with atlases built from other brains. That produces a hypothesis, not a demonstration. And reading the absence of spatial correspondence with the dopamine maps as a challenge to the role of dopamine would be an inference these data do not support, especially since the authors offer a purely methodological explanation for that absence.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“ What strikes me is that MRI does not see what is going on. We spent twenty years measuring volumes and thinking we were measuring the illness. That said, twenty-nine chronic patients, fifteen years of illness and years of treatment behind them: I will never know whether I am looking at schizophrenia or at the antipsychotic. And I will never have that scanner anyway. ”
What this means in practice: nothing to change today, but a mental model to correct. A normal MRI in a patient with psychosis does not mean the cortex is fine, and the absence of a structural abnormality offers no reassurance about anything.
What you can do with this
- Stop treating the volume loss seen on MRI as an indicator of what is happening at the synapse. On these data, the two do not vary together.
- Be ready to answer a patient or a family who asks why the MRI is normal. Normal structural imaging does not rule out brain involvement, it simply does not explore the right scale.
- Do not conclude from this study that dopamine no longer has a role in schizophrenia, nor that antipsychotics are poorly targeted. This work tests neither of these questions.
- Bear in mind that synaptic imaging is a research tool: this publication uses it only within a research protocol and describes no diagnostic use for it.
- Watch for replication in first-episode patients not exposed to antipsychotics. It is the only route that will separate the effect of the illness from that of treatment, and it is the one the authors recommend.
Frequently asked questions
Is synaptic density PET used in routine clinical practice?
Not on the basis of this publication, which uses it only within a research protocol. It belongs to research and to the few centres with the necessary equipment, and the publication describes no diagnostic indication for it. No patient should be offered it outside a research protocol.
Is schizophrenia a disorder of the synapse rather than of dopamine?
The study does not set these two levels against each other. It measures synaptic density and finds that, in the left hemisphere, its spatial distribution resembles that of certain neurotransmitter systems more than that of the dopamine maps. The authors in fact attribute this absence in part to their analysis being dominated by the cortex, whereas the dopamine maps are mainly striatal. A spatial correspondence is not a mechanism, and nothing here says that antipsychotics act in the wrong place.
Do antipsychotics damage synapses?
This study cannot settle the question. The sample is chronic and exposed over the long term, so the effect of treatment and that of the illness cannot be separated. The authors repeated the analysis adjusting for medication exposure and found the same pattern, and one of the studies they cite reports that the marker used here is unaffected by antipsychotics in rats. That is not enough to conclude in humans, and the authors call for studies in treatment-naive participants. In any case, there is no argument here for changing or stopping a treatment.
Why does the lack of correlation with MRI volume matter so much?
Because it challenges a common shortcut, the one that treats volume loss as the image of neuronal or synaptic loss. On these data, the two measures do not track each other, and this also holds against four independent MRI cohorts covering several stages of illness. It does not establish two independent processes, but it rules out using one to speak for the other.
Are twenty-nine patients enough for a synaptic PET study?
For a large group effect measured with a precise technique, yes, provided one goes no further. The authors rely on earlier estimates of nine to twenty-one participants per group depending on the region studied, which they present as lower bounds since their object is a spatial pattern. This sample size does not allow the signal to be linked to clinical severity, as the exploratory analysis on the PANSS yielded nothing after correction, nor subgroups to be compared, nor the result to be generalised to all patients.
Annotated bibliography
Source study. Chopra S, Worhunsky PD, Naganawa M, Zhang XH, Segal A, Labache L, Orchard E, Cropley V, Wood S, Angarita GA, Cosgrove K, Matuskey D, Nabulsi NB, Huang Y, Carson RE, Esterlis I, Skosnik PD, D’Souza DC, Holmes AJ, Radhakrishnan R. Widespread synaptic density loss in schizophrenia follows molecular and network architecture. Molecular Psychiatry. 2026. Published online, with no volume, issue or page numbers at the date of verification. doi: 10.1038/s41380-026-03717-x. Received 3 November 2025, revised 24 May 2026, accepted 16 June 2026, published open access under a Creative Commons Attribution 4.0 licence.
Funding and competing interests. Support is declared author by author: S.C. by the University of Melbourne (McKenzie Fellowship) and the Brain & Behavior Research Foundation; I.E. by the Nancy Taylor Foundation and the VA National Center for PTSD; D.M. by the National Institute of Neurological Disorders and Stroke (R01NS124819); A.S. by the Yale Wu Tsai Postdoctoral Fellowship; V.C. by the National Health and Medical Research Council of Australia (Investigator Grant 1177370) and the University of Melbourne (Dame Kate Campbell Fellowship); G.A.A. by the National Institute on Drug Abuse (R01 DA052454-03); K.C. by the National Institute on Alcohol Abuse and Alcoholism (U54 AA027989); R.R. by the Dana Foundation (David Mahoney Program), the National Center for Advancing Translational Sciences (UL1 TR001863) and the National Center for Homelessness Among Veterans (36C24820Q1276); A.J.H. by the National Institute of Mental Health (R01MH120080). Open access funding is provided by CAUL and its member institutions. On competing interests, the publication states that the authors declare no competing interests. The PubMed identifier 42350786 was established on 31 August 2026 by querying bibliographic registries, as the document consulted carried none. The supplementary material was consulted: it details the construction of the reference connectome (one hundred and fifty Human Connectome Project participants, optimised probabilistic tractography, consensus threshold of 37.7%, weighted 332 by 332 matrix), the parameters and the two null models of the network diffusion model, the construction of the receptor and transporter similarity matrix from eighteen maps, and the legends of the seven supplementary figures. The supplementary figures themselves are images and were not readable in the extracted text: the numerical values they carry, in particular the receptor-by-receptor canonical loadings, the distribution of antipsychotic types at the time of the scan and the sensitivity analysis maps, could therefore not be checked figure by figure. The mathematical symbols of the diffusion equation are also missing from the extracted text.
Context references. These references are those cited by the publication and from which it draws the elements used here. Finnema SJ, Nabulsi NB, Eid T, Detyniecki K, Lin SF, Chen MK, et al. Imaging synaptic density in the living human brain. Sci Transl Med. 2016;8:348ra96. Contribution: this is the validation work for the tracer used here, from which the publication takes the regional correlation above 0.95 with synaptophysin in primates. Limitation: this validation was carried out in animals and on small numbers, and it does not guarantee the correspondence in ill humans. Radhakrishnan R, Skosnik PD, Ranganathan M, Naganawa M, Toyonaga T, Finnema S, et al. In vivo evidence of lower synaptic vesicle density in schizophrenia. Mol Psychiatry. 2021;26:7690-8. Contribution: the first in vivo demonstration of lowered vesicular density in schizophrenia. Limitation: sample overlap, as thirteen of the twenty-nine patients in the current study come from this work, which rules out treating it as an independent replication. Onwordi EC, Halff EF, Whitehurst T, Mansur A, Cotel MC, Wells L, et al. Synaptic density marker SV2A is reduced in schizophrenia patients and unaffected by antipsychotics in rats. Nat Commun. 2020;11:246. Contribution: this is the only readable item in the publication that directly addresses the effect of antipsychotics on the marker used, and its title indicates an absence of effect. Limitation: the experimental arm is a rodent model, not transferable as it stands to the chronically ill patient. Chopra S, Segal A, Oldham S, Holmes A, Sabaroedin K, Orchard ER, et al. Network-based spreading of gray matter changes across different stages of psychosis. JAMA Psychiatry. 2023;80:1246-57. Contribution: this is the work whose diffusion model is reused, applied there to grey matter volume, and which identified the anterior hippocampus as the epicentre. Limitation: same team, same family of models, and epicentres different from those found here, which the authors interpret as consistent with the dissociation they report. Toyonaga T, Khattar N, Wu Y, Lu Y, Naganawa M, Gallezot JD, et al. The regional pattern of age-related synaptic loss in the human brain differs from gray matter volume loss: in vivo PET measurement with [11C] UCB-J. Eur J Nucl Med Mol Imaging. 2024;51:1012-22. Contribution: a dissociation of the same kind between synaptic loss and volume loss, but in ageing. Limitation: it concerns individuals without psychiatric illness, so it supports the plausibility of the result without replicating it. Schijven D, Postema MC, Fukunaga M, Matsumoto J, Miura K, de Zwarte SMC, et al. Large-scale analysis of structural brain asymmetries in schizophrenia via the ENIGMA consortium. Proc Natl Acad Sci. 2023;120:e2213880120. Contribution: this is the reference on which the publication bases the consistency of the left asymmetry with left-predominant cortical thinning in language regions. Limitation: it is structural MRI, precisely the modality that this work shows does not track the synaptic signal.