🧵 New publication from the PGC Anxiety Working Group. Our GWAS meta-analysis of anxiety disorders is now published in @natgenet.nature.com! 🔗: doi.org/10.1038/s415...

Big shout out to @clarajiang.bsky.social for leading this work, to all co-authors in France, Australia and UK, to @iesinria.bsky.social, @institutducerveau.bsky.social, #UQ, #IMB, as well as the Inria Associate team funding who made the collaboration possible by funding Clara' exchange in France

We concluded that depression does not have a strong signature in the grey matter, and that other brain MRI sequences (e.g. diffusion, or functional MRI) may be required to accurately predict depression.

A unified framework for association and prediction from vertex‐wise grey‐matter structure This manuscript introduces a set of analyses, that rely on linear mixed models to perform association and prediction, while being suited to tackle the challenges of big-data in neuroimaging. Our fram....

In our article, we estimated the morphometricity of depression to be 6%, which means that only 6% of the disease status variance is captured by all brain measurements. This implies that an optimal linear predictor would only reach an AUC=0.64.

See link for more on morphometricity (and BLUP)

Classification of major depressive disorder using vertex-wise brain sulcal depth, curvature, and thickness with a deep and a shallow learning model - Molecular Psychiatry Molecular Psychiatry - Classification of major depressive disorder using vertex-wise brain sulcal depth, curvature, and thickness with a deep and a shallow learning model

Our results highlight the difficulty to predict depression from structural brain MRI, which was previously reported in papers from the ENIGMA-MDD consortium.

Interesting validation: our BLUP predictor was associated with a polygenic risk score for depression, but also captured additional information not currently tagged by the genetic predictor.

Notably, the simple and efficient BLUP predictor outperformed the Deep Learning ResNet predictor, which is not unusual for brain-based predictors with these sample sizes.

The main result is that brain-based predictors can do better than chance, but prediction remains low (OR=1.28; AUC=0.57), too low for clinical applications (e.g., aid diagnosis, referral).

Happy to share our latest paper, the work of @clarajiang.bsky.social who trained brain-based predictors of depression on the UK Biobank (N=7,500 curated cases and matched controls) using AI/deeplearning as well as efficient statistical learning (BLUP).

See thread below for a summary of findings

NeuroStats2026

Save the date 👆🧵 Neuroimaging Statistics Workshop, 12-13 June, @ohbmofficial.bsky.social 2026 satellite meeting, w/ @ohbmossig.bsky.social's BrainHack. Keith Worsley lecture by Sir. John Aston, + great work at Neuro/Stats/ML/AI interface. Registration opens soon! sites.google.com/view/nsw2026

What other criteria should we use to benchmark processing? Elise also quantified the carbon footprint associated with computation in her PhD thesis.

Still, we hope our results can help researchers make informed decisions when selecting a processing. And that it can encourage others to evaluate other image processing options (e.g., different software versions, non-default settings).

Many questions remain. Would we see the same results in other age groups (e.g., ABCD data) or in other databases (we used @ukbiobank.bsky.social)?

To make things even more complicated, it may also depend on the trait you are interested in. On average FSL VBM does a good job, but for some traits like Alzheimer's disease or maternal smoking, FreeSurfer may yield more significant associations.

We also demonstrated that each processing step captures a unique signal. So we cannot claim that a single processing (among the ones we considered) is the best. Some are better, but none is perfect.

Elise's results also confirm that the choices of processing have important consequences on the results, which contributes to the reproducibility crisis. See results below for the same individuals and same trait (maternal smoking around birth - associated regions vary depending on processing)

Choice of Processing Pipelines for T1-Weighted Brain MRI Impacts Association and Prediction Analyses - PubMed The growing availability of large neuroimaging datasets, such as the UK Biobank, provides new opportunities to improve robustness and reproducibility in brain imaging research. However, little is known about the extent to which MRI processing pipelines influence results. Using 39,655 T1-weighted MRI …

Results (in brief) show that FSL VBM performs very well, capturing the most signal (morphometricity), yielding performant predictors, maximising power to detect associated brain regions, and yielding replicable results.
pubmed.ncbi.nlm.nih.gov/41163627/

I have always wondered if there is a best method/software to process T1w image (at a vertex/voxel wise level). Elise Delzant has some answers in her first published paper from her PhD. She compared FreeSurfer, CAT12 and FSL across several traits and analyses (association, prediction).

Post-doc job offer with Sonia Shah in Brisbane, working in statistical genetics within a great lab and ecosystem.
www.linkedin.com/jobs/view/41...

@institutducerveau.bsky.social

Lastly, we showed that the brain-based predictor of Alzheimer's could also predict early AD, mild cognitive impairment, cognition, tau levels or genetic risk. Our predictor is on par with some of the best ones that can predict progression to AD, which could help with early intervention

Interestingly, the brain regions associated with Alzheimer's, disease progression and cognition overlapped suggesting some of the same regions are implicated. Some regions stand out being associated with many traits.

We systematically attempted to replicate the results also showed the identified regions could help predict disease, progression and cognition in independent samples.

We have also estimated that there should be at least 5 times more regions that remain to be identified (difference between the morphometricity and variance currently explained by the identified regions). Even more data will be needed!

Even better with a GIF (showing the regions with reduced thickness associated with Alzheimer's across the left subcortical structures)! Made with our R package brainMapR. github.com/baptisteCD/b...

Our high-resolution approach makes it possible to identify sub-regions of the hippocampus (or of the amygdala) that are involved, which can orient research towards specific brain regions/networks or cell populations.

This is the largest study of its kind in terms of the number of participants (>9,000 subjects from 10 clinical cohorts+ the UK Biobank), and we have identified 103 grey matter regions associated with Alzheimer's, including some in ‘known’ regions, such as the hippocampus.

Grey‐Matter Structure Markers of Alzheimer's Disease, Alzheimer's Conversion, Functioning and Cognition: A Meta‐Analysis Across 11 Cohorts We meta-analysed results from 11 neuroimaging cohorts of elderly participants, to estimate the morphometricity (total association with brain measurements) of 17 traits of interest, and to identify th...

Our latest paper is out! A well powered high resolution mapping of the grey-matter regions associated with Alzheimer's, progression to Alzheimer's disease, cognition and functioning. @drbreaky.bsky.social @barkhof.bsky.social @michelleklupton.bsky.social

At this point, we do not know if the identified drugs or conditions directly cause neurodegenerative disorders or whether they are early symptoms of the disease or reflect more complex aetiology. However, the results could be useful for personalising and accelerating referrals to neurologists.