Do Brain Labels Match Biology? Three Preprints Give Three Different Answers

Autism splits, and the split is genetic

The strongest yes comes from autism. A preprint reports that machine-learning-derived autism subtypes — labeled HSN, EBSN, and LSN — are clinically distinct and reproducible across independent cohorts, not artifacts of a single dataset. m:…4465

The subtypes also differ genetically in ways that make the split harder to dismiss. The EBSN subtype exhibits the highest SNP-based heritability, at 0.224, and polygenic scores for ADHD and depression are consistently associated with higher EBSN scores, while higher burdens of de novo variants and rare copy-number variants are associated with the HSN subtype in the SPARK cohort. Different subtypes, different genetic architectures — the kind of result that makes stratification worth pursuing rather than merely fashionable. m:…4465

Brain structure does not predict binge drinking

The second preprint is a caution. Its authors tested whether structural and diffusion MRI features could predict binge drinking, and the answer was weak: predictive performance was poor in the ABCD cohort, with an area under the curve of 0.585, and only stronger — 0.695 — in the NCANDA cohort. For a screening tool, that gap between cohorts is itself a warning. m:…1938

The authors went further and tested causality, not just prediction. Targeted maximum likelihood estimation showed no causal effect of the top-ten morphological brain predictors on binge drinking in either cohort. Brain differences associated with drinking are not, on this evidence, brain differences that cause drinking — a distinction that matters for any intervention aimed at the brain rather than at behavior and environment. m:…1938

Delirium after stroke leaves a measurable trace

A fourth preprint adds a label whose consequences are measurable even when its mechanism is not fully understood. Post-stroke delirium is associated with more severe global post-stroke cognitive impairment at 90 days post-stroke, according to the preprint. m:…1427

The stakes extend beyond cognition: post-stroke delirium is associated with worse functional outcomes, higher mortality, and lower mental quality of life at 90 days post-stroke. And it can be detected early — acute-phase kinematic parameters derived from tablet-based tasks show 78.89% discriminatory performance for post-stroke delirium. Here the label earns its keep not through genetics but through prognosis. m:…1427

Lewy body disease splits by diagnosis

The third preprint lands between the two. Five loci — APOE4, GBA1, SNCA, TMEM175, and HLA — are associated with autopsy-confirmed Lewy body disease, giving the diagnosis a real genetic footing. m:…6729

But the details split by diagnosis in ways a single label hides. APOE4 is the only susceptibility locus associated with neocortical involvement in Lewy body disease, independent of Alzheimer's co-pathology, and the SNCA gene tells two stories depending on the variant: its 5-prime signal is associated with dementia with Lewy bodies and REM sleep behavior disorder, while its 3-prime variant is associated with Parkinson's disease. One gene, different clinical destinations. m:…6729

Taken together, the three preprints argue against both hype and nihilism about brain stratification. Sometimes the label maps to biology, sometimes it does not, and sometimes it maps only once the label is split finer. All three are preprints, so each result still awaits peer review, but the shared discipline — test the category, don't just use it — is the point. A fourth, on post-stroke delirium, shows a label earning its keep through prognosis rather than genetics: worse cognition, worse function, and higher mortality at 90 days. m:…4465 m:…1938 m:…6729