A reader’s guide to this study

Schizophrenia runs strongly in families, and large genetic studies have traced part of that heritability to hundreds of specific regions of the genome. Almost none of those regions contain a protein-breaking mutation. The standard explanation is that the risk variants instead adjust how strongly nearby genes are switched on or off in the brain. If that explanation is right, risk variants should coincide with variants already known to change gene activity in brain tissue. Mostly, they do not. That mismatch has been called the missing regulation problem^[3], and this project measures it, tests explanations for it, and asks which kind of molecular evidence closes the gap.

A genome-wide association study compares the genomes of people with and without a condition at millions of variants, single positions where DNA differs between people, and flags the variants that are reliably more common in one group. The schizophrenia study used here, PGC3^[12], compared 67,390 people with schizophrenia against 94,015 without and reported 287 significant regions; 281 of them are analyzable and appear on this site.

Each region, or locus, is a stretch of DNA containing many variants that travel together through generations because they are inherited as a block. That correlation, called linkage disequilibrium, means a GWAS can say a region matters but usually cannot say which single variant is responsible, and the flagged index variant is best read as a bookmark. A locus typically spans several genes, so the question “which gene does this locus act through?” has no automatic answer.

The large majority of risk variants are non-coding: they sit outside the parts of genes that specify proteins, so they cannot change a protein’s composition. What they can plausibly change is gene expression, the amount of RNA a gene produces. Genetics has a tool for finding such effects: an eQTL is a variant associated with the expression level of a nearby gene, measured across hundreds of donated tissue samples. Public brain eQTL catalogs^[6][8][4] cover 14 adult brain datasets used here, from 13 GTEx regions to the MetaBrain cortex meta-analysis of roughly 2,700 effective samples.

Finding a risk variant and an eQTL in the same region is not enough, because LD makes coincidental overlap common. Colocalization^[5] compares the full shape of the two association signals and asks which is more likely: one shared causal variant driving both, or two distinct variants that happen to be neighbors. The answer arrives as posterior probabilities, and the one that matters here is PP4, the probability of a shared causal variant. This site counts a result as colocalized when PP4 exceeds 0.8.

One caution applies everywhere on this site: colocalization is evidence that two signals share a cause, not proof that the molecular trait carries the disease effect. A shared variant might influence the disease through some other route entirely.

Applied to all 281 loci and all 14 bulk brain expression datasets, the direct test succeeds at 90 loci, or 32%. The failures are not explained by an absence of eQTLs: 94% of the non-colocalizing loci harbor a strong eQTL for some gene nearby, but the eQTL and the risk signal favor different variants. Larger eQTL studies help, roughly 2.5-fold higher odds of colocalization per tenfold increase in sample size, yet the trend flattens near 25% for single cortex datasets, so more of the same data will not close the gap on its own.

The project then widens the search beyond adult bulk expression. Six evidence channels are applied in a fixed order, from the most direct to the most indirect, and each locus is credited to the first channel that explains it. This sequential attribution is what the waterfall figure on the dashboard shows.

• 901. Bulk expression. eQTLs from 14 adult brain datasets: 13 GTEx regions plus the MetaBrain cortex meta-analysis. The most direct route from variant to gene.
• 232. Multi-signal rescue. The same expression data retested after fine-mapping separates each locus into independent signals, catching loci the single-variant test misjudged.
• 63. Single-cell eQTL. eQTLs measured within eight brain cell types^[1], recovering effects that bulk tissue averages away.
• 44. Fetal eQTL. eQTLs from developing cortex^[13], testing whether risk variants act during development rather than in adult tissue.
• 95. Splicing. Splicing QTLs from 17 datasets, catching variants that change transcript form without changing total expression.
• 846. Methylation. Brain methylation QTLs from roughly 1,160 brains^[10], a readout of regulatory activity that does not require knowing the target gene.

The counts above are loci newly explained by each channel. The expression-based channels together account for 132 loci. Methylation, tested last, adds 84 more, the largest single contribution; tested on its own it colocalizes at 193 of 281 loci, including 118 of the 191 loci that bulk expression missed, and its per-test success rate is six times that of expression. The final tally is 216 of 281 loci explained (77%), leaving 65 (23%) unexplained by every channel. For loci with no expression evidence at all, a two-step methylation chain nominates candidate genes; 44 such loci received a high-tier nomination.

• ✓Most schizophrenia risk loci cannot be tied to a gene through adult bulk brain eQTLs at strict thresholds; expression data of all kinds explain roughly half of the loci.
• ✓The regulatory hypothesis itself holds up: brain methylation QTLs colocalize at most loci, placing the risk variants on active regulatory DNA even where expression data are silent.
• ✓The shortfall lies mostly with current expression catalogs, their sample sizes, tissues, cell types, and developmental windows, rather than with the idea that these variants are regulatory.
• ✓For 44 expression-orphan loci, the methylation chain supplies specific, testable gene nominations.
• ✕Colocalization does not prove causation. A shared variant links two signals; it does not show that the molecular trait transmits the disease effect.
• ✕A methylation link does not establish mechanism or direction; methylation may mediate the effect, respond to it, or simply mark the same regulatory element.
• ✕Nominated genes are hypotheses for functional follow-up, not confirmed schizophrenia genes.
• ✕The 65 unexplained loci are not evidence of absent regulatory function; contexts this project could not measure, such as specific developmental windows or rare cell states, remain untested.

• The dashboard holds the headline figures: the waterfall of sequential attribution, the evidence matrix, and the full locus table. Click a waterfall bar to filter everything else to those loci.
• Every locus links to its own evidence card with per-channel PP4 values, fine-mapping annotations, and any chained gene nominations.
• The gene nominations page lists every chained candidate with both link strengths and filters for tier, orphan status, and CpG distance.
• The regional plots page shows three worked examples, including the FURIN positive control, at the level of individual variants.
• The about page lists data sources, methods, references, and the repository that regenerates every number shown here.

Each figure and table carries its own “How to read this” panel, and every dotted-underlined term opens a definition. The full list lives in the glossary.