A significant enrichment that is not: spatial nulls, co-expression, and the imaging transcriptomics of EEG alpha-power genetics

Researchers have long sought to bridge the gap between the electrical rhythms of the brain and the molecular machinery that drives them. While we know certain genes influence the strength of EEG (electroencephalography) oscillations—the rhythmic electrical patterns measured from the scalp—we do not fully know where in the cortex those genetic signals manifest. A common goal in imaging transcriptomics (the study of how gene expression correlates with brain imaging) is to determine if a set of trait-associated genes is "enriched," or over-expressed, in the specific brain regions that generate those rhythms.

However, a new study from Jacopo Schenetti reveals that many current claims of such enrichment may be illusory. By investigating the genetics of the alpha rhythm—a dominant brain wave occurring at 8–13 Hz—the authors demonstrate that what looks like a biological signal is often just a mathematical artifact. Specifically, the study finds that apparent enrichment can be entirely explained by the fact that genes within a set tend to be co-expressed (expressed together in the same regions) in the same brain territories. This is a nuance that standard statistical tests frequently miss.

The insufficiency of the spatial spin test

In imaging transcriptomics, researchers typically correlate brain-wide maps of gene expression with maps of genetic associations. To ensure that an observed overlap isn't just a fluke caused by the "smoothness" of brain maps—where neighboring regions look similar—the field standard is to use a "spin test." This involves rotating the brain map in various directions to create thousands of random spatial permutations. This ensures the observed enrichment is truly unique to the target region.

As shown in, the study utilizes a pipeline that moves from GWAS (genome-wide association study) statistics to regional gene expression profiles.

The researchers highlight a critical vulnerability. The spin test only corrects for spatial autocorrelation (the tendency of nearby regions to have similar values). It does not account for the internal structure of the gene sets themselves. If a group of genes is functionally related, they are likely to be expressed in the same cortical areas regardless of the trait being studied. This creates a "blind spot." A researcher might report a significant discovery that is actually a generic property of the cortex.

Testing alpha genetics against multiple nulls

To expose this flaw, the authors designed a rigorous multi-stage validation process. Instead of relying solely on the spatial spin test, they implemented three distinct layers of scrutiny:

1. A positive control: The authors first verified their pipeline by testing it against a known, massive expression gradient in the brain. The pipeline successfully recovered this signal with high significance ($p_{spin} = 2 \times 10^{-10}$) [Figure 4A]. This proves the machinery is sensitive enough to detect real effects.
2. An independent surrogate model: They cross-checked their custom spin implementation against an established model called brainsmash. This ensured the spatial permutations were mathematically sound.
3. A co-expression-aware gene-set null: This was the decisive architectural choice. Rather than rotating the brain, this test compares the target gene set against 10,000 random gene sets of the same size. This determines if the "alpha genes" are actually special, or if they behave just like any other arbitrary collection of genes.

When significance dissolves

If the researchers had stopped after the spin test, they would have published a classic success story. The authors report that, according to the field-standard spin test, alpha-power genes appeared significantly enriched in the 41 cortical alpha-generator regions ($p_{spin} = 0.022$ for continuous scores; $p_{spin} = 0.030$ for the top-100 genes) [Figure 6A]. This result was even corroborated by the brainsmash surrogate model.

However, the subsequent tests dismantled this conclusion. When the authors applied the co-expression-aware gene-set null, the alpha gene set became completely unremarkable ($p_{geneset} = 0.33$) [Figure 4B]. This value means the alpha set is indistinguishable from random genes. Furthermore, the signal lacked specificity. The enrichment was not unique to the alpha band. Theta, beta, and delta power frequencies showed comparable or even stronger nominal enrichment [Figure 6B]. Finally, the results failed to remain significant once the researchers adjusted for multiple comparisons. The minimum FDR (false discovery rate) corrected $q$-value was 0.127, which fails to meet standard thresholds for significance.

Limits of the current transcriptomic lens

The study acknowledges several factors that bound its conclusions. First, the underlying genetic data comes from a GWAS with modest sample sizes. This means the signal-to-noise ratio for individual genes is low. Only one gene, PRKG2, survived strict genome-wide correction .

There is also a fundamental "modality gap." The genetic associations are derived from scalp EEG, while the target regions were localized using MEG (magnetoencephalography). These two methods have different sensitivities to different types of neural sources. Additionally, the study notes that EEG "band power" measurements often conflate periodic oscillations with "aperiodic" activity (the background 1/f noise of the brain). If the genetic signal actually resides in the aperiodic component, a test focused on band power might be looking at the wrong biological phenomenon entirely.

The verdict: a new standard for enrichment

The verdict is clear. Reporting a spatial spin test is no longer sufficient to claim biological discovery in imaging transcriptomics. The authors demonstrate that a result can pass every conventional check. It can pass positive controls and surrogate validation. Yet, it can still be a false positive driven by gene co-expression.

For practitioners, the takeaway is a mandate for methodological rigor. If you are claiming that a gene set is enriched in a brain region, you must report a gene-set null alongside your spatial null. The computational cost of this extra step is negligible. It often takes just minutes of processing. However, it is the only way to ensure that your "breakthrough" is a property of the genes and not just a property of the cortex.

Figures from the paper