Predicted Regulatory Layers Split Complex-Trait Heritability by Disease and Tissue Context
A new preprint argues that not all heritable signal within a tissue is created equal. Using sequence-based variant-effect predictions from AlphaGenome Atlas, investigators compared how different predicted regulatory layers captured the heritability of 11 complex traits, asking a question that earlier partitioned-heritability studies could not answer cleanly: within the same tissue, which regulatory layer is most informative for a given trait?
The answer was not uniform. Traits localized to the expected lineages, but the most informative predicted layer differed sharply across phenotypes, with a trait-by-layer difference reaching P = 3.6 × 10-8. In practical terms, the same tissue biology did not map to the same regulatory mechanism across diseases and quantitative traits. Some traits appeared to lean on predicted chromatin accessibility effects, while others were better captured by predicted transcription-initiation effects.
One Tissue, Multiple Regulatory Signals
The analytic constraint driving this work is familiar to anyone who has tried to compare regulatory annotations across studies. Measured annotations for different regulatory layers usually come from different biosamples, which makes direct comparison messy and can blur the biological signal. Sequence-to-function models alter that equation by predicting the allelic effect of every variant on several regulatory layers using a single model, allowing the layers to be compared on the same variants within the same tissues.
Here, predicted effects were grouped into 12 lineages and six regulatory layers, then tested with stratified LD score regression across the trait set. The lineage results aligned with established biology, which lends credibility to the layer-specific findings. Height, bone mineral density, brain-related traits, and cholesterol did not drift into surprising tissue categories; instead, they largely landed where current genetic architecture would suggest. The sharper signal came one level deeper, at the regulatory layer itself.
One layer did not fit all traits.
Chromatin Accessibility Dominated Skeletal Traits
Predicted chromatin-accessibility, or DNase, effects were the most informative layer for height and bone mineral density. That pairing makes biological sense. Skeletal growth and mineralization depend on tightly timed enhancer and promoter accessibility programs in developmentally active cell states, where the ability of transcriptional machinery to reach DNA can shape downstream growth trajectories. A regulatory model that captures open-chromatin perturbations may therefore be especially useful for traits with strong developmental and structural components.
The result does not imply that chromatin accessibility is the only mechanism involved. It suggests something more specific: among the predicted regulatory signals tested here, accessibility carried the greatest heritability information for those traits. In genetic terms, the variants most relevant to skeletal phenotypes may act first by altering whether regulatory DNA is open or closed to transcriptional control.
Transcription Initiation Mapped More Strongly to Brain and Lipid Traits
For Alzheimer’s disease, most brain traits, and cholesterol levels, predicted transcription-initiation, or CAGE, effects were the most informative. This pattern is notable because transcription initiation sits at a bottleneck in gene regulation. Changes in promoter activity and initiation frequency can propagate into cell-type-specific expression changes, which is especially plausible in the brain, where precise expression timing and neuronal subtype specificity are central to function.
Cholesterol appearing in the same regulatory bucket as brain traits may reflect shared transcriptional control logic in metabolic tissues, where promoter-level regulation can influence lipid handling, hepatic transport pathways, and systemic homeostasis. The mechanism is not claimed as proven, but the genetic architecture points in that direction: among the tested predicted annotations, transcription initiation captured more of the heritable signal than chromatin accessibility for these outcomes.
The Layer Contrast Survived Multiple Checks
The pattern did not disappear when the investigators stress-tested it. The layer contrast persisted across sensitivity analyses, which reduces the likelihood that the result was a statistical artifact of one analytical choice. Even more interesting, the same predicted contrast held for three related traits whose layers had been predicted in advance, with P = 2 × 10-7. Pre-specification matters because it limits the temptation to explain away a pattern after the fact; here, the layer-specific separation remained visible even when the trait context was constrained before analysis.
The authors also conditioned predicted CAGE effects on fine-mapped eQTLs and measured regulatory annotations. CAGE remained informative. That finding matters because it implies the predicted transcription-initiation signal was not merely a proxy for already-known expression quantitative trait loci or for existing experimental annotations. Instead, it retained independent heritability information after accounting for those features, suggesting that sequence-based prediction may be capturing regulatory influence that conventional annotation sets do not fully recover.
What This Means for Complex-Trait Genetics
This preprint reframes a familiar heritability question. Tissue specificity remains important, but the more precise unit of analysis may be the regulatory layer within that tissue. A trait can localize to the same lineage while still drawing most of its inherited architecture from a different regulatory mechanism than another trait in that lineage. For investigators, the practical implication is straightforward: annotation choice can alter what looks biologically important, even when the tissue context is held constant.
For studies of complex traits, that opens a useful path. Predicted variant-effect models can be used to generate testable hypotheses about whether a phenotype is more likely driven by promoter-level disruption, chromatin accessibility shifts, or other regulatory effects. In a disease like Alzheimer’s, that could help prioritize promoter-centered experiments in relevant brain cell types. In skeletal biology, it may steer attention toward accessibility-sensitive enhancers and developmental windows.
The study remains a preprint and has not been peer reviewed, so the conclusions should be treated as provisional. Its scope was also limited to 11 traits and to the specific prediction and regression framework used here. Even so, the central result is concrete: heritability information is not distributed uniformly across predicted regulatory layers, and those differences appear to be trait-specific rather than generic across complex disease.