The standard genomics playbook still spends a lot of time watching protein-coding DNA, the roughly 1-2% of the genome that actually spells out proteins; this paper walks past that celebrity carpet and starts interrogating the noncoding 98% in the alley.

That sounds rude. It is also where a lot of disease risk appears to be hiding.

In “Decoding common and rare noncoding variant effects across cellular and developmental contexts,” Marderstein and colleagues built a very large prediction machine for a very small question: when one DNA letter changes outside a gene, what happens to gene regulation in a specific cell type? They generated about 3 billion predictions using deep learning models of chromatin accessibility across fetal and adult cellular contexts. Three billion. A normal person would maybe make tea first.

The Part of DNA That Does Not Wear a Name Tag

Noncoding DNA does not directly encode proteins, but plenty of it acts like switches, dimmers, timers, and badly labeled circuit breakers. Some regions help decide when genes turn on, where they turn on, and how strongly they shout into the cellular void.

The paper focuses on chromatin accessibility, which is a fancy way of asking: is this stretch of DNA physically open enough for regulatory proteins to reach it? Techniques like ATAC-seq measure these open regions across the genome. If DNA is a cookbook, chromatin accessibility tells you which pages are lying open on the counter and which ones are trapped under a stack of unpaid bills.

That matters because many disease-linked variants from genome-wide association studies land in noncoding regions. GWAS can point to a neighborhood. It often cannot tell you which house is on fire.

Common Variants Whisper Locally. Rare Ones Throw Furniture.

The main result is nicely weird.

The authors found that common variants tended to have more cell-type-specific regulatory effects. In other words, a common DNA change might matter in one cellular setting but not another. Very polite. Keeps its drama contained.

Ultra-rare variants, by contrast, showed larger and broader predicted effects across cell types. These were not tiny regulatory nudges. More like someone adjusted the thermostat with a hammer.

The strongest signal of purifying selection showed up in fetal neurons. Purifying selection is evolution’s way of saying, “No, absolutely not,” by removing harmful variants from the population over generations. Seeing that signal in fetal neurons suggests that regulatory disruptions during brain development may be especially costly. Evolution, famously, does not write friendly error messages.

FLARE Enters, Carrying a Clipboard

To make these predictions more useful, the team developed FLARE, short for Functional Lasso Analysis of Regulatory Evolution. It combines deep learning predictions with evolutionary constraint, including conservation signals such as PhyloP, to prioritize noncoding variants with unusually strong regulatory effects.

That combination is the key move. A model can predict that a variant changes chromatin accessibility. Evolutionary constraint can hint that the affected region has been protected over time because breaking it tends to go poorly. Put them together and you get a sharper list of variants worth investigating.

The authors applied FLARE across several problems: de novo mutations in childhood disorders, rare variants linked to outlier adult brain gene expression, and common variants enriched for schizophrenia heritability. That does not mean FLARE has solved these conditions. It means it gives researchers a better suspect list. In biology, that is already a luxury item.

Why This Lands Now

This paper sits inside a broader wave of genomic deep learning. Models like DeepSEA helped show that neural networks could predict regulatory effects from DNA sequence. Newer systems such as Borzoi predict RNA-seq coverage and variant effects across transcription, splicing, and polyadenylation. AlphaGenome pushes toward broader sequence-to-function prediction across thousands of molecular tracks. Reviews of the field now frame noncoding variant interpretation as one of the main jobs for genomic AI, not a side quest with funding paperwork.

The difference here is the scale and framing: common versus ultra-rare variants, across adult and fetal contexts, with single-cell chromatin accessibility and population genetics feeding the same machine. It is not just asking, “Does this variant matter?” It asks, “Where, when, how broadly, and does evolution seem annoyed by it?”

That is a much better question. Also a longer one. Science does that.

The Catch, Because Biology Charges Rent

These are predictions. Strong predictions, useful predictions, but still predictions. Deep learning models learn from available assays and cell states, which means blind spots travel with the training data like emotional baggage in a carry-on.

Chromatin accessibility is also only one layer of regulation. A variant might affect enhancer activity, RNA processing, chromatin contacts, transcription factor binding, or some combination that makes everyone’s whiteboard worse. Experimental validation remains the adult supervision.

Still, if the findings hold up and expand, this kind of framework could help prioritize rare disease variants, interpret psychiatric genetics, and guide functional experiments that would otherwise be fishing expeditions in a genome-sized lake.

The genome has billions of letters. Most do not code for proteins. Some still matter a lot. This paper gives researchers a better way to find the ones quietly rearranging the furniture.

References

1. Marderstein AR, Kundu S, Padhi EM, et al. Decoding common and rare noncoding variant effects across cellular and developmental contexts. Nature Genetics (2026). DOI: 10.1038/s41588-026-02619-6. PubMed: PMID 42298188.

2. Kathail P, Bajwa A. Leveraging genomic deep learning models for the prediction of non-coding variant effects. arXiv: 2411.11158.

3. Linder J, Srivastava D, Yuan H, Agarwal V, Kelley DR, et al. Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation. Nature Genetics 57, 949-961 (2025). DOI: 10.1038/s41588-024-02053-6.

4. Avsec Z, et al. Advancing regulatory variant effect prediction with AlphaGenome. Nature (2025). DOI: 10.1038/s41586-025-10014-0.

5. Pampari A, et al. ChromBPNet: bias factorized, base-resolution deep learning models of chromatin accessibility reveal cis-regulatory sequence syntax, transcription factor footprints and regulatory variants. bioRxiv (2024). DOI: 10.1101/2024.12.25.630221.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.