thank you!

thank you Sally!

If you reached till here, and you still find this interesting: soon we will open a position for a computational scientist to continue on some of this work. Feel free to reach out, follow us and spread the news! #functionalgenomics #noncoding #enhancer #STARR-seq

Also grateful to the input of the anonymous reviewers which further improved the work. Particular, Reviewer 1 provided comments whose combined length surpassed that of the manuscript, requiring >100 pages of rebuttal and many additional analysis for which there was no space in the paper.

We started working on this in 2017 when I set-up my lab in Rotterdam. Many involved over the years, but I am particular grateful to both shared first-authors @ruizhideng.bsky.social (dry-lab) and Elena Perenthaler (wet-lab) and our collaborators (@eskeww.bsky.social , @roshchupkin.bsky.social)

Together, our functional atlas and BRAIN-MAGNET AI model bridge experimental and computational genomics, helping decode how the non-coding genome shapes the brain. And importantly, allows us now to move this knowledge into analysis of patient genomes aiming for novel diagnoses

This identified amongst others a novel enhanceropathy, caused by a rare single nucleotide variant in the enhancer of the gene RAB7A, causing a novel type of Charcot-Marie-Tooth disease. We show in cells and zebrafish models how this variant affects RAB7A expression.

Then we moved to the field of rare disease. Making use of the @genomicsengland.bsky.social 100,000 Genomes Project data, genomes from collaborators and from our genetics department @erasmusmc.bsky.social, we use BRAIN-MAGNET to filter for potential non-coding disease causing variants.

We then asked can we apply this model to interpret variants? We first tried on GWAS data 🧩:
BRAIN-MAGNET pinpointed which SNPs at disease loci actually affect enhancer function — correctly prioritizing previously experimentally validated variants for schizophrenia & depression.

Then came the AI. 🤖
We trained a convolutional neural network, BRAIN-MAGNET, directly on our experimental data.
It predicts enhancer activity from DNA sequence alone and highlights key nucleotides required for enhancer activity. Those predictions held up rigorous experimental validation

We then tested the activity of the same regions in embryonic stem cells, a developmentally earlier cell state. Comparing embryonic vs neural stem cells, we found “primed” enhancers: marked early in ESCs, activated later during neural differentiation. Capturing the early wiring of brain gene control

Ranking the functional enhancers according to their activity and linking them to their target genes provided insights into the regulatory grammar of gene regulation, for example showing enrichment of transcription factors at highly active non-coding regulatory elements (NCREs)

First, we used ChIP-STARR-seq to functionally test >148,000 candidate enhancers in human neural stem cells.
This represents one of the largest experimental atlas of quantitatively tested brain regulatory regions to date, providing novel insights in gene regulation during early brain development

Here we developed BRAIN-MAGNET (brain-focused artificial intelligence method to analyze genomes for non-coding regulatory element mutation targets), an AI model trained on functional genomics data to make it more easy to find those needles. How does this work?

Finding genetic variation in the non-coding genome is no longer so difficult; but interpreting their functional effects often is. Thus, trying to find disease-causing variants in the non-coding genome is like finding needles in a haystack. Our solution: bring a magnet!

Most of the human genome does not directly encode for proteins; yet it harbors most genetic variation. In routine human genetics, we usually focus on protein coding genes trying to find disease causing variants in humans. Why are we ignoring all that non-coding space so often?