Users are thrilled by the Kundaje Lab's FLARE tool publication because it delivers billions of variant effect predictions across brain and heart contexts and they anticipate further discoveries by the community.
18/ FLARE: https://github.com/drewmard/FLARE Models and Predictions: https://www.synapse.org/Synapse:syn64693551/files/ and https://www.synapse.org/Synapse:syn73770440/files/ Analysis Code: https://github.com/kundajelab/neuro-variants
19/ Thanks to all of our collaborators, and we look forward to seeing what others discover with these tools!
And here’s a full-text sharable link to the paper: https://rdcu.be/fopFM
2/ This was first released as a preprint last year (https://www.biorxiv.org/content/10.1101/2025.02.18.638922v2), and we're thrilled to now see it published, which brings together machine learning, statistical genetics, evolution, & gene regulation to better understand functional impacts of non-coding variation.
14/ One of the other applications of FLARE is to identify mutations that underlie outlier gene expression.
In 791 adult brain samples with paired WGS and brain RNA-seq, top-scoring FLARE-brain variants were enriched near genes showing outlier underexpression.
4/ To address this, we used: - single-cell ATAC-seq across adult and fetal brain and heart (132 contexts in all), - trained deep learning-based DNA sequence models of chromatin accessibility (ChromBPNet), - and made over 3 billion variant effect predictions
3/ Whole genome sequencing has led to a huge catalog of variants, and GWAS has linked many of the common ones to traits and disease.
But a large fraction are rare or de novo, and we have few tools to assess functional impact (especially in the non-coding genome!)
5/These predictions captured the effects of common, fine-mapped disease variants, which had larger effects in relevant contexts
(e.g. predicted brain effects for fine-mapped GTEx brain eQTLs, or heart effects for CAD loci).
17/ We're excited to share FLARE, >3 billion variant effect predictions across 132 adult and fetal brain and heart contexts, and all associated code and resources with the community.
15/ The paired RNA-seq also helped interpret FLARE variants.
For example, two SNPs are only 78 bp apart but affect different genes: - one breaks an NFIL3 motif (lowering SENP3), - the other creates a ZEB/SNAI repressor motif (lowering TNFSF13).
6/ Fine-mapping + ATAC peak + ChromBPNet helped narrow down GWAS loci to the causal variant, TF motif, and cell type.
For example, our approach identified a single variant in an Alzheimer’s locus, which creates a new ZEB/SNAI repressor motif to lower chromatin accessibility.
10/ This is particularly striking in fetal neurons, which fits Medawar's theory of aging:
selection wanes after reproduction,
so fetal variants get purged while later ones escape.
8/ These models were powerful for studying ultra-rare variants (MAF<0.1%), where methods like GWAS are underpowered.
This allowed us to ask about the regulatory properties that differ between common and rare variants.
9/ Accounting for gene distance and constraint, ultra-rare variants had larger predicted effects than common ones while affecting a broader number of cellular and developmental contexts.
16/ Finally, FLARE informed common variant architecture.
As a schizophrenia heritability annotation in S-LDSC, FLARE-fetal brain beat the strongest region-based annotations.
7/ We then used CRISPRi to knockdown the candidate enhancer in iPSC-derived microglia, which supported RASGEF1C as the target gene.
This demonstrates how computational predictions can directly guide experimental validation!
12/ Applying FLARE-fetal brain to de novo mutations in ~2,000 autism families, 14 of the 16 highest-scoring mutations near syndromic autism genes occurred in probands rather than unaffected siblings, with several hits near CNTNAP2.
13/ We then trained FLARE on heart data and applied it to de novo mutations in congenital heart disease, where most top scoring mutations were found in cases.
11/ Leveraging these observations, we built FLARE (Functional Lasso Analysis of Regulatory Evolution). FLARE: - predicts conservation (PhyloP), - uses regulatory and genomic features, - can be trained on any specific context.
@amarderstein @soumyakundu_ @anshulkundaje Congratulations 🎉