AI Model Maps How Genetic Variants Rewire Gene Regulation in the Developing Human Brain

A new computational framework can predict, directly from DNA sequence, how genetic variants change gene expression in the developing human brain — down to individual cell types and developmental stages, including transient fetal cell states that are difficult or impossible to capture experimentally.

The tool, called Seal, is reported in Nature Neuroscience as a Technical Report by a team led by computational biologist Olga G. Troyanskaya at the Center for Computational Biology, Flatiron Institute, with Yun Hao as first author. Seal uses transfer learning: it is pretrained on abundant gene expression data from adult brain regions and cell lines (drawing on GTEx, ENCODE and Roadmap data), then fine-tuned on scarcer developmental data by freezing most of the network and updating only the output layer. That strategy lets it make predictions in data-poor settings where standard models struggle.

Seal models 802 brain contexts — 30 cell types across 23 brain regions, plus 26 brain regions — spanning seven developmental stages from early fetal life through adulthood. According to the authors, that is roughly eightfold broader coverage of brain-specific contexts than existing sequence-to-expression models such as Enformer and AlphaGenome.

On held-out chromosomes, Seal's predicted expression correlated with observed expression at an average Spearman correlation of 0.81 across all 802 contexts, with an AUROC of 0.99 for distinguishing highly from lowly expressed genes. Tested without any additional training on two independent single-cell RNA-seq datasets from the fetal brain, Seal reached AUROC values of 0.97 and 0.91, compared with 0.85 and 0.81 for Enformer and 0.84 and 0.80 for AlphaGenome on the same benchmarks.

Why this matters: most genetic risk variants for brain disorders fall in noncoding DNA, and genome-wide association studies rarely say which cell type, brain region or developmental window is affected. Seal aims to fill that gap by predicting variant effects ab initio, independent of linkage disequilibrium, and by tracing predictions back to specific transcription factors, histone marks and chromatin accessibility features.

The authors applied Seal to 743,810 brain-associated expression quantitative trait loci (eQTLs) and found that fine-mapped eQTLs carried stronger predicted effects than non-eQTL variants. They then used the predictions as genome-wide annotations in LD score regression to partition heritability for brain traits and psychiatric conditions. Cortical surface area showed higher heritability enrichment among variants predicted to perturb expression in fetal brain regions, while cortical mean thickness was enriched in adult-stage contexts. Among six neuropsychiatric conditions, autism spectrum disorder (ASD) and ADHD showed significant enrichment only at fetal, neonatal and infancy stages, whereas bipolar disorder, schizophrenia, major depressive disorder and Alzheimer's disease showed effects across later stages as well — a pattern broadly consistent with their ages of onset. In addition to known cell types, this GWAS analysis flagged two cell states not previously linked to ASD — VIP-KIRREL3 interneurons and SST-BRINP3 interneurons — as candidates for follow-up.

The most striking application used whole-genome sequencing from the Simons Simplex Collection. The team analyzed 27,926 de novo mutations near gene transcription start sites from 1,781 ASD probands and 1,735 unaffected siblings across 82 prefrontal cortex cell states. Restricting to 1,704 mutations near high-confidence ASD risk genes (SFARI categories 1 and 2), they found the strongest predicted regulatory burden in excitatory neurons, concentrated in fetal-stage cell states: 13 excitatory neuron states showed significant burden, seven of them fetal, including L2/3-CUX2, L5/6-TLE4 and L5/6-THEMIS neurons previously implicated in ASD.

Feature attribution highlighted transcription factors such as p300, YY, CTCF, FOXP2, JARID1A, GATA3 and ZNF263 among those most affected, suggesting a hypothesis that ASD-associated regulatory variants act partly by disrupting transcription factor binding in fetal excitatory neurons.

Important caveats apply. This is a computational study: the ASD burden result concerns predicted regulatory effects on expression, not measured expression in patient tissue, and the interpretations are hypotheses rather than demonstrated mechanisms. Every prediction will need experimental validation. The authors also note a key limitation — the current framework does not model distal enhancer–promoter interactions, a direction they identify for future work. Seal's accuracy outside brain contexts and species is likewise untested.

Still, the work offers a template for squeezing mechanistic insight out of sparse, context-specific data — and a way to generate testable, cell-type- and stage-specific hypotheses about how noncoding variation contributes to brain development and disease.

For social sharing: "A new AI framework predicts how genetic variants alter gene expression across 802 contexts of the developing human brain — and points to fetal excitatory neurons as a hotspot for autism-associated regulatory burden."

Reference: Hao, Y., Park, C.Y., Theesfeld, C.L. & Troyanskaya, O.G. Spatiotemporal AI-based modeling of gene regulation and variant effects in the developing human brain using Seal. Nature Neuroscience (2026). https://doi.org/10.1038/s41593-026-02470-2