A DNA variant does not have one universal effect in the brain. Its consequence can depend on the region, the cell type and the developmental moment in which the surrounding regulatory machinery is active. Seal, a new sequence-based AI framework, tries to answer that contextual question across 802 states of human brain development.[1]
The model spans 26 brain regions, 30 cell types and seven developmental stages. Those dimensions could form 5,460 simple mathematical combinations, but the paper reports 802 modelled biological contexts—not a complete grid in which every cell type exists or is measured in every region and stage. The difference is the point: development is structured, sparse and transient.[1]
A context is the biological question attached to the same sequence
The scale of Seal
| Measure | Value | Context |
|---|---|---|
| Modelled contexts | 802 | Brain-specific combinations of spatial, cellular and developmental state.Source 1 |
| Brain regions | 26 | Spatial context changes which regulatory programmes are active.Source 1 |
| Cell types | 30 | A variant can matter in one cell type and be quiet in another.Source 1 |
| Developmental stages | 7 | Transient fetal states are especially difficult to measure directly.Source 1 |
| Neuropsychiatric conditions analysed | 6 | The study compares shared and distinct regulatory architectures across conditions.Source 1 |
One variant, several possible answers
| Sequence question | Context added by Seal | Potential output |
|---|---|---|
| Does this base change alter expression? | Region | A predicted effect can be localised to a brain area rather than treated as brain-wide.Source 1 |
| Which cells are affected? | Cell type | The same variant can rank differently in excitatory neurons, glia or other populations.Source 1 |
| When does it matter? | Developmental stage | A transient fetal effect may disappear from adult tissue and ordinary expression atlases.Source 1 |
| How might it act? | Transcriptional regulator link | The model connects sequence change to regulatory programmes, offering a mechanistic hypothesis.Source 1 |
Transfer learning is doing the work where biology has missing rows
Directly measuring gene expression and variant effects across every developing brain state is impossible. Tissue is scarce, cell states are short-lived and the relevant combinations are unevenly sampled. Seal uses transfer learning to build sequence-to-expression models from richer data and adapt them to the sparse developmental contexts.[1]
The Seal analysis pipeline
| Stage | Question | Output |
|---|---|---|
| Sequence model | What expression pattern does the surrounding DNA support in this context? | A context-specific baseline prediction.Source 1 |
| Variant perturbation | How does changing one allele alter that prediction? | A signed and ranked regulatory-effect estimate.Source 1 |
| Mechanistic interpretation | Which transcriptional regulators are implicated? | A hypothesis about the regulatory programme through which the effect operates.Source 1 |
| Trait integration | Do high-scoring effects concentrate in genetic studies of a condition? | Shared and condition-specific cell types and developmental windows.Source 1 |
| Cohort analysis | Are de novo variants in affected children enriched in particular contexts? | A burden signal to investigate biologically, not a diagnostic test.Source 1 |
The paper tests whether predicted effects are stronger in the biological context where a variant is known to act. Across matched cell-state comparisons, Seal produced higher effects for astrocytes, oligodendrocytes and microglia, while one oligodendrocyte-precursor comparison did not meet the conventional false-discovery threshold. That mixed detail matters: context sensitivity is supported, but not every cell class delivers equally strong evidence.[1]
The autism result is a developmental map, not a variant verdict
In whole-genome data from the Simons Simplex Collection, Seal found a significant burden of de novo regulatory variants in transient fetal excitatory neurons among autistic probands relative to sibling controls. The result directs attention to a cell state and developmental period that ordinary adult-tissue analysis could miss.[1]
It does not mean that a Seal score can diagnose autism, that every high-scoring variant is causal or that transient excitatory neurons are the sole mechanism. Burden analyses compare distributions across groups. Each individual variant still needs genetic, molecular and functional evidence.[1]
What Seal adds—and what remains outside the model
| Contribution | Boundary |
|---|---|
| Context-specific ranking of regulatory variants | A prediction is not direct measurement of expression in an individual fetal brain.Source 1 |
| Mechanistic links to transcriptional regulators | The link is a testable hypothesis, not proof of the molecular chain.Source 1 |
| Trait and cohort analyses across sparse developmental states | Association and burden do not establish an intervention target.Source 1 |
| A reusable framework for other sequence-regulation questions | Performance can change with ancestry, assay, cell-state definition and training data.Source 1 |
Seal's most important contribution is conceptual. It replaces the question 'Is this variant harmful?' with a set of more precise questions: in which cells, in which region, at what stage and through which regulator might this sequence change matter? For developmental disease, that is not added complexity. It is the biology the simpler question leaves out.[1]
Sources & further reading
- Spatiotemporal AI-based modeling of gene regulation and variant effects in the developing human brain using Seal Nature Neuroscience · 15 Oct 2026 · Accessed 16 Oct 2026