A change in DNA can have no detectable effect in one cell and a consequential effect in another. It may matter during fetal development and become quiet in adulthood, or act in one brain region while leaving a neighbouring region largely unchanged. A single universal “harmfulness” score cannot express that timing and location.[1]
Seal is designed around that missing context. The interpretable transfer-learning framework predicts gene expression and regulatory variant effects across 26 regions, 30 cell types and seven developmental stages. The published model provides 802 brain-development contexts: the combinations actually represented and modelled, not the full 26 × 30 × 7 Cartesian product.[1]
From a DNA sequence to a contextual effect
The question changes at each layer
| Layer | Question | Seal’s role |
|---|---|---|
| Sequence | What bases surround the gene or variant? | Reads genomic sequence near regulatory regionsSource 1 |
| Expression | How strongly is the gene expected to be active? | Predicts expression for a specified brain stateSource 1 |
| Variant effect | How does changing one sequence alter the prediction? | Compares the reference and alternative sequence in the same contextSource 1 |
| Mechanism | Which regulators may explain the change? | Links predicted effects to transcriptional regulatorsSource 1 |
The model first learns transferable sequence-to-expression relationships, then adapts them to brain regions and cell states where experimental data may be sparse. That matters in development because many transient cell populations are difficult to sample repeatedly. The paper reports that transfer learning lowered prediction error, particularly for genes involved in neurodevelopment, compared with training all context-specific profiles in one step.[1]
Why 802 contexts are more informative than one score
Imagine a variant near a gene used by radial glia early in cortical development. A context-free system might predict a moderate regulatory effect. Seal can instead ask whether the effect is large in radial glia, small in mature neurons and different again in another region. The useful output is a pattern across states, which can guide the next experiment towards the cells and developmental window where a signal is strongest.[1]
The model’s published scope
| Measure | Value | Context |
|---|---|---|
| Brain regions | 26 | Regional contexts represented in the frameworkSource 1 |
| Cell types | 30 | Cell classes include transient developmental populationsSource 1 |
| Developmental stages | 7 | Predictions span fetal and later stages described in the studySource 1 |
| Available contexts | 802 | Observed or modelled brain states; not every theoretical region–cell–stage combinationSource 1 |
| Conditions compared | 6 | Neuropsychiatric conditions used to examine shared and distinct regulatory architecturesSource 1 |
The autism finding is a prioritisation result
The researchers applied Seal to whole-genome data from the Simons Simplex Collection. They report a significant burden of de novo regulatory variants around autism-associated genes in transient fetal excitatory neurons. The result points to a developmental cell state in which non-coding mutations may have concentrated functional effects.[1]
That does not make Seal an autism diagnostic test. It does not observe a child’s developing brain directly, and a predicted regulatory effect is not the same as a confirmed causal mechanism. The model narrows a vast search space: it identifies variants, regulators and cellular windows that deserve functional experiments.[1]
Open code helps, but validation still has to follow the context
The authors provide code and trained parameters. That supports reproduction of the computational workflow and lets researchers inspect predictions for their own variants. Experimental validation remains context-specific: a result in fetal excitatory neurons needs evidence in a biologically appropriate model, rather than confirmation in an unrelated adult cell line.[1][2]
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
- Seal framework code FunctionLab · Accessed 16 Oct 2026