Use CasesEnterpriseResearchBlog

Company

Pricing
Biomni Labarrow

Decoding Neural Crest Grammar with AI Agents

How the Naqvi Lab used Biomni to test whether closely related neural crest cells follow different DNA grammar rules.

Decoding Neural Crest Grammar with AI Agents

By the Naqvi Lab (PI: Sahin Naqvi) Boston Children's Hospital and Harvard Medical School Phylo Academic Lab Program

Nov 19, 2026 · 7 min read

Every cell in your body carries the same DNA, yet a facial bone cell and a pigment cell behave completely differently. Part of the answer lies in the "grammar" of DNA. Short sequences called motifs act as docking sites for regulatory proteins called transcription factors that help switch genes on and off. It matters which motifs are present, but also how they are arranged: their order, spacing and orientation.

Scientists have learned a lot about this grammar across very different cell types. But we wanted to ask a subtler question: when closely related cells use the same set of transcription factors, do those factors follow different grammatical rules?

Why neural crest cells?

Neural crest cells are a small, short-lived population in the early embryo, and developmental biologists have long been fascinated by them. They start out along the developing neural tube, which gives rise to the brain and spinal cord, then migrate throughout the body. They then transform into an astonishing range of cell types and tissues: facial bone and cartilage, peripheral nerves and glia, and the pigment cells in our skin.

They come in four flavors, named for where they arise along the head-to-tail axis: cranial, vagal, trunk and sacral. All four rely on the same core set of "master regulator" proteins, including SOX9/10, TFAP2A/B and PAX3. Yet reducing those proteins causes defects in only some of these four subpopulations. That matters for health, because losing about half of one of these proteins can cause a rare disorder. For example, reduced levels of TFAP2A/B disrupt facial bone and cartilage (derived from cranial neural crest) with no major effects on peripheral nerves and glia. Our hypothesis was that the same proteins follow slightly different grammatical rules in each one.

Studying this in real embryos is hard because there are so few of these cells. Instead, we used neural crest cells grown from human stem cells in the dish. That gave us data on chromatin accessibility (measured by ATAC-seq, which maps DNA regions accessible to cellular machinery) for all four subpopulations, from public studies and our own lab.

Experiments on a computer

To investigate those rules, we trained separate deep learning models on the accessibility data from each subpopulation. Each model learned to predict accessibility from DNA sequence in that cellular context. Then we ran virtual experiments on the trained models. We inserted a motif, or a pair of motifs, into background DNA sequences and predicted the change in accessibility relative to the background alone (Figure 1). For pairs, we shifted the spacing one base at a time and tested both orientations. We then compared the predicted effects of the same motifs and motif arrangements across the models, asking which responses were shared and which differed between populations.

Figure 1. How the virtual experiments work. (A) Each motif is inserted into background DNA sequences, and the model predicts the change in accessibility. (B) Two motifs are tested at spacings from 0 to 150 base pairs, in both orientations, to find where they cooperate. The six models represent the four subpopulations along the head-to-tail axis.

How Biomni helped

We gave it a list of accession numbers, and got back analysis-ready data. We gave Biomni only the GEO accession numbers of the public studies. It downloaded and mapped the data for all four neural crest subpopulations, then merged biological replicates to get deeper coverage.

We gave it a paper's methods section, and got back a full analysis. We wanted to reproduce an approach from a published study (Liu et al., Nature 2026). It places pairs of motifs into background DNA sequences across a range of spacings and both orientations, then asks the model how much the pair increases predicted accessibility. We pasted the paper's methods text into Biomni and pointed it to tangermeme, a recent toolkit for virtual experiments. Biomni wrote the code, planned and ran the GPU jobs, and analyzed the results. It flagged which motif pairs cooperate, and classified each as flexible ("soft" syntax) or strictly spaced ("hard" syntax).

Along the way, Biomni also distilled 22 consensus neural crest motifs from the models' outputs. It then estimated how much each motif contributes to predicted accessibility in each subpopulation. Together, these analyses let us systematically compare motif effects and interactions across the four populations.

Where we stepped in. A few steps needed our own setup. Some datasets required training models with custom bias correction. We ran that step on our own cluster with existing scripts, along with model interpretation (DeepLIFT and TF-MoDISco), then handed the trained models and their motif results back to Biomni for everything downstream. Biomni can now run all of these steps itself, but at the time we did them on our own.

What we found

The grammar is largely shared across the four subpopulations, but with notable differences.

  • The core regulators are active everywhere. Motifs for SOX9/10, TFAP2A/B, PAX3, NR2F1/2 and TEAD matter in nearly all four subpopulations. Together with earlier expression data, this supports these transcription factors being functionally active across the neural crest (Figure 2).
Figure 2. How much each consensus motif raises predicted accessibility in each of the six models. Darker green means a bigger effect.
  • Two kinds of cooperation. When two motifs sit near each other, they can boost DNA activity more than either does alone. Some pairs cooperate across a wide range of spacings, called soft syntax. Others cooperate only at one exact spacing and orientation, called hard syntax. TFAP2 and THAP/ZNF motifs, for example, cooperate sharply when they sit 4 base pairs apart in cranial cells, while SOX and TFAP2 motifs cooperate anywhere from 20 to 150 base pairs apart (Figure 3). Most of these rules are shared across subpopulations (Figure 4).
Figure 3. One example of each. Left: TFAP2 and THAP/ZNF motifs cooperate sharply at a spacing of 4 base pairs in one cranial model (hard syntax). Right: SOX and TFAP2 motifs cooperate across 20 to 150 base pairs (soft syntax).
Figure 4. The share of motif pairs showing each kind of cooperation, for five anchor motifs across the six models.
  • Different transcription factors follow different rules. SOX motifs cooperate with partners at almost any spacing, which fits SOX9/10's known role as "pioneers" that open up tightly packed DNA (Figure 5). PAX motifs are pickier: specific spacings and orientations with other motifs produce strong synergy (Figure 6).
Figure 5. How the SOX motif cooperates with each partner motif in each model. S marks soft syntax, H hard syntax, H+S both, and a dash no cooperation.
Figure 6. The same view for the PAX motif. Syn marks cooperation that fits neither class.
  • A head-to-tail gradient. Cranial and vagal cells show the most motif cooperation, trunk cells less, and sacral cells the least. The grammar seems to shift along the body axis.

What's next

These model predictions give us specific hypotheses to test at the bench. Our lab has built stem cell tools to dial the core neural crest transcription factors up and down. We're using them to investigate the shared and the subpopulation-specific grammar rules Biomni helped us uncover.

See it for yourself: replay the full Biomni session to see every step Biomni took.

Try it yourself. Academic labs can run a smaller version of this workflow on public data from human sacral neural crest cells with this prompt:

Download the ATAC-seq data for samples GSM5973057 and GSM5973058 from GEO series GSE199440 and map it to hg38. Merge the two replicates and train a ChromBPNet model. Run DeepLIFT and TF-MoDISco to find the motifs that drive accessibility, then insert each top motif into 100 background sequences and report how much it raises predicted accessibility.

Supporting the work of scientists

We're excited to collaborate with more scientists and research institutions to push the boundaries of what's possible with AI-powered research. Whether you're looking to accelerate computational analyses, explore new hypotheses, or bring AI-powered research capabilities to your broader team, we'd love to learn about your work and explore how Biomni Lab can support it.

If your organization is interested in deploying Biomni Lab at scale to accelerate drug discovery and development, fill out the form below to explore enterprise solutions with our team!

Technical details

For readers who want the specifics. Adapted from the lab's original write-up.

Data and models. ATAC-seq data from hPSC-derived cranial, vagal, trunk and sacral NCCs came from public GEO studies plus data newly generated in our lab. Biomni downloaded and mapped the public data and merged biological replicates. It initiated ChromBPNet training on each merged dataset. Some studies required custom bias models, so we trained those ChromBPNet models on our local cluster with existing scripts. We then ran DeepLIFT for base-pair contribution scores and TF-MoDISco for motif discovery.

Consensus motifs. Using MotifCompendium, Biomni derived 22 consensus NCC motifs from the union of all TF-MoDISco results.

Single-motif marginalization. Each consensus motif was inserted into 100 background sequences. The increase in predicted counts quantified its contribution to predicted accessibility in each subpopulation (Fig. 1A). Motifs for SOX9/10, TFAP2A/B, PAX3, NR2F1/2 and TEAD increased predicted accessibility in almost all subpopulations (Fig. 2).

Pairwise motif synergy. Following Liu et al. (2026), one motif was inserted at the center of background regions. A second motif was inserted in either orientation, sliding at 1 bp spacing steps, and model predictions were computed for each instance. Pairs with joint effects exceeding the expectation for independently acting motifs were classified as soft syntax (cooperativity across a broad range of spacings and orientations), hard syntax (one or a few favored arrangements) or mixed. Biomni implemented this from the paper's methods text using tangermeme, planned and ran the GPU jobs, and classified the results (examples in Fig. 3; overview in Fig. 4).

Key results. A set of “core” motifs were active across all subpopulations. Syntax involving these core motifs was mostly shared, with some specific differences. SOX motifs showed almost exclusively soft syntax (Fig. 5). PAX motifs showed multiple instances of hard syntax with a range of partner motifs (Fig. 6). Sacral NCCs showed the least motif synergy and cranial and vagal the most, suggesting an anterior-posterior gradient.

Reference. Liu, B.B., Jessa, S., Kim, S.H. et al. Multiomics and deep learning dissect regulatory syntax in human development. Nature 653, 1240–1253 (2026). https://doi.org/10.1038/s41586-026-10326-9

Acknowledgment. Parts of this post were drafted with the help of AI (e.g., Biomni). The Naqvi Lab reviewed the full text and figures for accuracy.