2026-08-25
IMP and UC Irvine used compact CNNs plus transfer learning to design 15 heart, limb and CNS enhancers; all 15 fired in the intended E11.5 mouse embryo tissue.
Enhancers are short DNA stretches that turn genes on in some tissues and not others. Deep learning already predicts and designs them in fly embryos, zebrafish, and mammalian cell lines. Whether the same playbook survives the longer genomes and messier tissues of a living mammal was still untested.
The Stark lab at IMP in Vienna and the Kvon lab at UC Irvine pinned that question to E11.5 mouse embryos. Heart, limb, and central nervous system. Design enhancers from sequence, put them into live mice, and see if they light up where they should.
Four steps, built around a small model and a thin labeled set.
They first trained sequence-to-accessibility CNNs on published ATAC-seq from E11.5 mouse heart, limb, and midbrain. Input is 1,001 bp of one-hot DNA. Output is chromatin accessibility. The network is a compact DeepSTARR-style CNN: four conv layers (256/120/60/60 filters) plus two fully connected layers. Chromosome 18 was held out. Pearson correlation on that chromosome is at least 0.76, at least 0.88 on a class-balanced test set, and at least 0.62 on tissue-specific peak differences.
Those models were then fine-tuned on VISTA embryonic mouse and human enhancers, turning the head into a sigmoid classifier of in vivo activity. Each tissue only had 311 to 432 tissue-specific positives, so they scored models on positive predictive value (PPV), the fraction of predicted enhancers that are real. Transgenic assays in mice are slow and expensive; a false positive costs more than a miss.
Design used Ledidi, a gradient-based sequence editor, pushing both accessibility and activity scores up at once. Seeds matched VISTA dinucleotide frequencies. An edit penalty of 0.1 kept the optimizer from rewriting the sequence too aggressively. BLAST then dropped anything with a significant match to the mouse or human genome.
Five sequences per tissue (15 total, 1,001 bp each) went into a site-specific H11 transgenic reporter and were scored by LacZ at E11.5. Per VISTA convention, a construct counts as positive only if at least three independent embryos show signal in the target tissue.
Transfer learning is doing real work here. Training on ATAC-seq alone, or training on VISTA without pretraining, drops PPV by 20.9 to 52.1 percentage points depending on tissue. After the fine-tune, motif contribution of tissue master regulators (MEF2 in heart, TWIST1 in limb, SOX3 in CNS) holds or rises, while broadly used insulator motifs such as CTCF get down-weighted. The model is learning which open chromatin is actually an enhancer.
On held-out VISTA sequence, PPV reaches 70.6% and above in all three tissues. Figure 1d peaks at 71.5% (heart), 70.6% (limb), and 80.2% (CNS/midbrain), still with at least 100 predicted positives on the curve. The same models score about 300,000 random sequences and 300,000 inaccessible genomic bins near zero.
All 15 designed enhancers were reproducibly active in the intended tissue. Specificity is less uniform.
| Tissue | Designed | On-target | Specificity |
| Heart | 5 | 5/5 | 4 heart-only; enhancer 5 weaker in heart, some brain leak |
| Limb | 5 | 5/5 | Strong in limb, weaker in other TWIST1+ mesenchyme (craniofacial, trunk) |
| CNS | 5 | 5/5 | 4 CNS-specific; one weak heart leak |
Embryo counts in the representative set: heart 7/14, 7/7, 6/6, 3/3, 5/8; limb 4/5, 9/9, 15/15, 15/15, 7/10; CNS 5/5, 6/7, 8/8, 6/6, 11/11. None of the 15 has a significant BLAST hit against mm10, hg19, or VISTA (E-value > 0.05). These are not lightly edited natural enhancers.
For tissues without a VISTA-scale gold set, three proxy labels (distal, non-CTCF ATAC peaks; H3K27ac plus H3K4me1 peaks; tissue-specific ATAC peaks) used in place of VISTA still lift PPV over accessibility-only models. Cross-validation puts expected design success at least at 60%.
Subregion design is harder. Two forebrain enhancers both fired in forebrain (7/8 and 4/4); one leaked into limb, a tissue the optimizer was not asked to suppress. Of two midbrain designs, one was dead and the other uninterpretable because of a 419 bp duplication in the reporter. Subregion targeting looks feasible in principle. The current sample size does not support picking any brain subdomain on demand.
For gene therapy, synthetic biology, and functional genomics, this moves tissue-specific mammalian enhancer design from fly-style feasibility to a 15/15 result in the mouse embryo. The cost of entry is low: a compact CNN, public ATAC-seq, and a few hundred validated enhancers per tissue. No Enformer-scale long-context genome model, no huge MPRA.
Weights are on Hugging Face as DeepSTARR-Mouse; code is on Zenodo. A lab with a new tissue can fine-tune. For tissues without VISTA, tissue-specific ATAC peaks as weak labels are the more realistic number, around 60% expected success.
Read the 15/15 in context: candidates were drawn from a high-specificity, low-recall slice of sequence space, then five per tissue were hand-picked. That is not the same as claiming the model has fully absorbed mammalian regulatory grammar.
The authors already fence the claim. They designed single enhancer elements in a standardized reporter and promoter. Native output still depends on promoter identity, chromatin neighborhood, and helper elements. The work does not take on multi-enhancer architectures at genome scale, the problem long-sequence models are aimed at. Only three tissues and one stage (E11.5). They did not target defined expression strength, arbitrary lengths, or complex multi-tissue patterns.
In vivo scoring was done by two people, unblinded. The positive rule is visible signal in at least three embryos. Limb off-target tracks the TWIST1 expression domain, so the model learned the master regulator and has not yet carved limb away from neighboring mesenchyme. One midbrain failure is a cloning artifact; two sequences are too thin for a strong subregion claim.
The 70%+ PPV is on held-out VISTA sequences. The 15/15 in vivo result is five hand-picked high-scoring designs per tissue. The paper does not report what happens if you draw 15 sequences at random from the high-score pool. A higher-throughput in vivo MPRA that can live with more false positives will almost certainly post a lower hit rate.