CREsted

Cell identity is written into non-coding enhancers as combinations of transcription factor binding sites, and sequence-to-function neural networks have become the standard way to read that grammar. Until recently each such model was a bespoke artifact: one dataset, one architecture, one paper. CREsted (Cis-Regulatory Element Sequence Training, Explanation, and Design) standardizes that workflow end to end and distributes a large zoo of pretrained enhancer models that apply to new DNA sequences immediately.

CREsted comes from Stein Aerts's Laboratory of Computational Biology at the VIB Center for AI & Computational Biology and the Department of Human Genetics at KU Leuven, led by Niklas Kempynck and Seppe De Winter. It was posted to bioRxiv in April 2025 and published in Nature Methods. The package covers four stages — preprocessing single-cell ATAC-seq into per-cell-type accessibility targets, training a sequence model, explaining it at nucleotide resolution, and designing synthetic enhancers — and is built on Keras 3, so it runs on TensorFlow or PyTorch and integrates with scverse.

The released checkpoints are what make CREsted more than a training harness. crested.get_model(name) downloads a fixed pretrained model, crested.utils.load_model instantiates it, and crested.tl.predict scores any sequence with no retraining. The registry ships roughly 58 checkpoints — among them DeepPBMC, DeepZebraFish, DeepBICCN and DeepBICCN2, DeepMouseBrain1–3, DeepHumanCortex1–2, DeepChickenBrain1–2, DeepFlyBrain, DeepMEL1/2, DeepGlioma, and the DeepLiver models — plus CREsted-hosted Enformer and Borzoi weights and a BorzoiBICCN transfer-learned variant.

#Key Features

• Pretrained model zoo: Dozens of checkpoints across mouse, human, zebrafish, fly, and chicken are fetched by name and applied directly; DeepPBMC, for example, is an 18.9 M-parameter dilated CNN mapping a 2,114 bp sequence to accessibility across seven human immune cell types.
• Peak-height normalization: A constitutive-peak rescaling step corrects the systematic differences in pseudobulk peak height between cell types that plain CPM normalization leaves behind.
• Nucleotide-resolution explanation: Contribution scores and in silico mutagenesis, with tfmodisco-lite pattern mining and motif-to-transcription-factor matching, recover the motifs driving each cell type's predictions.
• Synthetic enhancer design: In silico evolution and motif implantation optimize sequences against an L2 objective on the full prediction vector, pushing a designed enhancer to be active in the target cell type and silent elsewhere.
• Foundation model transfer: Enformer and Borzoi serve as pretrained trunks, fine-tuned onto a new scATAC-seq dataset through the same API as models trained from scratch.

#Technical Details

CREsted supports topic modeling from pycisTopic and pseudobulk peak-height regression, with a default loss combining cosine similarity and log mean squared error. DeepBICCN2, trained on mouse motor cortex, was pretrained on 440,993 consensus peaks and fine-tuned on 73,326 variable regions; on held-out chromosomes it reached an average Spearman correlation of 0.79 and Pearson r of 0.82 across cell types, significantly outperforming both 6 M- and 22 M-parameter gReLU models. Against 171 in vivo validated BICCN brain enhancers it achieved 0.77 average precision and 0.93 recall, and sliding-window locus scoring generalized to inaccessible intergenic sequence (r = 0.75) and across species, scoring a chicken locus with a mouse interneuron class at r = 0.62.

DeepZebraFish was trained on a developmental atlas of 639 cell-type-by-timepoint classes, using 793,273 consensus peaks with fine-tuning on 89,637 specific peaks, and predicted the correct class for 76% of 54 validated zebrafish enhancers. Enhancers designed with it were tested by Tol2 reporter assay in zebrafish embryos: all three cardiac muscle designs and all three somatic muscle designs drove specific activity at 48 hours post-fertilization, and one of three endothelial designs was strongly specific. The package is distributed on PyPI and GitHub under an academic non-commercial license.

#Applications

CREsted serves two audiences. Groups with their own scATAC-seq atlas get a standardized path from fragments to a trained, interpretable enhancer model and on to designed sequences — useful for cell-type-specific gene therapy vectors, regulatory variant interpretation, or comparing enhancer logic between tumor states, as the paper does for mesenchymal-like cancer cells. Groups without such data get the zoo: any sequence can be scored for accessibility in mouse cortical subclasses, human PBMC types, fly brain types, or zebrafish developmental classes, and any locus swept to nominate candidate enhancers, including in species with no ATAC data of their own.

#Impact

CREsted consolidates a line of enhancer-modeling work from the Aerts lab — DeepFlyBrain, DeepMEL, and the in silico evolution design strategy — into one maintained package with reproducible checkpoints, and its Nature Methods publication and in vivo zebrafish validation put it among the few enhancer design methods verified in a whole organism. The constraints are practical rather than scientific: an academic non-commercial license that excludes industrial use without separate terms, checkpoints hosted on a lab server rather than a model hub, and uneven coverage, so a cell type absent from the zoo still requires training your own — which is what the rest of the package is for.