CREATE

A stretch of non-coding DNA can silence a gene in one cell type and activate it in another, and its sequence is identical in both. Silencers and enhancers share most of their sequence grammar, which is why methods that read sequence alone — and that typically answer one binary question, "is this an enhancer?" — cannot say which role a region plays in the tissue in front of you. CREATE (Cis-Regulatory Element identificAtion via discreTe Embedding) lets the cell type enter through the data instead: each 1000-bp candidate arrives as one-hot DNA stacked with per-base-pair chromatin accessibility and chromatin loop scores measured in that cell type, and the model returns one of five labels — silencer, enhancer, promoter, insulator, or background.

The second idea is the discrete embedding. Rather than compressing a region into a continuous latent vector, CREATE quantizes the fused representation against a shared, dynamically updated codebook, so every element is a combination of a small number of reusable code words. How often each code word appears in each CRE class yields what the authors call a CRE-specific feature spectrum — a quantitative account of what distinguishes a silencer from an enhancer, instead of a saliency map over a black box.

CREATE was developed in Tsinghua University's Ministry of Education Key Laboratory of Bioinformatics and Department of Automation, with collaborators at Nankai University and Stanford University, and published in Nature Communications in May 2025 after an October 2024 preprint. It sits alongside sequence-only chromatin predictors such as DeepSEA and shares its design posture with UniversalEPI: one fixed checkpoint, with cell-type specificity supplied by the input tracks rather than per-cell-type weights.

#Key Features

• Five-way classification, not five binary classifiers: silencers, enhancers, promoters, insulators and background are resolved in one model, forcing the hardest pair — silencers against enhancers — apart explicitly.
• Cell-type context through the inputs: accessibility comes from DNase/ATAC read counts and interaction from HiChIP loop counts, both cell-type-matched, so a new cell type needs new tracks rather than new training.
• Discrete, interpretable embeddings: a 200-entry codebook with split quantization gives each region a code-word composition readable as a feature spectrum.
• Configurable input modalities: three released checkpoints cover sequence only, sequence plus accessibility, and sequence plus accessibility plus loops, so the model degrades gracefully when a track is unavailable.
• Genome-scale scoring: the released weights run as a forward pass over user-supplied regions, making whole-genome screens practical.

#Technical Details

CREATE is a convolutional VQ-VAE. Modality-specific encoders (convolution, max-pooling, ReLU, dropout) feed an integration encoder; the latent is split into 16 sub-vectors of dimension 8, each replaced by its nearest neighbour in a shared 200-entry codebook updated by exponential moving average. Decoders reconstruct all three modalities and a three-layer head (3328 → 1024 → 512 → 5) reads the quantized embedding. Training begins with 50 reconstruction-only epochs before the classification loss is added; the full checkpoint holds roughly 6.8 million parameters.

Labels come from SilencerDB, FANTOM5, EPD and ENCODE CTCF peaks on hg19 — in K562, 6,754 silencers, 10,528 enhancers, 15,699 promoters, 18,631 insulators and 20,000 background regions, with a silencer-poorer set in HepG2. Under 10-fold cross-validation on K562, CREATE reaches a macro-averaged auROC of 0.964 ± 0.002 against 0.928 ± 0.002 for ES-transition, and a macro auPRC of 0.848 ± 0.004 against 0.743 ± 0.003 for DeepICSH; on silencers alone the auPRC is 0.549 ± 0.017 versus 0.410 ± 0.012. HepG2 shows a 9.1% auPRC improvement, and the model was also applied to GM12878 and HeLa-S3, which lack validated silencers.

#Applications

The intended use is screening the non-coding genome of a cell type that already has accessibility and interaction data. The authors scored 270,259 candidate regions in K562 and 232,456 in HepG2; the calls carried the expected histone signatures — H3K27me3 in predicted silencers, H3K27ac and H3K4me1 in predicted enhancers — and the expected transcription factor motifs. Comparing the two cell types surfaced 2,409 dual-function elements acting as silencers in K562 and enhancers in HepG2. Predicted elements were enriched for rare SNPs and whole-blood eQTLs and carried heritability for blood traits, a usable prior for prioritizing non-coding variants.

#Impact

CREATE demonstrates that multi-class CRE identification is tractable and that discrete embeddings buy interpretability without costing accuracy — particularly for silencers, which are under-annotated relative to enhancers and which most existing tools do not model at all. Its scope is genuinely narrow: training and benchmarking rest on two cell lines, the input window is fixed at 1000 bp, and the only demonstrated output is the five-way label, with no transfer to other tasks shown. A new cell type requires matched accessibility and HiChIP data, though the sequence-only checkpoint is a fallback when those are missing. Code and the three checkpoints are released under the MIT license on GitHub, with the paper's version archived on Zenodo.