Enformer transfer-learned on single-cell ATAC-seq to predict chromatin accessibility in 151 human cell types directly from DNA sequence.
No providers recorded yet. Browse all providers
Bulk DNase-seq and ATAC-seq profiles average accessibility over every cell in a sample, so an enhancer that opens only in one lineage is diluted by the cells in which it is closed. Enformer learned to predict hundreds of such bulk tracks from DNA sequence, which gives it a strong general model of regulatory grammar but leaves its accessibility outputs tied to tissue-level mixtures. REnformer asks whether that learned grammar can be redirected at cleaner, cell-type-resolved targets: it keeps Enformer's sequence encoder and retrains the output layer on pseudobulked single-cell ATAC-seq (scATAC-seq) profiles for 151 human cell types.
REnformer was developed by Simone G. Riva, Jim R. Hughes and colleagues in the MRC Molecular Haematology Unit at the MRC Weatherall Institute of Molecular Medicine, University of Oxford, with a co-author at Bocconi University. It was posted to bioRxiv in June 2025 and published at the IEEE Conference on Computational Intelligence in Bioinformatics and Computational Biology (CIBCB 2025). The project README notes that it was developed in collaboration with Nucleome Therapeutics, which Hughes co-founded.
The model is the accessibility-prediction member of the Hughes lab's regulatory-genomics toolkit. Its later sibling CREP also builds on Enformer but emits discrete cis-regulatory element classes rather than continuous accessibility signal; the two are separate models with separate checkpoints and papers.
REnformer retains Enformer's architecture of 7 convolutional blocks with pooling, 11 transformer blocks, and a cropping layer followed by a pointwise convolution. The authors converted the TensorFlow Enformer weights released by Avsec et al. into PyTorch and replaced Enformer's output head with a new head predicting 151 tracks. Input is a one-hot-encoded 196,608 bp sequence; output covers the central 114,688 bp (after cropping 40,960 bp from each side) in 896 bins of 128 bp.
Training data were collected from 151 GRCh38 scATAC-seq cell types: 7 from Turner et al., 107 from the Zhang et al. single-cell chromatin atlas, 15 from Granja et al., 18 PBMC types from public 10x Genomics datasets, and 4 in-house samples (three erythroid differentiation stages and HUVEC). The genome was chunked into 9,406 non-blacklisted 196,608 bp regions, split into 7,099 training and 2,307 test regions, and the model was trained on an NVIDIA L40S GPU. On 2,184 non-empty test regions divided into six subsets, REnformer's error against observed scATAC-seq was significantly lower than Enformer's error against its own DNase targets (one-sided Mann–Whitney U tests with Bonferroni correction) on all three NRMSE metrics in every subset, with the single exception of the min–max-scaled metric in one subset.
REnformer is aimed at regulatory genomics and human genetics groups who want to ask how a DNA change alters chromatin accessibility in a particular cell type. The paper's case study is the α-thalassaemia variant at chr16:159710 (hg38), a T>C gain-of-function SNP that creates a new promoter acting as an orientation-dependent enhancer blocker; subtracting REnformer's reference prediction from its altered-sequence prediction in erythroid cells produced a new accessibility peak at the variant, matching the experimentally established effect. The authors position the model for in silico mutagenesis, prioritising candidate causal variants within GWAS haplotypes, and linking non-coding variants to regulatory elements in combination with the lab's REgulamentary annotation pipeline.
REnformer shows that a thin, cheaply trained head on a frozen Enformer trunk can reorient a bulk-trained sequence model toward single-cell-derived accessibility targets across a wide panel of cell types. Its main contributions are the normalisation scheme for making heterogeneous scATAC-seq data trainable and the NRMSE benchmark for comparing Enformer derivatives. The evaluation compares REnformer only against Enformer, and against different target assays (scATAC-seq versus DNase), so it does not establish how the model ranks against other accessibility predictors such as ChromBPNet. Training and analysis code are public under an academic non-commercial licence, but the trained REnformer checkpoint is not distributed with the repository.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.