Design of tissue-specific mammalian enhancers that function in vivo

This repository contains all the source code, trained models, and model training and testing datasets associated with the study. The source code used to train models, perform transfer learning across different data modalities, generate predictions, compute contribution tracks, and design synthetic enhancers is provided in DeepSTARR_Mouse_codes.zip. The archive preserves the original project directory structure and contains the following modules: bin/ – auxiliary functions and shared utilities Accessibility_model_training/ – scripts for training chromatin accessibility prediction models Enhancer_activity_model_training/ – scripts for training enhancer activity prediction models Ledidi_enhancer_design/ – scripts for Ledidi-based synthetic enhancer design Model_evaluation/ – scripts for model evaluation and control model training replacement_transfer_learning/ – scripts for replacement-based transfer learning experiments motif_syntax_rules_analysis/ – scripts used to analyze motif syntax rules Additional code used for initial data curation, preprocessing, data augmentation, figure generation, and downstream analyses is provided in chen_loubiere_2025_git-master.zip. The trained models are provided in DeepSTARR_Mouse_models.zip, which preserves the original project directory structure. The archive contains: accessibility_models/ – trained chromatin accessibility prediction models enhancer_activity_models/ – trained enhancer activity prediction models K27Ac_K4me1_models/ – enhancer activity prediction models trained using datasets defined by H3K27ac and H3K4me1 signals promoter5kb_noCTCF_models/ – enhancer activity prediction models trained using promoter-distal regions excluding CTCF-associated sequences tissueSpecATAC/ – enhancer activity prediction models trained using tissue-specific accessible chromatin regions The datasets used for model training and testing are provided in DeepSTARR_Mouse_dataset.zip, which preserves the original project directory structure. The archive contains: accessibility_model_dataset/ – datasets used to train chromatin accessibility prediction models enhancer_activity_model_dataset/ – datasets used to train enhancer activity prediction models K27Ac_K4me1_model_dataset/ – histone modification-derived datasets used to train enhancer activity prediction models promoter5kb_noCTCF_model_dataset/ – datasets derived from promoter-distal regions excluding CTCF-associated sequences tissue_specific_ATAC_seq_model_dataset/ – datasets derived from tissue-specific accessible chromatin regions testing_dataset/ – datasets used for model testing and evaluation Software requirements and usage instructions are provided in the README files included in the corresponding archives. For easier access and browsing, the repository is also available on GitHub. Code and tutorials for model training, prediction, and evaluation are available at: https://github.com/Shenzhi-Chen/DeepSTARR-Mouse. Additional code for initial data curation, preprocessing, data augmentation, figure generation, and downstream analyses is available at: https://github.com/vloubiere/chen_loubiere_2025_git. The complete datasets are available on Hugging Face at: https://huggingface.co/datasets/Shenzhi-Chen/DeepSTARR-Mouse-dataset, and the trained models are available at: https://huggingface.co/Shenzhi-Chen/DeepSTARR-Mouse.