Enformer Celltyping

TL;DR - A deep learning model which incorporates distal effects of DNA interactions, up to 100,000 base-pairs away, to predict epigenetic signals in previously unseen cell types using DNA and chromatin accessibility data.

Detailed Explanation

Enformer Celltyping is a tensorflow, multi-headed attention based model that incorporates distal effects of Deoxyribonucleic Acid (DNA) interactions to predict histone marks across diverse cell types.

Enformer Celltyping uses DNA and chromatin accessibility information (derived from ATAC-Seq experiments) to predict histone mark signals (H3k27ac, H3k4me1, H3k4me3, H3k9me3, H3k27me3, H3k36me3).

The prediction of histone marks, transcription factors and gene expression from DNA was progressed by Enformer. A multi-headed attention model which can consdier long-range DNA interactions (up to ~100k base-pairs away). However, a key limitation is that Enformer can only predict on the cell types upon which it was trained and thus is primarily useful for interpretation of its output, for example in silico saturated mutagenesis of DNA sequence. Here, we extend these long-range capabilities through a transfer learning approach but also enable the prediction in cell types outside of the training set by incorporating chromatin acessibility information.

Installation

First, clone this repository and move to the directory.

git clone https://github.com/neurogenomics/EnformerCelltyping && cd EnformerCelltyping

To use Enformer Celltpying, dealing with it's dependencies, you should install conda package manager.

Once conda is installed, place the conda executable in PATH, use the following command to create an environment named EnformerCelltyping with all necessary dependencies and activate it. Note this may take some time to complete:

Create the necessary environments using the yml files, then activate the Enformer Celltyping environment:

conda env create -f ./environment/enformer_celltyping.yml &&\
make renv &&\ 
make pyanalyenv &&\
conda activate EnformerCelltyping &&\
pip install -e .

Next all use-cases for Enformer Celltyping involve transfer learning from the enformer model (with frozen weights) and require some other files to run. Also, if you want to use the trained version of Enformer Celltyping, the weights will need to be downloaded. So download these with the following with the EnformerCelltyping conda env activated:

python bin/download_Enformer_Celltyping_dependencies.py

See using_enformer_celltyping for a detailed explanation of all of these files. Downloaded data can also be viewed at source on figshare.

Enformer Celltyping - how can it be used for research?

Enformer Celltyping has two main usage categories:

1. You want to use the trained Enformer Celltyping model to predict. For example, you may have ATAC-seq (bulk or scATAC-seq which you have pseudobulked) for a cell type/tissue of interest and you would like to predict histone mark tracks for it. See Using & interpretting Enformer Celltyping section below for this and other uses for the trained Enformer Celltyping model.
2. You are interested in training an Enformer Celltyping model on your own data so would like to see the trianing process. See Training Enformer Celltyping.

Training Enformer Celltyping

To understand how to train Enformer Celltyping on your own data, see training_demo.ipynb which steps through training the model on demo data. Or see full_training_recreation.ipynb for the full set of scripts to download, preprocess and train Enformer Celltyping as outlined in our manuscript.

Using & interpretting Enformer Celltyping

To see all use cases for the pre-trained Enformer Celltyping model, see using_enformer_celltyping. The use cases included are:

• Predict histone marks for a cell type of interest.
• Testing the receptive field of a model.
• Visualise Enformer Celltyping's cell type embedding.
• Predict the effect of a genetic variant.

Reproducing hQTL SNP effect prediction analysis

Reproducing the hQTL SNP effect predicition analysis from our paper includes predicting the effect of a genetic variants as explained in using_enformer_celltyping. However, since this requires predicting the effect for >850k SNPs for every cell type of interest, we used code to checkpoint partial runs in case of any errors, precomputed DNA embeddings for all SNPs and set up runs in parallel. The analysis also required preprocessing of the hQTL datasets. See reproducing_hQTL_SNP_effect_prediction for details and related scripts.

Reproducing cell type motif enrichment

Reproducing the cell type motif enrichment analysis from our paper includes predicting across the entire genome for cell types as explained in using_enformer_celltyping. We then calculcate the gradient of all peak predicted regions on the global input and compare to identify peaks that rely the most on the global signal, do motif enrichment and finally inspect the cell type specificity of the identifed transcription factors. See reproducing_cell_type_motif_enrichment for details and related scripts.

Reproducing LDSC enrichment

Reproducing the cell type LDSC enrichment analysis from our paper includes predicting across the entire genome for cell types as explained in using_enformer_celltyping. See reproducing_ldsc_enrichment for details and related scripts.

Cite

If you use this work, please cite our paper.