DeepCT

DeepCT can learn complex interconnections of epigenetic features and infer unmeasured data from any available input. Furthermore, it can learn cell type-specific properties, build biologically meaningful vector representations of cell types, and utilize these representations to generate cell type-specific predictions of the effects of non-coding variations in the human genome.

Our preprint: Sindeeva et al. Cell type-specific interpretation of noncoding variants using deep learning-based methods.

Tutorial

You can now refer to our Google Colab tutorial to play with our model. For inference on a list of variants please refer to our inference helper repositoryhttps://github.com/AIRI-Institute/DeepCT-inference-helper.

How to run:

1. Install requirements

pip install -r requirements.txt

If you run into any problems with installation of selene, you can also consult selene's official instruction

2. Run scripts specified by config files, e.g.:

python -m selene_sdk model_configs/config_file.yml

Data

There are several scripts to download and process the data.

• data/download_boix_et_al_data.sh to download raw *.bigWig files
• data/process_boix_et_al_data.py to preprocess the *.bigWig files
• data/annotate_biox_data.py to produce selene-style .bed file containing intervals of present track peaks.
• data/generate_target_files.py to produce dataset files to specify in configuration files to run training (see data/README.md for more info).

Training the models

1. Make sure you have prepared your data.
2. Run training using your configuration file, for example model_configs/train_example.yml.

Config files for benchmarks mentioned in the paper (most configs assume necessary data files have already been generated and placed into a folder dataset_data):

• cell type specificity: model_configs/boix_train_masked_ct.yml (classification).

  Note: Before running this config you should also generate all necessary data files for tracks forming a maximal clique and place them into a folder dataset_data/maximal_clique_with0 and additionally run src/split_intervals.py and src/split_cell_types.npy and place their results in the same folder;

• sequence specificity: model_configs/boix_all_multi_ct_train.yml (classification) and model_configs/boix_allNonTreated_highqual_multi_ct_qDeepCT_train_2021-11-15-03-30-59.yml (regression);

• masked tracks prediction: model_configs/boix_allNonTreated_highqual_masked_multi_ct_qDeepCT_train.yml.

Inference

1. Prepare pre-trained model and input files.
2. Run inference using your configuration file, for example model_configs/inference_example.yml.

Alternatively, you might want to use our helper pipeline. It comes with the trained model and scripts to yield simple outputs that can be used to annotate VCFs.

Config file for variant effect predictions mentioned in the paper: model_configs/boix_all_multi_ct_predict.yml

Copyright

Provided under Apache License 2.0.

© 2022 Autonomous Non-Profit Organization "Artificial Intelligence Research Institute" (AIRI). All rights reserved.