CpGenie

A deep learning method for predicting DNA methylation level of CpG sites from the sequence context, and predicting non-coding variants' effects on DNA methylation.

Dependencies

• Docker
• GPU support
  • NVIDIA-docker
  • NVIDIA CUDA: currently we support CUDA 7.0 and CUDA 8.0.

Note: You can run CpGenie on CPU by replacing all the nvidia-docker below with docker, but it will be extremely slow and thus highly not recommended.

Predict DNA methylation level of CpG sites

• Prepare a FASTA file (example) of the 1001 bp sequence context centered at the CpG you wish to predict, one sequence for one CpG. The 501 nucleotide should be the 'C' of the CpG.

• Process the FASTA file and make predictions with 50 CpGenie models trained on RRBS datasets from ENCODE immortal cell lines.

   	docker pull haoyangz/cpgenie:CUDA_VER
   	mkdir -p OUTPUT_DIR
   	nvidia-docker run -u $(id -u) -v FASTA_FILE:/in.fa -v OUTPUT_DIR:/outdir \
   			--rm haoyangz/cpgenie:CUDA_VER python main.py ORDER -cpg_fa /in.fa -cpg_out /outdir

  • CUDA_VER: 'cuda7.0' or 'cuda8.0' depending on your NVIDIA driver version.
  • FASTA_FILE: the absolute path to the FASTA file of sequences to predict.
  • OUTPUT_DIR: the absolute path to the output directory, under which the prediction from each of the 50 CpGenie models will be saved.
  • ORDER: the following orders can be both used and seperated by space:
    • -embed: data preprocessing. The output will be saved under $OUTPUT_DIR/embedded.h5.
    • -cpg: make prediction and save under $OUTPUT_DIR/CpGenie_pred.
  • If you wish to only predict for a subset of the 50 CpGenie models

    • Download the 50 CpGenie models to a customized directory (e.g. YOUR_MODEL_DIR, an absolute path)

       	mkdir -p YOUR_MODEL_DIR
       	cd YOUR_MODEL_DIR
       	wget http://gerv.csail.mit.edu/CpGenie_models.tar.gz
       	tar -zxvf CpGenie_models.tar.gz

    • Keep only the models you want

    • Run the following instead:

       	docker pull haoyangz/cpgenie:CUDA_VER
       	mkdir -p OUTPUT_DIR
       	nvidia-docker run -u $(id -u) -v FASTA_FILE:/in.fa -v OUTPUT_DIR:/outdir \
       			-v YOUR_MODEL_DIR:/modeldir --rm haoyangz/cpgenie:CUDA_VER \
       			python main.py ORDER -cpg_fa /in.fa -cpg_out /outdir \
       				-modeltop /modeldir/models

Predict the functional score of sequence variants

• Prepare the variants to score in VCF file (example).

• Score each variant by the predicted impact on DNA methylation in 50 RRBS datasets.

   	docker pull haoyangz/cpgenie:CUDA_VER
   	mkdir -p OUTPUT_DIR
    	nvidia-docker run -u $(id -u) -v VCF_FILE:/in.vcf -v OUTPUT_DIR:/outdir \
   			--rm haoyangz/cpgenie:CUDA_VER python main.py ORDER -var_vcf /in.vcf -var_outdir /outdir

  • CUDA_VER: 'cuda7.0' or 'cuda8.0' depending on your NVIDIA driver version.
  • VCF_FILE: the absolute path to the VCF file to score.
  • OUTPUT_DIR: the absolute path to the output directory.
  • ORDER: the following orders can be both used and seperated by space:
    • -var_prep: find all CpG sites within 500 bp to each variant and prepare the right format under $OUTPUT_DIR/CpGenie_processed.
    • -var_score: make predictions on the data preprocessed in the previous step. For each variant, the predicted absolute change of the following metrics are generated for each of the 50 RRBS datasets, resulting in a 250-dim feature vector. The output will be saved as $OUTPUT_DIR/CpGenie_var_pred, where the first line is a header of feature names and then one line for each variant's 250-dim feature vector.
      • sum of methylation within 500 bp
      • max methylation within 500 bp
      • log odds of the max methylation within 500 bp
      • mean methylation within 500 bp
      • log odds of the mean methylation within 500 bp