DeepViFi pipeline detects viral infections in NGS data using the Transformer architecture. It can be trained on specific viral families.
Rajkumar, U. et al. DeepViFi: Detecting Oncoviral Infections in Cancer Genomes using Transformers. ACM BCB. (2022)
| File | Description |
|---|---|
| config.yaml | File to set parameters for different tasks |
| Makefile | Makefile to run the different scripts |
| environment.yml | Requirement file for setting up environment |
| setup.sh | Script to install ecSeg package |
...
| Folder | Description |
|---|---|
| src | Contains python scripts |
| models | Contains transformer models trained on HPV and HBV references. Download the models.zip from here and and unzip the folder inside the deepvifi/ folder |
git clone https://github.com/ucrajkumar/DeepViFi
cd DeepViFi
conda env create -f environment.yml
conda activate deepvifi
The training data must be a single fasta file containing all known reference genoems of the virus of interest. It must be in the form:
> name_of_ref
CACACATACAGACATA...
> name_of_ref2
...
The genomes do not need to be the same length.
Reads the input fasta file containing the reference genomes and creates the training data to train the transformer. Set parameters in config.yaml under create_data:
datapath: Path to folder containing the viral references
viral_refs: Name of fasta file
read_len : Length of read
coverage : Simulate reads at coverage. (recommended value 0.5)
mask_rate : What portion of each input read to mask. (recommended value 0.2)
Trains the transformer given the hyperparameters. Set parameters in config.yaml under train_transformer:
datapath: Path to folder containing the viral references and training data
model_path : Path to folder where you want the model saved
embed_dim : Dimensions to embed the input read. (recommended value 128 or 256)
num_heads : Number of attention heads. (recommended value 8 or 16)
num_layers : Number of transformer blocks. (recommended value 8)
ff_dim : Dimensions of feed forward layer in each transformer block. (recommended value 256)
batch_size : Batch size
Trains a random forest model given the hyperparameters. Set parameters in config.yaml under train_rf:
datapath: Path to folder containing the viral references and training data
model_path : Path to folder where you want the model saved
ck_name : Path to checkpoint folder containing the transformer model. If no checkpoint folder, leave as ''
transformer_model : Name of the trained transformer model. Typically something like "embed128_heads8_h256_fd5_nl8"
num_trees : Number of trees to be trained in the RF model. Recommended value 500
Read/process a test file, and run the transformer and RF on the test data. Set parameters in config.yaml under test_pipeline:
datapath: Path to folder containing the viral references and training data
model_path : Path to folder where you want the model saved
ck_name : Path to checkpoint folder containing the transformer model. If no checkpoint folder, leave as ''
transformer_model : Name of the trained transformer model. Typically something like "embed128_heads8_h256_fd5_nl8"
rf_model : Name of the trained RF model.
test_file : Name of the file to test
file_type : Type of the test file. Can be 'fasta' or 'fastq'.
test_pipeline produces a file called <test_file>_rfpredictions.csv The first column is the read IDs and the second column is the rf prediction. A value close to 1 indicates that the read is highly likely to be viral.