Train

The train function will train a maxATAC model using the supplied ATAC-seq and ChIP-seq inputs. The inputs are organized by a tab-delimited text file described below:

Column Name	Description
Cell_Line	Sample Cell Type
TF	Gene Symbol for TF
ATAC_Signal_File	Path to ATAC-seq bigwig signal track
Binding_File	Path to ChIP-seq bigwig signal track
ATAC_Peaks	Path to ATAC-seq peak bed file
CHIP_Peaks	Path to ChIP-seq peak bed file
Train_Test_Label	Train or Test label

**Note: maxATAC was built with version 2.5.0 of tensorflow – newer versions of tensorflow may not be compatible. Therefore, if you experience errors in running train, check the version of tensorflow installed in your environment.

Approach

The meta file described above is used to locate all the input files for all the training data.

General Steps:

1) Initialize the training regions of interest pools 2) Initialize a Keras.Sequence object. Each sequence object is specific for training or prediction. The object will then create random batches of regions of interest from the input ROI pool with the correct ATAC-seq and ChIP-seq signals. 3) Fit the model for the given # of epochs 4) Select best model based on dice coefficient and save the results of training

Training, Validation, and Test Data Splits

For each TF model, all the ATAC-seq and ChIP-seq peaks are pooled into training, validation, and test groups.

Each model is trained on 100 batches of 1,000 examples per epoch. Each batch is composed of randomly chosen peaks that are then randomly assigned to cell types.

Training on ATAC-seq and ChIP-seq peaks is considered “peak-centric” training.

Training on multiple cell types per batch that are randomly assigned peaks is called “pan cell” training.

For every TF model, one cell type and 2 chromosomes are held out for independent testing.

Example

maxatac train --arch DCNN_V2 --sequence hg38.2bit --meta_file CTCF_meta.tsv --output ./CTCF_DCNN --prefix CTCF_DCNN --shuffle_cell_type --rev_comp

Required Arguments

--genome

Specify which genome build this task is specified for (i.e. hg38).

--sequence

This argument specifies the path to the 2bit DNA sequence for the genome of interest

--meta_file

This argument specifies the path to the meta file that describes the training data available. This meta file described above.

Optional Arguments

--prefix

This argument is reserved for the prefix used to build the output filename. This can be any string. The extension .bw will be added to the filename prefix. Default: maxatac_model

--train_roi

This argument is used to input the bed file that you want to use to define the training regions of interest. If you set this option you will randomly select regions from this file for training instead of using the meta data to build the training data pool.

--validate_roi

This argument is used to input the bed file that you want to use to define the validation regions of interest. If you set this option you will randomly select regions from this file for validation instead of using the meta data to build the validation data pool.

--target_scale_factor

The scaling factor for scaling model targets signal. Used only for quantitative models. Default: 1

--output_activation

The output activation to use for the model. No other options are considered in the publication. Test at your own risk. Default: sigmoid

--chroms

The list of chromosomes to limit the study to. These include the training and validation chromosomes. Default: ["chr2", "chr3", "chr4", "chr5", "chr6", "chr7", "chr9", "chr10", "chr11", "chr12", "chr13", "chr14", "chr15", "chr16", "chr17", "chr18", "chr19", "chr20", "chr21", "chr22", "chrX"]

--tchroms

The list of chromosomes to use for training only. Default: ["chr3", "chr4", "chr5", "chr6", "chr7", "chr9", "chr10", "chr11", "chr12", "chr13", "chr14", "chr15", "chr16", "chr17", "chr18", "chr20", "chr21", "chr22"]

--vchroms

The list of chromosomes to use for validation only. Default: ["chr2", "chr19"]

--arch

The architecture to use for the neural network. Default: DCNN_V2

--rand_ratio

The proportion of random regions to use per training and validation batch. This corresponds to the number of regions that are randomly selected form the genome as opposed to being created based on the ATAC-seq or ChIP-seq peaks. Default: 0

--seed

The seed to use for the model in case of reproducibility. Default: random.randint(1, 99999)

--weights

The weights to use to initialize a model. Default: do not initialize with weights

--epochs

The number of epochs to train the model for. Default: 20

--batches

The number of batches to use per stochastic gradient descent step. Default: 100

--batch_size

The number of examples to use per training batch. Default: 1000

--val_batch_size

The number of examples to use per validation batch. Default 1000

--output

The output directory name to save results to. Default: ./training_results

--plot

Whether to plot the model structure or training history. Default: True

--dense

Used if you want to use a dense layer at the end of the neural network. Default: False

--threads

Use to change the number of threads used per job. Default: get available

--loglevel

This argument is used to set the logging level. Currently, the only working logging level is ERROR.

--rev_comp

If rev_comp_train, then use the reverse complement sequence in addition to the reference sequence. Default: False

--shuffle_cell_type

If shuffle_cell_type, then shuffle training ROI cell type label. This is related to “pan-cell” training as described in the maxATAC manuscript. Default: True