Fibertools: training supervised and semi-supervised CNN models for m6A detection from Fiber-seq HiFi reads.

This repository is for training supervised and semi-supervised convolutional neural network for m6A detection from Fiber-seq reads.

Install

python -m pip install git+https://github.com/mrvollger/m6a-calling

Reinstall

pip uninstall m6a-calling && pip install git+https://github.com/mrvollger/m6a-calling

Make training data

m6adata \
  --threads 20 \ 
  --hifi fiberseq.bam \ # input fiberseq bam file
  -o output.ml.npz \ # training data for a ML model
  --train \ # must be included to make training data
  -s 0.03 \ # sample just 3% of the data
  --is_u16 \ # has u16 kinetics values 
  -m 244 \ # min ML value to include in training
  --ec 6 \ # minimum CCS coverage to use in training

Download data for existing models

Download training and validation data for all three chemistries from Zenodo and set up requisite folders for saving models and results with the following instructions:

mkdir -p paper_v1/models
mkdir -p paper_v1/results
mkdir -p paper_v1/figures

mkdir -p paper_v1/data
cd paper_v1/data

wget https://zenodo.org/record/7809229/files/fibertoolsMLTrain.tar.gz
tar -zxvf fibertoolsMLTrain.tar.gz

Ensure that the sup_train_data, sup_val_data, semi_train_data and semi_val_data variables in the config.yml file are pointing to the correct folder.

Train supervised CNN model

You can run all three versions of the fully supervised CNN model with the following commands:

m6a_supervised_cnn --config_file paper_v1/config.yml --train_chem train_2_2_chemistry

m6a_supervised_cnn --config_file paper_v1/config.yml --train_chem train_3_2_chemistry

m6a_supervised_cnn --config_file paper_v1/config.yml --train_chem train_revio_chemistry

All required resources are defined in the config.yml file. See Train new Fiber-seq chemistry models section for more details on the resources.

Train semi-supervised CNN model

To run the semi-supervised CNN models, run the following commands:

m6a_semi_supervised_cnn --config_file paper_v1/config.yml --train_chem train_2_2_chemistry

m6a_semi_supervised_cnn --config_file paper_v1/config.yml --train_chem train_3_2_chemistry

m6a_semi_supervised_cnn --config_file paper_v1/config.yml --train_chem train_revio_chemistry

Inference on semi-supervised CNN model

To generate precision at different CNN scores for all chemistries, run the following commands:

m6a_semi_supervised_cnn_predict --config_file paper_v1/config.yml --train_chem train_2_2_chemistry

m6a_semi_supervised_cnn_predict --config_file paper_v1/config.yml --train_chem train_3_2_chemistry

m6a_semi_supervised_cnn_predict --config_file paper_v1/config.yml --train_chem train_revio_chemistry

Train new Fiber-seq chemistry models

To train semi-supervised CNN model for a new chemistry, you first need to generate a config file entry. The configuration file is divided into three sections, one section for each new chemistry. For training a model for a new chemistry, simply add the required resources in the config.yml file under a new header. Then run the following commands:

m6a_supervised_cnn --config_file paper_v1/config.yml --train_chem <train_new_chemistry>

m6a_semi_supervised_cnn --config_file paper_v1/config.yml --train_chem <train_new_chemistry>

m6a_semi_supervised_cnn_predict --config_file paper_v1/config.yml --train_chem <train_new_chemistry>

Here is an example resources section:

[train_new_chemistry]
fdr=0.05
device=cuda
sup_lr=0.0001
semi_lr=0.0001
input_size=6
input_length=15
semi_batch_size=32
semi_num_workers=2
sup_batch_size=32
sup_num_workers=8
train_sample=1
train_sample_fraction=0.4
val_sample=1
val_sample_fraction=0.25
min_pos_proportion=0.0001
add_m6a_percent=1.0
total_m6a_percent=70.0
supervised_train_epochs=30
semi_supervised_train_epochs=20
sup_train_data=paper_v1/data/PS00109_2.npz
sup_val_data=paper_v1/data/PS00109_3_val.npz
semi_train_data=paper_v1/data/PS00109_2_semi_train.npz
semi_val_data=paper_v1/data/PS00109_3_semi_val.npz
best_supervised_model_name=paper_v1/models/m6A_2_2_supervised_cnn.best.torch.pickle
final_supervised_model_name=paper_v1/models/m6A_2_2_supervised_cnn.final.torch
pretrain_model_name=paper_v1/models/m6A_2_2_supervised_cnn.best.torch.pickle
best_semi_supervised_model_name=paper_v1/models/m6A_2_2_semi_supervised_cnn.best.torch.pickle
final_semi_supervised_model_name=paper_v1/models/m6A_2_2_semi_supervised_cnn.final.torch.pickle
save_pos=paper_v1/results/semi_train_2_2_chemistry.npz
score_ap_json=paper_v1/results/semi_2_2_score_precision.json
score_ap_table=paper_v1/results/semi_2_2_score_precision.tsv

And here are the explanations for each variable:

fdr: acceptable false discovery rate for training the semi-supervised model
device: training device for models, cpu or cuda
input_size: number of input channels, currently 4 nucleotides, IPD and PW. 
input_length: length of every input, currently 15 nucleotides. 
semi_batch_size: Batch size for training semi-supervised CNN model. 
semi_num_workers: Number of worker threads for preprocessing data for semi-supervised CNN model. 
sup_batch_size: Batch size for training supervised CNN model. 
sup_num_workers: Number of worker threads for preprocessing data for supervised CNN model.
train_sample: Whether to subsample training data, 0-> false, 1 -> True
train_sample_fraction: Fraction of training data to subsample. 
val_sample: Whether to subsample validation data, 0-> false, 1 -> True
val_sample_fraction: Fraction of validation data to subsample. 
min_pos_proportion: minimum proportion of m6A calls above this threshold, ensures a good initialization for the semi-supervised training. 
add_m6a_percent: minimum proportion of addition m6A calls an iteration of semi-supervised model should find to continue training 
total_m6a_percent: Minimum proportion of total m6A calls the semi-supervised model needs to stop training.
supervised_train_epochs: number of training epochs for the supervised approach
semi_supervised_train_epochs: number of training epochs for the semi-supervised approach
sup_train_data: path to supervised training data
sup_val_data: path to supervised validation data
semi_train_data: path to semi-supervised training data
semi_val_data:  path to semi-supervised validation data
best_supervised_model_name: path and name for best supervised model
final_supervised_model_name: path and name for final supervised model
best_semi_supervised_model_name: path and name for best semi-supervised model
final_semi_supervised_model_name: path and name for final semi-supervised model
save_pos: path to save validation AUPR and # positives identified in validation set for semi-supervised approach
score_ap_json: json for cnn score to u8 precision conversion
score_ap_table: table with cnn scores, precision and u8 precision

fiberseq/train-m6A-calling