Improving Compound Activity Classification via Deep Transfer and Representation Learning

This repository is the official implementation of Improving Compound Activity Classification via Deep Transfer and Representation Learning.

Requirements

Operating systems: Red Hat Enterprise Linux Server 7.9

To install requirements:

pip install -r requirements.txt

Installation guide

Download the code and dataset with the command:

git clone https://github.com/ninglab/TransferAct.git

Data Processing

1. Use provided processed dataset

One can use our provided processed dataset in ./data/pairs/: the dataset of pairs of processed balanced assays $\mathcal{P}$ . Check the details of bioassay selection, processing, and assay pair selection in our paper in Section 5.1.1 and Section 5.1.2, respectively. We provided our dataset of pairs as data/pairs.tar.gz compressed file. Please use tar to de-compress it.

2. Use own dataset

We provide necessary scripts in ./data/scripts/ with the processing steps in ./data/scripts/README.md.

Training

1. Running TAc

• To run TAc-dmpn,

python code/train_aada.py --source_data_path <source_assay_csv_file> --target_data_path <target_assay_csv_file> --dataset_type classification --extra_metrics prc-auc precision recall accuracy f1_score --hidden_size 25 --depth 4 --init_lr 1e-3 --batch_size 10 --ffn_hidden_size 100 --ffn_num_layers 2 --epochs 40 --alpha 1 --lamda 0 --split_type index_predetermined --crossval_index_file <index_file> --save_dir <chkpt_dir> --class_balance --mpn_shared

• To run TAc-dmpna, add these arguments to the above command

--attn_dim 100 --aggregation self-attention --model aada_attention

source_data_path and target_data_path specify the path to the source and target assay CSV files of the pair, respectively. First line contains a header smiles,target. Each of the following lines are comma-separated with the SMILES in the 1st column and the 0/1 label in the 2nd column.

dataset_type specifies the type of task; always classification for this project.

extra_metrics specifies the list of evaluation metrics.

hidden_size specifies the dimension of the learned compound representation out of GNN-based feature generators.

depth specifies the number of message passing steps.

init_lr specifies the initial learning rate.

batch_size specifies the batch size.

ffn_hidden_size and ffn_num_layers specify the number of hidden units and layers, respectively, in the fully connected network used as the classifier.

epochs specifies the total number of epochs.

split_type specifies the type of data split.

crossval_index_file specifies the path to the index file which contains the indices of data points for train, validation and test split for each fold.

save_dir specifies the directory where the model, evaluation scores and predictions will be saved.

class_balance indicates whether to use class-balanced batches during training.

model specifies which model to use.

aggregation specifies which pooling mechanism to use to get the compound representation from the atom representations. Default set to mean: the atom-level representations from the message passing network are averaged over all atoms of a compound to yield the compound representation.

attn_dim specifies the dimension of the hidden layer in the 2-layer fully connected network used as the attention network.

Use python code/train_aada.py -h to check the meaning and default values of other parameters.

2. Running TAc-fc variants and ablations

• To run Tac-fc,

python code/train_aada.py --source_data_path <source_assay_csv_file> --target_data_path <target_assay_csv_file> --dataset_type classification --extra_metrics prc-auc precision recall accuracy f1_score --hidden_size 25 --depth 4 --init_lr 1e-3 --batch_size 10 --ffn_hidden_size 100 --ffn_num_layers 2 --local_discriminator_hidden_size 100 --local_discriminator_num_layers 2 --global_discriminator_hidden_size 100 --global_discriminator_num_layers 2 --epochs 40 --alpha 1 --lamda 1 --split_type index_predetermined --crossval_index_file <index_file> --save_dir <chkpt_dir> --class_balance --mpn_shared

• To run TAc-fc-dmpna, add these arguments to the above command

--attn_dim 100 --aggregation self-attention --model aada_attention

Ablations

• To run TAc-f, add --exclude_global to the above command.
• To run TAc-c, add --exclude_local to the above command.
• Adding both --exclude_local and --exclude_global is equivalent to running TAc.

3. Running Baselines

DANN

python code/train_aada.py --source_data_path <source_assay_csv_file> --target_data_path <target_assay_csv_file> --dataset_type classification --extra_metrics prc-auc precision recall accuracy f1_score --hidden_size 25 --depth 4 --init_lr 1e-3 --batch_size 10 --ffn_hidden_size 100 --ffn_num_layers 2 --global_discriminator_hidden_size 100 --global_discriminator_num_layers 2 --epochs 40 --alpha 1 --lamda 1 --split_type index_predetermined --crossval_index_file <index_file> --save_dir <chkpt_dir> --class_balance --mpn_shared

• To run DANN-dmpn, add --model dann to the above command.
• To run DANN-dmpna, add --model dann_attention --attn_dim 100 --aggregation self-attention --model to the above command.

Run the following baselines from chemprop as follows:

FCN-morgan

python chemprop/train.py --data_path <assay_csv_file> --dataset_type classification --extra_metrics prc-auc precision recall accuracy f1_score --init_lr 1e-3 --batch_size 10 --ffn_hidden_size 100 --ffn_num_layers 2 --epochs 40 --features_generator morgan --features_only --split_type index_predetermined --crossval_index_file <index_file> --save_dir <chkpt_dir> --class_balance

FCN-morganc

python chemprop/train.py --data_path <assay_csv_file> --dataset_type classification --extra_metrics prc-auc precision recall accuracy f1_score --init_lr 1e-3 --batch_size 10 --ffn_hidden_size 100 --ffn_num_layers 2 --epochs 40 --features_generator morgan_count --features_only --split_type index_predetermined --crossval_index_file <index_file> --save_dir <chkpt_dir> --class_balance

FCN-dmpn

python chemprop/train.py --data_path <assay_csv_file> --dataset_type classification --extra_metrics prc-auc precision recall accuracy f1_score --hidden_size 25 --depth 4 --init_lr 1e-3 --batch_size 10 --ffn_hidden_size 100 --ffn_num_layers 2 --epochs 40 --split_type index_predetermined --crossval_index_file <index_file> --save_dir <chkpt_dir> --class_balance

FCN-dmpna

Add the following to the above command:

--model mpnn_attention --attn_dim 100 --aggregation self-attention

For the above baselines, data_path specifies the path to the target assay CSV file.

FCN-dmpn(DT)

python chemprop/train.py --data_path <source_assay_csv_file> --target_data_path <target_assay_csv_file> --dataset_type classification --extra_metrics prc-auc precision recall accuracy f1_score  --hidden_size 25 --depth 4 --init_lr 1e-3 --batch_size 10 --ffn_hidden_size 100 --ffn_num_layers 2 --epochs 40 --split_type index_predetermined --crossval_index_file <index_file> --save_dir <chkpt_dir> --class_balance

FCN-dmpna(DT)

--model mpnn_attention --attn_dim 100 --aggregation self-attention

For FCN-dmpn(DT)and FCN-dmpna(DT), data_path and target_data_path specify the path to the source and target assay CSV files.

Use python chemprop/train.py -h to check the meaning of other parameters.

Testing

1. To predict the labels of the compounds in the test set for Tac*, DANN methods:

   python code/predict.py --test_path <test_csv_file> --checkpoint_dir <chkpt_dir> --preds_path <pred_file>

   test_path specifies the path to a CSV file containing a list of SMILES and ground-truth labels. First line contains a header smiles,target. Each of the following lines are comma-separated with the SMILES in the 1st column and the 0/1 label in the 2nd column.

   checkpoint_dir specifies the path to the checkpoint directory where the model checkpoint(s) .pt filles are saved (i.e., save_dir during training).

   preds_path specifies the path to a CSV file where the predictions will be saved.

2. To predict the labels of the compounds in the test set for other methods:

   python chemprop/predict.py --test_path <test_csv_file> --checkpoint_dir <chkpt_dir> --preds_path <pred_file>
   

Compound Prioritization using dmpna:

Please refer to the README.md in the comprank directory.