Utilizing AI-driven approaches for drug–target interaction (DTI) prediction require large volumes of training data which are not available for the majority of target proteins. In this study, we investigate the use of deep transfer learning for the prediction of interactions between drug candidate compounds and understudied target proteins with scarce training data. The idea here is to first train a deep neural network classifier with a generalized source training dataset of large size and then to reuse this pre-trained neural network as an initial configuration for re-training/fine-tuning purposes with a small-sized specialized target training dataset. To explore this idea, we selected six protein families that have critical importance in biomedicine: kinases, G-protein-coupled receptors (GPCRs), ion channels, nuclear receptors, proteases, and transporters. In two independent experiments, the protein families of transporters and nuclear receptors were individually set as the target datasets, while the remaining five families were used as the source datasets. Several size-based target family training datasets were formed in a controlled manner to assess the benefit provided by the transfer learning approach.

Here, we present a systematic evaluation of our approach by pre-training a feed-forward neural network with source training datasets and applying different modes of transfer learning from the pre-trained source network to a target dataset. The performance of deep transfer learning is evaluated and compared with that of training the same deep neural network from scratch. We found that when the training dataset contains fewer than 100 compounds, transfer learning outperforms the conventional strategy of training the system from scratch, suggesting that transfer learning is advantageous for predicting binders to under-studied targets.

The source code and datasets are available in this repository. Our web-based service containing the ready-to-use pre-trained models is accessible at https://tl4dti.kansil.org.

Figure. Overview of the model training phase. We first trained a source neural network model using the bioactivity dataset of the selected source protein family (Stage I). This pre-trained source model is then used in the context of transfer learning for further trainin (i.e., fine tuning) with a small-sized target bioactivity dataset that belong to another protein family (Stage II). We also trained an FNN having exactly the same configuration from scratch (the reference model), as well as a shallow classifier (base model), using this same target training dataset.

• bin folder includes the source code of TransferLearning4DTI.

• training_files folder contains various traininig/test datasets mostly formatted for observational purposes and for employment in future studies

  • gpcr contains training data points, features, data splits for GPCR dataset. compound_feature_vectors/ecfp4.tsv includes ecfp4 features of ligands. compound_feature_vectors/chemprop.tsv includes chemprop features of ligands. comp_targ_binary.tsv is a csv file where each line is formatted as <compound_id>,<target_id>,<active/inactive>.
  • ionchannel contains training data points, features, data splits for Ion Channel dataset. compound_feature_vectors/ecfp4.tsv includes ecfp4 features of ligands. compound_feature_vectors/chemprop.tsv includes chemprop features of ligands. comp_targ_binary.tsv is a csv file where each line is formatted as <compound_id>,<target_id>,<active/inactive>.
  • kinase contains training data points, features, data splits for Kinase dataset. compound_feature_vectors/ecfp4.tsv includes ecfp4 features of ligands. compound_feature_vectors/chemprop.tsv includes chemprop features of ligands. comp_targ_binary.tsv is a csv file where each line is formatted as <compound_id>,<target_id>,<active/inactive>.
  • nuclearreceptor contains training data points, features, data splits for Nuclear Receptor dataset. compound_feature_vectors/ecfp4.tsv includes ecfp4 features of ligands. compound_feature_vectors/chemprop.tsv includes chemprop features of ligands. comp_targ_binary.tsv is a csv file where each line is formatted as <compound_id>,<target_id>,<active/inactive>.
  • protease contains training data points, features, data splits for Protease dataset. compound_feature_vectors/ecfp4.tsv includes ecfp4 features of ligands. compound_feature_vectors/chemprop.tsv includes chemprop features of ligands. comp_targ_binary.tsv is a csv file where each line is formatted as <compound_id>,<target_id>,<active/inactive>.
  • transporter contains training data points, features, data splits for Kinase dataset based on Setting-2. compound_feature_vectors/ecfp4.tsv includes ecfp4 features of ligands. compound_feature_vectors/chemprop.tsv includes chemprop features of ligands. comp_targ_binary.tsv is a csv file where each line is formatted as <compound_id>,<target_id>,<active/inactive>.

• Download the remaining training files here

• Clone the Git Repository

• Run the below commands for each dataset

--chln: number of neurons in compound hidden layers (default: 1200_300)

--lr:learning rate (default: 0.0001)

--bs: batch size (default: 256)

--td: the name of the target dataset (default: transporter)

--sd: the name of the source dataset (default: kinase)

--do: dropout rate (default: 0.1)

--en: the name of the experiment (default: my_experiment)

--model: model name (default: fc_2_layer)

--epoch: number of epochs (default: 100)

--sf: subset flag (default: 0)

--tlf: transfer learning flag (default: 0)

--ff: freeze flag (default: 0)

--fl: hidden layer to be frozen (default: 1)

--el: layer to be extracted (default: 0)

--ss: subset size (default: 10)

--cf: compound features separated by underscore character (default: chemprop)

--setting: Determines the setting (1: train_val_test, 2:extract layer train_val_test, 3:training_test, 4:only training, 5:extract layer train and test) (default: 1)

--et: external test dataset (default: -)

--nc: number of result classes (default: 2)

--train: (1) train or (0) extract features(default: 1)

python create_small_dataset.py --d transporter --ss 6

python baseline_training.py --setting 2 --tlf 0 --td transporter --ss 6 --en 0 --sf 1

python main_training.py --setting 3 --epoch 50 --ss 6 --en 0 --tlf 0 --sf 1 --td transporter

python main_training.py --setting 5 --train 0 --epoch 50 --ss 6 --en 0 --el 1 --tlf 1 --sf 1 --td transporter --sd kinase

python baseline_training.py --setting 2 --tlf 1 --el 1 --td transporter --sd kinase --ss 6 --en 0 --sf 1

python main_training.py --setting 3 --epoch 50 --ss 6 --en 0 --tlf 1 --sf 1 --td transporter --sd kinase

python main_training.py --setting 3 --epoch 50 --ss 6 --en 0 --ff 1 --fl 1 --tlf 1 --ff 1 --sf 1 --td transporter --sd kinase

--trainisc: the name of the training file. it should contains id, smiles and compound columns (default: input/train.csv)

--name: the name of the your protein or protein family (default: new_family)

--sc: the name of the source checkpoint (default: kinase)

python convert_chemprop_ftune.py --trainisc input/train.csv --name your_family_name --sc kinase

python main_training.py --setting 4 --td your_family_name 

python main_training.py --setting 4 --td your_family_name --sd kinase --tlf 1

--testisf: the name of the test file. it should contains id and smiles columns (default: input/test.csv)

--sc: the name of the source checkpoint (default: kinase)

python convert_chemprop_predict.py --testisf input/test.csv --sc kinase

python main_training.py --setting 6 --sd kinase --et output/test_chemprop.tsv --tlf 1

python main_training.py --setting 4 --td transporter --sd kinase --et output/test_chemprop.tsv --tlf 1

--trainisc: the name of the training file. it should contains id, smiles and compound columns (default: input/train.csv)

--name: the name of the your protein or protein family (default: new_family)

--testisf: the name of the test file. it should contains id and smiles columns (default: input/test.csv)

--sc: the name of the source checkpoint (default: kinase)

python convert_chemprop_ftune_predict.py --trainisc input/train.csv --name your_family_name --sc kinase --testisf input/test.csv

python main_training.py --setting 4 --td your_family_name --sd kinase --et output/test_chemprop.tsv --tlf 1

python main_training.py --setting 4 --td your_family_name --sd kinase --et output/test_chemprop.tsv --tlf 1 --ff 1 --fl 1

main_training.py creates a folder under named experiment_name (given as argument --en) under result_files folder. One file is created under results_files/<experiment_name>: performance_results.txt which contains the best performance results for test dataset. Sample output files for Transporter dataset is given under results_files/transporter.

Datasets for AKT1

Candidate drugs for AKT1.

Datasets for CDK1

Candidate drugs for CDK1.

Datasets for CDK2

Candidate drugs for CDK2.

Datasets for LOX

Candidate drugs for LOX.