ADME-DTI: Augmented Deep Meta Ensemble for Drug–Target Interaction Prediction

Overview

ADME-DTI is a deep learning ensemble for drug-target interaction prediction. It integrates multiple embeddings for drugs and proteins, along with physicochemical metadata, achieving competitive performance across benchmark datasets in both regression and classification tasks.

This repository extends our previous work on WAE-DTI by introducing the following key modifications:

• Use a metamodel rather than applying a mathematical weighted average ensemble of the submodels.
• Introduce drug/target metadata as auxiliary features.
• Add ESM Cambrian as an additional target descriptor.
• Remove MinHash fingerprint due to causing stability issues.

For WAE-DTI implementation, refer to this repository.

Getting Started

Clone the project from GitHub and install the necessary dependencies.

git clone https://github.com/tariqshaban/adme-dti
cd adme-dti
pip install -r requirements.txt

Pytorch needs to be installed (preferably utilizing GPU acceleration "CUDA-enabled").

The program will attempt to clone the ESM repository if it is not present automatically. Either install Git on your machine or manually download the repository and place it in ../esm.

Tip

You can also clone the ESM repository beforehand by running the following command:

git clone https://github.com/facebookresearch/esm ../esm

Project Structure

├── assets
│   └── images                        <- Store images for the README file.
├── config
│   └── config.py                     <- Store terminal arguments from entry files.
├── data
│   ├── embeddings
│   │   ├── drug_embedding            <- Pre-trained drug embeddings (Parquet format).
│   │   └── target_embedding          <- Pre-trained target embeddings (Parquet format).
│   └── raw
│       ├── classification            <- Raw classification datasets for training and evaluation (Parquet format).
│       └── regression                <- Raw regression datasets for training and evaluation (Parquet format).
├── results                           <- Store trained model, predictions, and metrics.
├── saved_models
│   ├── classification                <- Trained models and their performance on the classification task.
│   └── regression                    <- Trained models and their performance on the regression task.
├── src
│   ├── dti_dataset.py                <- Data preprocessing and mounting prior to training for the submodels.
│   ├── dti_dataset_combined.py       <- Data preprocessing and mounting prior to training for the metamodel.
│   ├── dti_model.py                  <- Neural network definition for the ADME-DTI submodels architecture.
│   ├── dti_model_combined.py         <- Neural network definition for the ADME-DTI metamodel architecture.
│   ├── evaluate.py                   <- Mathematical methods to evaluate model performance.
│   ├── metadata.py                   <- Extract metadata features from drugs (SMILES string)/targets (FASTA string)
│   ├── predict.py                    <- Provide predictions of a model given a dataloader.
│   ├── test.py                       <- Generate and evaluate predictions of the model given external examples.
│   └── train.py                      <- ADME-DTI model trainer.
├── utils
│   ├── embedding
│   │   ├── extractor
│   │   │   ├── drug_extractor.py     <- Extract and save embeddings from a list of drugs (Parquet format).
│   │   │   └── target_extractor.py   <- Extract and save embeddings from a list of targets (Parquet format).
│   │   └── loader
│   │       ├── drug_loader.py        <- Load drug embeddings from a Parquet file.
│   │       └── target_loader.py      <- Load target embeddings from a Parquet file.
│   ├── eda.py                        <- Perform exploratory data analysis (EDA) on all datasets.
│   ├── results_plotter.py            <- Generate residual plots (regression) and AUC/AUPRC plots (classification)
│   └── terminal_command_runner.py    <- Run external terminal commands in real time.
├── entry_inference.py                <- Entry point for inference.
├── entry_train.py                    <- Entry point for training.
├── train_all.py                      <- Perform all the experiments.
├── README.md                         <- README file and documentation.
└── requirements.txt                  <- List of dependencies required to run the project.

Usage

Training

entry_train.py [-h] [--use-wandb USE_WANDB] --task {classification,regression} [--dataset DATASET] [--seed SEED] [--epochs EPOCHS] [--patience PATIENCE]
               [--eda EDA] [--learning-rate LEARNING_RATE] [--batch-size BATCH_SIZE] [--torch-device TORCH_DEVICE]

Argument	Description	Default	Notes
-h, --help	Display a help message
--use-wandb USE_WANDB	Log training and validation metrics into W&B	False	If set to True, the code will initialize a project named by the selected dataset
--task	Supervised learning algorithm type		Must be set to either classification or regression
--dataset DATASET	Dataset used for training	'davis'	The dataset must be in the data/raw path with train and test folders containing Parquet files
--seed SEED	Ensures reproducibility (on the same device)
--epochs EPOCHS	Number of training epochs	3000
--patience PATIENCE	Number of epochs to elapse without improvement for the training to stop	200
--eda EDA	Conduct a quick EDA on startup	False	EDA is applied to all datasets regardless of the specified --dataset
--learning-rate LEARNING_RATE	Learning rate	5e-4
--batch-size BATCH_SIZE	Number of examples in each batch	1024
--torch-device TORCH_DEVICE	Device used for training (e.g. cuda:0, cpu)		If not specified, GPU will be utilized (if any)

Inference

entry_inference.py [-h] --models-path MODELS_PATH --input-file INPUT_FILE 
                   [--batch-size BATCH_SIZE] [--torch-device TORCH_DEVICE]

Argument	Description	Default	Notes
-h, --help	Display a help message
--models-path MODELS_PATH	Folder path which contains the models trained on each drug fingerprint		You can use the pretrained models of any dataset within models/saved_models
--input-file INPUT_FILE	CSV file path containing "drug", "target", and "label" columns (label is optional)		If "label" column is specified, you must have enough examples to satisfy the concordance index calculation requirement
--task	Supervised learning algorithm type		Must be set to either classification or regression
--batch-size BATCH_SIZE	Number of examples in each batch	1024
--torch-device TORCH_DEVICE	Device used for training (e.g. cuda:0, cpu)		If not specified, GPU will be utilized (if any)

Note

When running entry_train.py and entry_inference.py, missing embeddings from drugs and targets are automatically extracted and saved into data/embeddings

Important

During training, you may notice that tqdm reports monotonically increasing total value, this is caused by taking into account the dynamic fluctuation of the early stopping counter. So, rather than always having the total value equal to the number of epochs, the progress bar adjusts to display the smallest number of epochs needed to finish the training, which translates to the following formula:

total = min(epochs, elapsed_epochs + patience - elapsed_patience)

Results

The following tables are the result of training the model using nine drug descriptors and two target embeddings on six datasets while repeating the experiment five times. The values represent the mean and standard deviation of each metric.

Davis

Drug Descriptor	Protein Descriptor	Davis
		rm2	CI	MSE
Avalon fingerprint	ESM-2	0.730 ± 0.019	0.903 ± 0.003	0.203 ± 0.005
	ESM Cambrian	0.720 ± 0.017	0.902 ± 0.003	0.207 ± 0.004
Morgan fingerprint	ESM-2	0.725 ± 0.015	0.902 ± 0.004	0.207 ± 0.006
	ESM Cambrian	0.739 ± 0.011	0.905 ± 0.003	0.203 ± 0.005
Topological torsion fingerprint	ESM-2	0.733 ± 0.014	0.904 ± 0.005	0.201 ± 0.004
	ESM Cambrian	0.726 ± 0.011	0.902 ± 0.002	0.205 ± 0.005
MACCS keys fingerprint	ESM-2	0.730 ± 0.014	0.902 ± 0.004	0.202 ± 0.004
	ESM Cambrian	0.730 ± 0.012	0.903 ± 0.003	0.208 ± 0.006
LDP	ESM-2	0.672 ± 0.012	0.881 ± 0.003	0.252 ± 0.006
	ESM Cambrian	0.679 ± 0.005	0.880 ± 0.003	0.255 ± 0.004
RDKit fingerprint	ESM-2	0.736 ± 0.011	0.905 ± 0.002	0.201 ± 0.005
	ESM Cambrian	0.728 ± 0.018	0.902 ± 0.004	0.205 ± 0.007
Atom pair fingerprint	ESM-2	0.744 ± 0.010	0.904 ± 0.004	0.203 ± 0.006
	ESM Cambrian	0.739 ± 0.012	0.903 ± 0.004	0.202 ± 0.005
SEC fingerprint	ESM-2	0.728 ± 0.020	0.900 ± 0.005	0.205 ± 0.007
	ESM Cambrian	0.728 ± 0.008	0.901 ± 0.002	0.205 ± 0.006
Metamodel		0.759 ± 0.008	0.916 ± 0.002	0.186 ± 0.003

Kiba

Drug Descriptor	Protein Descriptor	Kiba
		rm2	CI	MSE
Avalon fingerprint	ESM-2	0.787 ± 0.003	0.889 ± 0.004	0.140 ± 0.001
	ESM Cambrian	0.784 ± 0.007	0.891 ± 0.004	0.140 ± 0.001
Morgan fingerprint	ESM-2	0.780 ± 0.012	0.892 ± 0.006	0.142 ± 0.003
	ESM Cambrian	0.779 ± 0.008	0.887 ± 0.004	0.143 ± 0.002
Topological torsion fingerprint	ESM-2	0.778 ± 0.005	0.893 ± 0.004	0.141 ± 0.002
	ESM Cambrian	0.780 ± 0.008	0.891 ± 0.004	0.141 ± 0.003
MACCS keys fingerprint	ESM-2	0.755 ± 0.003	0.871 ± 0.003	0.163 ± 0.002
	ESM Cambrian	0.745 ± 0.005	0.870 ± 0.005	0.168 ± 0.002
LDP	ESM-2	0.509 ± 0.017	0.797 ± 0.006	0.321 ± 0.008
	ESM Cambrian	0.490 ± 0.018	0.790 ± 0.005	0.333 ± 0.011
RDKit fingerprint	ESM-2	0.783 ± 0.005	0.893 ± 0.005	0.141 ± 0.001
	ESM Cambrian	0.784 ± 0.004	0.891 ± 0.003	0.141 ± 0.001
Atom pair fingerprint	ESM-2	0.780 ± 0.007	0.890 ± 0.004	0.144 ± 0.002
	ESM Cambrian	0.786 ± 0.005	0.890 ± 0.005	0.143 ± 0.001
SEC fingerprint	ESM-2	0.781 ± 0.006	0.891 ± 0.005	0.141 ± 0.003
	ESM Cambrian	0.781 ± 0.008	0.893 ± 0.004	0.141 ± 0.001
Metamodel		0.815 ± 0.007	0.902 ± 0.003	0.118 ± 0.002

DTC

Drug Descriptor	Protein Descriptor	DTC
		rm2	CI	MSE
Avalon fingerprint	ESM-2	0.842 ± 0.007	0.893 ± 0.003	0.158 ± 0.003
	ESM Cambrian	0.837 ± 0.005	0.892 ± 0.004	0.160 ± 0.002
Morgan fingerprint	ESM-2	0.836 ± 0.009	0.894 ± 0.002	0.158 ± 0.000
	ESM Cambrian	0.832 ± 0.010	0.893 ± 0.004	0.160 ± 0.001
Topological torsion fingerprint	ESM-2	0.837 ± 0.010	0.894 ± 0.002	0.158 ± 0.002
	ESM Cambrian	0.835 ± 0.007	0.893 ± 0.005	0.160 ± 0.002
MACCS keys fingerprint	ESM-2	0.806 ± 0.005	0.874 ± 0.004	0.189 ± 0.001
	ESM Cambrian	0.807 ± 0.005	0.874 ± 0.001	0.189 ± 0.001
LDP	ESM-2	0.656 ± 0.005	0.810 ± 0.004	0.336 ± 0.005
	ESM Cambrian	0.644 ± 0.017	0.808 ± 0.004	0.340 ± 0.008
RDKit fingerprint	ESM-2	0.837 ± 0.007	0.896 ± 0.003	0.157 ± 0.003
	ESM Cambrian	0.827 ± 0.003	0.894 ± 0.002	0.159 ± 0.002
Atom pair fingerprint	ESM-2	0.830 ± 0.007	0.893 ± 0.002	0.160 ± 0.002
	ESM Cambrian	0.835 ± 0.005	0.893 ± 0.002	0.161 ± 0.003
SEC fingerprint	ESM-2	0.836 ± 0.004	0.895 ± 0.001	0.161 ± 0.002
	ESM Cambrian	0.833 ± 0.012	0.893 ± 0.004	0.161 ± 0.003
Metamodel		0.862 ± 0.006	0.904 ± 0.002	0.134 ± 0.001

Metz

Drug Descriptor	Protein Descriptor	Metz
		rm2	CI	MSE
Avalon fingerprint	ESM-2	0.653 ± 0.008	0.809 ± 0.002	0.295 ± 0.004
	ESM Cambrian	0.654 ± 0.007	0.810 ± 0.001	0.295 ± 0.003
Morgan fingerprint	ESM-2	0.631 ± 0.012	0.804 ± 0.002	0.312 ± 0.004
	ESM Cambrian	0.632 ± 0.011	0.803 ± 0.002	0.313 ± 0.006
Topological torsion fingerprint	ESM-2	0.634 ± 0.006	0.804 ± 0.001	0.309 ± 0.002
	ESM Cambrian	0.645 ± 0.009	0.805 ± 0.001	0.307 ± 0.002
MACCS keys fingerprint	ESM-2	0.631 ± 0.007	0.803 ± 0.001	0.315 ± 0.002
	ESM Cambrian	0.637 ± 0.018	0.802 ± 0.002	0.316 ± 0.006
LDP	ESM-2	0.402 ± 0.006	0.731 ± 0.002	0.529 ± 0.006
	ESM Cambrian	0.400 ± 0.008	0.730 ± 0.002	0.532 ± 0.006
RDKit fingerprint	ESM-2	0.637 ± 0.008	0.806 ± 0.002	0.306 ± 0.005
	ESM Cambrian	0.625 ± 0.008	0.804 ± 0.001	0.311 ± 0.002
Atom pair fingerprint	ESM-2	0.620 ± 0.016	0.802 ± 0.002	0.318 ± 0.007
	ESM Cambrian	0.640 ± 0.009	0.805 ± 0.001	0.311 ± 0.004
SEC fingerprint	ESM-2	0.629 ± 0.022	0.802 ± 0.002	0.316 ± 0.006
	ESM Cambrian	0.635 ± 0.020	0.803 ± 0.003	0.315 ± 0.008
Metamodel		0.710 ± 0.004	0.824 ± 0.000	0.262 ± 0.002

ToxCast

Drug Descriptor	Protein Descriptor	ToxCast
		rm2	CI	MSE
Avalon fingerprint	ESM-2	0.568 ± 0.009	0.918 ± 0.001	0.312 ± 0.001
	ESM Cambrian	0.562 ± 0.006	0.916 ± 0.001	0.316 ± 0.002
Morgan fingerprint	ESM-2	0.556 ± 0.003	0.918 ± 0.002	0.315 ± 0.002
	ESM Cambrian	0.552 ± 0.008	0.916 ± 0.002	0.320 ± 0.003
Topological torsion fingerprint	ESM-2	0.566 ± 0.005	0.917 ± 0.001	0.314 ± 0.001
	ESM Cambrian	0.552 ± 0.008	0.917 ± 0.001	0.317 ± 0.002
MACCS keys fingerprint	ESM-2	0.564 ± 0.009	0.915 ± 0.001	0.318 ± 0.001
	ESM Cambrian	0.569 ± 0.008	0.915 ± 0.001	0.318 ± 0.002
LDP	ESM-2	0.517 ± 0.007	0.905 ± 0.001	0.354 ± 0.002
	ESM Cambrian	0.518 ± 0.006	0.903 ± 0.001	0.356 ± 0.002
RDKit fingerprint	ESM-2	0.560 ± 0.005	0.918 ± 0.001	0.315 ± 0.002
	ESM Cambrian	0.557 ± 0.006	0.918 ± 0.002	0.316 ± 0.003
Atom pair fingerprint	ESM-2	0.567 ± 0.008	0.919 ± 0.001	0.311 ± 0.002
	ESM Cambrian	0.564 ± 0.016	0.919 ± 0.001	0.314 ± 0.004
SEC fingerprint	ESM-2	0.559 ± 0.009	0.917 ± 0.001	0.314 ± 0.003
	ESM Cambrian	0.550 ± 0.006	0.916 ± 0.002	0.319 ± 0.002
Metamodel		0.584 ± 0.010	0.922 ± 0.003	0.300 ± 0.002

STITCH

Drug Descriptor	Protein Descriptor	STITCH
		rm2	CI	MSE
Avalon fingerprint	ESM-2	0.436 ± 0.003	0.769 ± 0.014	1.016 ± 0.017
	ESM Cambrian	0.434 ± 0.005	0.769 ± 0.010	1.013 ± 0.013
Morgan fingerprint	ESM-2	0.451 ± 0.007	0.786 ± 0.011	0.984 ± 0.009
	ESM Cambrian	0.458 ± 0.005	0.791 ± 0.006	0.965 ± 0.007
Topological torsion fingerprint	ESM-2	0.425 ± 0.005	0.768 ± 0.005	1.051 ± 0.007
	ESM Cambrian	0.432 ± 0.009	0.763 ± 0.005	1.045 ± 0.009
MACCS keys fingerprint	ESM-2	0.426 ± 0.004	0.749 ± 0.003	1.041 ± 0.004
	ESM Cambrian	0.424 ± 0.004	0.740 ± 0.003	1.046 ± 0.006
LDP	ESM-2	0.260 ± 0.005	0.669 ± 0.005	1.416 ± 0.013
	ESM Cambrian	0.265 ± 0.007	0.667 ± 0.006	1.403 ± 0.016
RDKit fingerprint	ESM-2	0.416 ± 0.006	0.775 ± 0.019	1.059 ± 0.007
	ESM Cambrian	0.428 ± 0.006	0.767 ± 0.013	1.044 ± 0.007
Atom pair fingerprint	ESM-2	0.420 ± 0.009	0.773 ± 0.019	1.058 ± 0.002
	ESM Cambrian	0.431 ± 0.008	0.769 ± 0.011	1.043 ± 0.006
SEC fingerprint	ESM-2	0.426 ± 0.009	0.780 ± 0.010	1.040 ± 0.007
	ESM Cambrian	0.437 ± 0.014	0.778 ± 0.010	1.027 ± 0.012
Metamodel		0.586 ± 0.003	0.799 ± 0.006	0.791 ± 0.005

Classification Task Results

In addition to the mentioned results for the regression task, three datasets were used to evaluate the model for the classification task.

BioSNAP

Drug Descriptor	Protein Descriptor	BioSNAP
		AUC	AUPRC	Sensitivity	Specificity	Threshold
Avalon fingerprint	ESM-2	0.940 ± 0.001	0.943 ± 0.002	0.870 ± 0.006	0.882 ± 0.008	0.503 ± 0.264
	ESM Cambrian	0.938 ± 0.001	0.940 ± 0.001	0.863 ± 0.003	0.888 ± 0.008	0.701 ± 0.201
Morgan fingerprint	ESM-2	0.926 ± 0.002	0.932 ± 0.002	0.845 ± 0.008	0.880 ± 0.008	0.814 ± 0.121
	ESM Cambrian	0.927 ± 0.002	0.936 ± 0.002	0.845 ± 0.008	0.876 ± 0.012	0.727 ± 0.129
Topological torsion fingerprint	ESM-2	0.929 ± 0.002	0.935 ± 0.003	0.863 ± 0.006	0.870 ± 0.009	0.539 ± 0.146
	ESM Cambrian	0.930 ± 0.001	0.937 ± 0.002	0.856 ± 0.006	0.873 ± 0.008	0.595 ± 0.166
MACCS keys fingerprint	ESM-2	0.936 ± 0.001	0.940 ± 0.002	0.870 ± 0.006	0.881 ± 0.007	0.789 ± 0.055
	ESM Cambrian	0.938 ± 0.001	0.941 ± 0.001	0.870 ± 0.009	0.881 ± 0.010	0.673 ± 0.153
LDP	ESM-2	0.863 ± 0.001	0.865 ± 0.003	0.774 ± 0.010	0.805 ± 0.013	0.547 ± 0.039
	ESM Cambrian	0.865 ± 0.002	0.871 ± 0.004	0.775 ± 0.007	0.804 ± 0.007	0.545 ± 0.062
RDKit fingerprint	ESM-2	0.930 ± 0.002	0.936 ± 0.003	0.854 ± 0.005	0.883 ± 0.005	0.701 ± 0.063
	ESM Cambrian	0.934 ± 0.001	0.939 ± 0.001	0.855 ± 0.005	0.881 ± 0.009	0.814 ± 0.057
Atom pair fingerprint	ESM-2	0.933 ± 0.003	0.937 ± 0.002	0.852 ± 0.005	0.886 ± 0.009	0.769 ± 0.121
	ESM Cambrian	0.932 ± 0.001	0.938 ± 0.002	0.858 ± 0.003	0.883 ± 0.006	0.687 ± 0.098
SEC fingerprint	ESM-2	0.925 ± 0.001	0.931 ± 0.003	0.853 ± 0.010	0.865 ± 0.009	0.524 ± 0.171
	ESM Cambrian	0.922 ± 0.004	0.930 ± 0.004	0.839 ± 0.006	0.873 ± 0.010	0.715 ± 0.187
Metamodel		0.955 ± 0.002	0.958 ± 0.006	0.897 ± 0.006	0.901 ± 0.008	0.289 ± 0.177

Davis

Drug Descriptor	Protein Descriptor	Davis
		AUC	AUPRC	Sensitivity	Specificity	Threshold
Avalon fingerprint	ESM-2	0.928 ± 0.004	0.438 ± 0.008	0.870 ± 0.013	0.859 ± 0.013	0.536 ± 0.034
	ESM Cambrian	0.929 ± 0.006	0.470 ± 0.024	0.879 ± 0.015	0.851 ± 0.009	0.594 ± 0.018
Morgan fingerprint	ESM-2	0.927 ± 0.003	0.456 ± 0.017	0.869 ± 0.011	0.868 ± 0.013	0.559 ± 0.023
	ESM Cambrian	0.929 ± 0.003	0.445 ± 0.008	0.871 ± 0.007	0.861 ± 0.011	0.612 ± 0.024
Topological torsion fingerprint	ESM-2	0.925 ± 0.002	0.448 ± 0.017	0.865 ± 0.011	0.859 ± 0.009	0.554 ± 0.058
	ESM Cambrian	0.927 ± 0.002	0.434 ± 0.009	0.864 ± 0.014	0.858 ± 0.022	0.569 ± 0.078
MACCS keys fingerprint	ESM-2	0.924 ± 0.004	0.438 ± 0.014	0.869 ± 0.016	0.860 ± 0.009	0.543 ± 0.075
	ESM Cambrian	0.930 ± 0.003	0.471 ± 0.015	0.867 ± 0.010	0.862 ± 0.017	0.604 ± 0.102
LDP	ESM-2	0.870 ± 0.015	0.331 ± 0.010	0.807 ± 0.034	0.793 ± 0.020	0.514 ± 0.085
	ESM Cambrian	0.878 ± 0.013	0.321 ± 0.006	0.820 ± 0.022	0.804 ± 0.020	0.444 ± 0.123
RDKit fingerprint	ESM-2	0.926 ± 0.002	0.467 ± 0.019	0.863 ± 0.013	0.862 ± 0.010	0.549 ± 0.053
	ESM Cambrian	0.931 ± 0.004	0.459 ± 0.019	0.871 ± 0.015	0.859 ± 0.011	0.585 ± 0.057
Atom pair fingerprint	ESM-2	0.927 ± 0.007	0.438 ± 0.016	0.865 ± 0.020	0.861 ± 0.013	0.642 ± 0.133
	ESM Cambrian	0.926 ± 0.001	0.460 ± 0.009	0.868 ± 0.009	0.854 ± 0.008	0.530 ± 0.080
SEC fingerprint	ESM-2	0.927 ± 0.005	0.460 ± 0.021	0.862 ± 0.009	0.866 ± 0.010	0.543 ± 0.054
	ESM Cambrian	0.930 ± 0.002	0.452 ± 0.008	0.867 ± 0.014	0.865 ± 0.008	0.543 ± 0.072
Metamodel		0.935 ± 0.011	0.510 ± 0.015	0.877 ± 0.005	0.886 ± 0.004	0.495 ± 0.041

BindingDB

Drug Descriptor	Protein Descriptor	BindingDB
		AUC	AUPRC	Sensitivity	Specificity	Threshold
Avalon fingerprint	ESM-2	0.924 ± 0.002	0.654 ± 0.007	0.882 ± 0.006	0.834 ± 0.009	0.433 ± 0.018
	ESM Cambrian	0.923 ± 0.001	0.660 ± 0.009	0.878 ± 0.011	0.835 ± 0.010	0.495 ± 0.056
Morgan fingerprint	ESM-2	0.924 ± 0.002	0.658 ± 0.006	0.890 ± 0.007	0.836 ± 0.006	0.514 ± 0.016
	ESM Cambrian	0.923 ± 0.002	0.661 ± 0.010	0.882 ± 0.004	0.843 ± 0.006	0.470 ± 0.047
Topological torsion fingerprint	ESM-2	0.922 ± 0.005	0.655 ± 0.005	0.871 ± 0.011	0.838 ± 0.011	0.523 ± 0.023
	ESM Cambrian	0.924 ± 0.001	0.668 ± 0.003	0.886 ± 0.010	0.833 ± 0.005	0.482 ± 0.025
MACCS keys fingerprint	ESM-2	0.923 ± 0.002	0.648 ± 0.004	0.883 ± 0.010	0.839 ± 0.006	0.454 ± 0.102
	ESM Cambrian	0.924 ± 0.001	0.657 ± 0.009	0.887 ± 0.007	0.833 ± 0.006	0.537 ± 0.054
LDP	ESM-2	0.896 ± 0.006	0.558 ± 0.015	0.862 ± 0.006	0.808 ± 0.015	0.374 ± 0.108
	ESM Cambrian	0.893 ± 0.004	0.568 ± 0.008	0.848 ± 0.013	0.803 ± 0.010	0.488 ± 0.037
RDKit fingerprint	ESM-2	0.924 ± 0.002	0.661 ± 0.009	0.885 ± 0.005	0.835 ± 0.004	0.508 ± 0.043
	ESM Cambrian	0.924 ± 0.001	0.667 ± 0.008	0.883 ± 0.006	0.838 ± 0.005	0.491 ± 0.032
Atom pair fingerprint	ESM-2	0.923 ± 0.002	0.659 ± 0.004	0.884 ± 0.011	0.836 ± 0.007	0.450 ± 0.065
	ESM Cambrian	0.922 ± 0.001	0.661 ± 0.008	0.876 ± 0.010	0.835 ± 0.008	0.497 ± 0.028
SEC fingerprint	ESM-2	0.924 ± 0.002	0.651 ± 0.009	0.888 ± 0.011	0.834 ± 0.004	0.506 ± 0.019
	ESM Cambrian	0.921 ± 0.002	0.654 ± 0.004	0.868 ± 0.012	0.838 ± 0.006	0.508 ± 0.036
Metamodel		0.936 ± 0.001	0.716 ± 0.007	0.891 ± 0.005	0.854 ± 0.005	0.410 ± 0.092

Citation

Tariq Sha’ban, Ahmad M. Mustafa, Mostafa Z. Ali, Talal Z. Ali, ADME-DTI: Augmented Deep Meta Ensemble for Drug–Target Interaction Prediction, Molecular Informatics, Volume 45, 2026, e70033.

@article{https://doi.org/10.1002/minf.70033,
    title = {ADME-DTI: Augmented Deep Meta Ensemble for Drug–Target Interaction Prediction},
    journal = {Molecular Informatics},
    volume = {45},
    number = {5},
    pages = {e70033},
    year = {2026},
    doi = {https://doi.org/10.1002/minf.70033},
    url = {https://onlinelibrary.wiley.com/doi/abs/10.1002/minf.70033},
    author = {Sha’ban, Tariq and Mustafa, Ahmad M. and Ali, Mostafa Z. and Ali, Talal Z.}
}

---