DDINet: Drug-drug interaction prediction network based on multi-molecular fingerprint features and multi-head attention centered weighted autoencoder.

Abstract

Drug-drug interactions (DDIs) pose a major concern in polypharmacy due to their potential to cause unexpected side effects that can adversely affect a patient's health. Therefore, it is crucial to identify DDIs effectively during the early stages of drug discovery and development. In this paper, a novel DDI prediction network (DDINet) is proposed to enhance the predictive performance over conventional DDI methods. Leveraging the DrugBank dataset, drugs are represented using the Simplified Molecular Input Line-Entry System (SMILES), with the RDKit software pre-processing the SMILES strings into their canonical forms. Multiple molecular fingerprinting techniques such as Extended Connectivity Fingerprints (ECFPs), Molecular ACCess System keys (MACCSkeys), PubChem Fingerprints, 3D molecular fingerprints (3D-FP), and molecular dynamics fingerprints (MDFPs) are employed to encode drug chemical structures into feature vectors. Drug similarities are computed using the Tanimoto coefficient (TC), and the final Structural Similarity Profile (SSP) is obtained by averaging the five molecular fingerprint types. The novelty of the approach lies in the integration of a Multi-head Attention centered Weighted Autoencoder (Mul_WAE) as the interaction prediction module, which leverages the Multi-head Attention (MHA) layer to focus on the most significant input features. Furthermore, we introduce the Upgraded Bald Eagle Search Optimization (UBesO) algorithm, which optimally selects the learnable parameters of the Mul_WAE based on cross-entropy loss, improving the model's convergence and performance. The proposed DDINet model achieves an accuracy of 99.77%, 99.66% of AUC, 99.5% average precision, 99.4% precision, and 99.49% recall, providing a comprehensive evaluation of the model's robustness. Beyond high accuracy, DDINet offers advantages in scalability, making it well suited for handling large datasets due to its efficient feature extraction and optimization processes. The unique combination of multiple molecular fingerprinting methods with the MHA layer and UBesO algorithm highlights the innovative aspects of our model and significantly improves prediction performance compared to existing approaches.