Robust Prediction of Drug Interactions using Chemical Descriptors

Description:

Dataset Details: This dataset (DDI2025) was curated in 2025 from DrugBank v5.1.12 (released 2024-04-13). It covers 868,069 drug pairs across 2,957 unique drugs, annotated with 178 interaction types spanning pharmacokinetic (PK) and pharmacodynamic (PD) mechanisms. The dataset was constructed by extracting canonical SMILES strings from DrugBank and computing 3,780 QSAR descriptors per drug pair using PyBioMed (PyInteraction module) and RDKit (version 2017.09.3) under Python 2.7.18. The distribution is highly imbalanced: the most frequent interaction type accounts for approximately 13% of all pairs, while over 121 types appear in fewer than 0.1% of samples.

Splits: Training (60%, 520,841 pairs), validation (20%, 173,614 pairs), and test (20%, 173,614 pairs), stratified randomly by class.

File Content: Each CSV file contains the following columns:

• drugid_drug_a / drugid_drug_b: DrugBank IDs of the two drugs
• drugname_drug_a / drugname_drug_b: Generic drug names
• drugsmiles_drug_a / drugsmiles_drug_b: Canonical SMILES strings
• 3,780 QSAR descriptor columns spanning seven base families and their pairwise cross-family interaction terms:
  • MR_VSA: Molar Refractivity van der Waals Surface Area (steric bulk)
  • EState_VSA: E-state van der Waals Surface Area (electronic topology)
  • SlogP_VSA: Octanol/water partition coefficient-based VSA (lipophilicity)
  • LabuteASA: Labute Atomic Surface Area (solvent-accessible surface)
  • MTPSA: Molecular Topological Polar Surface Area (polarity, membrane permeability)
  • PEOE_VSA: Partial Equalization of Orbital Electronegativity VSA (ionization, partial charges)
  • VSA_EState: van der Waals Surface Area-based E-state (electronic-spatial combined)
  • Cross-family interaction terms (pairwise products across families, e.g., SlogP_VSA*PEOE_VSA), which account for the majority of the 3,780 features
• class: Integer 0–177 encoding the DDI type

Usage: This dataset is the primary input for training and evaluating the T-DDI model. For source code and implementation details, see: https://github.com/HienKha/tddi. A live web application is available at: https://projectxddi-tddi-docker.hf.space/