Drug-drug interaction prediction with learnable size-adaptive molecular substructures

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence

Background:

• Drug-drug interactions (DDIs) pose significant health risks due to adverse effects from combining incompatible medications.
• Drug chemical composition is a primary cause of these interactions.
• Accurate prediction of DDIs is crucial for patient safety.

Purpose of the Study:

• To introduce a novel deep learning model, the gated message passing neural network (GMPNN), for predicting drug-drug interactions.
• To enable the model to learn relevant chemical substructures from drug molecular graphs for improved DDI prediction.
• To evaluate the efficacy of the GMPNN-CS method on established DDI datasets.

Main Methods:

• Utilized a gated message passing neural network (GMPNN) architecture to process molecular graph representations of drugs.
• Employed edges as learnable gates to control message passing and delimit chemical substructures of varying sizes and shapes.
• Developed a prediction module (GMPNN-CS) that calculates DDI likelihood based on learned substructure interactions and their relevance scores.

Main Results:

• The GMPNN-CS model was evaluated on two real-world DDI datasets.
• Achieved competitive results on one dataset.
• Demonstrated improved performance compared to existing methods on the other dataset.

Conclusions:

• The proposed GMPNN-CS method effectively predicts drug-drug interactions by learning salient chemical substructures.
• The gating mechanism in GMPNN allows for flexible and learnable substructure identification.
• GMPNN-CS offers a promising advancement in computational approaches for DDI prediction, enhancing drug safety.