Multimodal Deep Neural Network for Predicting Drug‑Drug Interactions

Today’s talk shares recent progress of a multimodal DNN model for drug‑drug interaction (DDI) prediction, originally published at IJCAI 2021.

Background : Drug discovery is costly and time‑consuming; AI‑driven DDI prediction can accelerate candidate selection and reduce toxicity.

DDI is represented as a matrix linking 57 2 drugs to 65 interaction outcomes, forming a Drug Knowledge Graph (DKG) with 57 2 nodes and 157 relation types.

The model incorporates heterogeneous features (HF) – targets, enzymes, and molecular structures – to capture drug‑specific information.

MDNN Architecture :

DKG module: builds a graph of drugs and their components, encoding semantic and topological information.

HF module: vectorizes target, enzyme, and substructure data, computing similarity matrices via Jaccard similarity.

Multimodal fusion layer: concatenates DKG embeddings (E) and HF embeddings (E') and feeds them to a softmax output layer.

The fusion improves performance over using DKG or HF alone.

Results : MDNN achieves state‑of‑the‑art scores on Acc, AUC, F1, AUPR, Precision, and Recall compared with baseline methods (excluding GNNs). Fusion of DKG and HF yields the best results.

Parameter sensitivity analysis shows how layer depth (l) and node count (N) affect evaluation metrics.

Multi‑task experiments indicate a performance drop when training and test drug sets are strictly separated.

Conclusion : MDNN effectively combines drug knowledge graphs and heterogeneous drug features, achieving superior DDI prediction performance, though further optimization is needed for real‑world deployment.

Q&A highlighted that the DrugBank dataset is publicly available and that similar multimodal‑graph techniques can be applied to e‑commerce and other domains.