How ImageDDI Boosts Drug‑Drug Interaction Prediction with Motif Sequences and Molecular Images

Background

Accurate prediction of drug‑drug interactions (DDIs) is critical for drug safety and new‑drug development. Existing methods rely on whole‑molecule or substructure representations, which cannot fully capture functional motif interactions that drive DDIs.

ImageDDI Overview

ImageDDI combines molecular motif sequences with 2‑D/3‑D molecular images to obtain richer drug representations. The framework consists of three components:

Motif vocabulary construction : Molecules are fragmented into functional motifs using the BRICS algorithm; a motif vocabulary is built to enable fine‑grained representation.

Image‑enhanced motif sequence modeling : Motif sequences of a drug pair are concatenated and fed into a Transformer encoder. Adaptive Feature Fusion injects visual features into the attention mechanism, allowing simultaneous modeling of local motif information and global image cues.

Molecular image feature extraction : 2‑D topological images (generated with RDKit) and multi‑view 3‑D renderings are encoded by a ResNet‑18 backbone. Features from the two drugs are concatenated to form a joint visual representation.

Method Details

Given a drug pair (dx, dy), their motif sequences Sdx and Sdy are extracted and concatenated into S(dx,dy). The Transformer processes this sequence, and the attention scores are modified by a learnable bias ΦIxy derived from the visual features Ixy:

Attention(Q,K,V) = softmax((QKᵀ / √d) + ΦIxy) V

The attention output passes through a feed‑forward network with residual connections and layer normalization, producing the final drug‑pair embedding used for DDI classification.

Experimental Setup

ImageDDI was evaluated on three public DDI datasets: Deng’s dataset, Ryu’s dataset, and DrugBank. Both conventional random‑split and inductive cold‑start scenarios were tested. Baselines included graph‑based models (e.g., MRCGNN) and image‑based models (e.g., ImageMol).

Results

In conventional settings, ImageDDI achieved the highest Accuracy and Macro‑F1, improving Macro‑F1 by >10 % on Deng’s dataset and >6 % on Ryu’s dataset compared with the best baselines. In inductive experiments, ImageDDI outperformed competitors by 9.4 % (Macro‑F1) and 7.4 % (Accuracy) in the new‑drug‑vs‑known‑drug (S1) setting and also led in the new‑drug‑vs‑new‑drug (S2) setting.

Hyperparameter Sensitivity

Learning rate, weight decay, and motif sequence length significantly affect performance, confirming the robustness of the approach.

Visual Interpretability

Grad‑CAM heatmaps show that ImageDDI consistently attends to chemically relevant motifs in both 2‑D and 3‑D visualizations, even for unseen drugs, providing transparent explanations.

References

Paper: ImageDDI: Image‑enhanced Molecular Motif Sequence Representation for Drug‑Drug Interaction Prediction (ScienceDirect) – https://www.sciencedirect.com/science/article/abs/pii/S1566253525006463

Code repository: https://github.com/1hyq/ImageDDI