Abstract

Subgraph-based GNNs can improve molecular representation learning, but existing approaches often lack a unified way to incorporate motifs across architectures. This paper introduces Uni-Motif, a universal motif integration framework that treats each motif as both a chemical identifier and a structural context carrier. The method is built to be compatible with different GNN backbones rather than tied to a single message-passing design.

Key ideas

• Decouple each motif into functional encoding and learnable structural encoding.
• Use functional encoding as a unique motif identifier and structural encoding as local graph context for nodes inside the motif.
• Support both message-passing workflows and global-attention-style workflows within the same motif-aware framework.
• Provide theoretical analysis showing stronger expressiveness than standard local aggregation alone.

Empirical findings

• The paper evaluates seven motif extraction techniques spanning topology-driven, ring-driven, and rule-driven categories.
• Experiments reported in the paper show that ring-driven motifs have a particularly strong effect on downstream performance.
• Across benchmark molecular graph datasets, Uni-Motif achieves competitive or superior results while remaining compatible with different GNN architectures.

Why it matters

Molecular motifs are not only structural patterns but also chemically meaningful units. By separating their functional role from their structural role, Uni-Motif gives GNNs a more expressive representation of subtle molecular differences without rewriting the underlying architecture.

Code

• GraphMoLab/Uni-Motif

Recommended citation: Runze Wang, Yuting Ma, *Xingyue Liu*, Zijie Xing, and Yanming Shen, “Motif-driven molecular graph representation learning,” Expert Systems with Applications, vol. 269, p. 126484, 2025
Download Paper | Download Bibtex