PCEvo: Path-Consistent Molecular Representation via Virtual Evolutionary

TL;DR

PCEvo tackles the data-scarce regime in molecular property prediction by moving beyond static endpoint supervision to path-consistent learning over virtual evolutionary edit paths. It constructs multiple chemically feasible transformation paths between similar molecules using a maximum common subgraph-based edit set and topological constraints, then encodes incremental changes with a differential evolutionary path encoder. Training combines a static endpoint loss with a path summation consistency loss, enforcing that per-edit contributions accumulate to match endpoint changes, thereby improving generalization under few-shot data. Theoretical analysis links the approach to tighter generalization bounds via an effective sample size $n_{eff}$ and reduced hypothesis complexity, and empirical results on QM9 and MoleculeNet show consistent, sometimes state-of-the-art, improvements across backbones and data regimes, validating the scalability and robustness of path-consistent molecular representations.

Abstract

Molecular representation learning aims to learn vector embeddings that capture molecular structure and geometry, thereby enabling property prediction and downstream scientific applications. In many AI for science tasks, labeled data are expensive to obtain and therefore limited in availability. Under the few-shot setting, models trained with scarce supervision often learn brittle structure-property relationships, resulting in substantially higher prediction errors and reduced generalization to unseen molecules. To address this limitation, we propose PCEvo, a path-consistent representation method that learns from virtual paths through dynamic structural evolution. PCEvo enumerates multiple chemically feasible edit paths between retrieved similar molecular pairs under topological dependency constraints. It transforms the labels of the two molecules into stepwise supervision along each virtual evolutionary path. It introduces a path-consistency objective that enforces prediction invariance across alternative paths connecting the same two molecules. Comprehensive experiments on the QM9 and MoleculeNet datasets demonstrate that PCEvo substantially improves the few-shot generalization performance of baseline methods. The code is available at https://anonymous.4open.science/r/PCEvo-4BF2.

Figures (4)

• Figure 1: Comparison of molecular representation paradigms: static end-to-end modeling, multi-stage learning, and virtual path modeling. P denotes the molecular property.
• Figure 2: Virtual evolutionary editing uses the HOMO energy from the QM9 dataset to illustrate two evolution paths between a pair of molecules. Despite exhibiting different intermediate fluctuations, both paths inevitably converge to the same final HOMO value, highlighting the path‐independent nature of the property.
• Figure 3: Overview of PCEvo method. Our method constructs virtual evolutionary paths between molecular states by identifying a sequence of valid graph edit operations. These discrete operations are then mapped into a continuous representation space to guide the learning of chemically interpretable molecular representations.
• Figure 4: Hyperparameter sensitivity analysis on QM9 with SchNet under the 100-shot setting. (a) fixes $N{=}10$ and varies $P_{\max}$, (b) fixes $P_{\max}{=}50$ and varies $N$.

Theorems & Definitions (1)

• Theorem 1: Edit decomposition yields a tighter generalization bound