CLRe: A Synergistic Dual‐Engine Framework for One‐Step Retrosynthesis Prediction
TL;DR
A label‐free contrastive metric that quantifies intrinsic molecular complexity to optimize training pacing and adapt label smoothing to act as a topological buffer, which preserves the search entropy required for complex multi‐path chemical reasoning.
Abstract
ABSTRACT One‐step retrosynthesis prediction is fundamentally limited by the random training order of sequence‐to‐sequence models and the inherent mismatch between local text generation and global chemical topology. Here we present CLRe (Contrastive curriculum Learning for Retrosynthesis), a framework that integrates self‐supervised curriculum learning with topological buffering to resolve these bottlenecks. We introduce a label‐free contrastive metric that quantifies intrinsic molecular complexity to optimize training pacing. Furthermore, we adapt label smoothing to act as a topological buffer, which preserves the search entropy required for complex multi‐path chemical reasoning. We demonstrate that CLRe consistently improves performance on the USPTO‐50K and USPTO‐MIT datasets, significantly reducing accuracy disparities across historically challenging reaction classes. By capturing fine‐grained structural complexity orthogonal to standard reaction rules, CLRe offers a robust strategy for bridging data‐driven sequence generation with intrinsic chemical intuition.