Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Copy-Augmented Representation for Structure Invariant Template-Free Retrosynthesis

Jiaxi Zhuang, Yu Zhang, Aimin Zhou, Ying Qian

TL;DR

This work tackles template-free retrosynthesis by explicitly modeling structural invariance, a property where large molecular scaffolds persist through reactions. It introduces C-SMILES, a decomposed SMILES representation that reduces editing distance between reactants and products, and augments generation with a copy mechanism governed by attention over product tokens. An atom-mapping based SMILES alignment guidance further constrains the model to chemically coherent transformations. On USPTO-50K and USPTO-FULL, the method delivers state-of-the-art performance among template-free approaches, with high validity and strong forward-planning plausibility, demonstrating the practical potential for structure-aware computational retrosynthesis.

Abstract

Retrosynthesis prediction is fundamental to drug discovery and chemical synthesis, requiring the identification of reactants that can produce a target molecule. Current template-free methods struggle to capture the structural invariance inherent in chemical reactions, where substantial molecular scaffolds remain unchanged, leading to unnecessarily large search spaces and reduced prediction accuracy. We introduce C-SMILES, a novel molecular representation that decomposes traditional SMILES into element-token pairs with five special tokens, effectively minimizing editing distance between reactants and products. Building upon this representation, we incorporate a copy-augmented mechanism that dynamically determines whether to generate new tokens or preserve unchanged molecular fragments from the product. Our approach integrates SMILES alignment guidance to enhance attention consistency with ground-truth atom mappings, enabling more chemically coherent predictions. Comprehensive evaluation on USPTO-50K and large-scale USPTO-FULL datasets demonstrates significant improvements: 67.2% top-1 accuracy on USPTO-50K and 50.8% on USPTO-FULL, with 99.9% validity in generated molecules. This work establishes a new paradigm for structure-aware molecular generation with direct applications in computational drug discovery.

Copy-Augmented Representation for Structure Invariant Template-Free Retrosynthesis

TL;DR

This work tackles template-free retrosynthesis by explicitly modeling structural invariance, a property where large molecular scaffolds persist through reactions. It introduces C-SMILES, a decomposed SMILES representation that reduces editing distance between reactants and products, and augments generation with a copy mechanism governed by attention over product tokens. An atom-mapping based SMILES alignment guidance further constrains the model to chemically coherent transformations. On USPTO-50K and USPTO-FULL, the method delivers state-of-the-art performance among template-free approaches, with high validity and strong forward-planning plausibility, demonstrating the practical potential for structure-aware computational retrosynthesis.

Abstract

Retrosynthesis prediction is fundamental to drug discovery and chemical synthesis, requiring the identification of reactants that can produce a target molecule. Current template-free methods struggle to capture the structural invariance inherent in chemical reactions, where substantial molecular scaffolds remain unchanged, leading to unnecessarily large search spaces and reduced prediction accuracy. We introduce C-SMILES, a novel molecular representation that decomposes traditional SMILES into element-token pairs with five special tokens, effectively minimizing editing distance between reactants and products. Building upon this representation, we incorporate a copy-augmented mechanism that dynamically determines whether to generate new tokens or preserve unchanged molecular fragments from the product. Our approach integrates SMILES alignment guidance to enhance attention consistency with ground-truth atom mappings, enabling more chemically coherent predictions. Comprehensive evaluation on USPTO-50K and large-scale USPTO-FULL datasets demonstrates significant improvements: 67.2% top-1 accuracy on USPTO-50K and 50.8% on USPTO-FULL, with 99.9% validity in generated molecules. This work establishes a new paradigm for structure-aware molecular generation with direct applications in computational drug discovery.
Paper Structure (22 sections, 8 equations, 4 figures, 6 tables, 2 algorithms)

This paper contains 22 sections, 8 equations, 4 figures, 6 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Method
  3. Overview
  4. C-SMILES Representation
  5. Copy-Augmented Generation
  6. Training Strategy
  7. Experiments
  8. Experimental Setup
  9. Datasets
  10. Evaluation Metrics
  11. Baselines
  12. Implementation Details
  13. Performance On USPTO-50K
  14. Top-k Accuracy
  15. Top-k Validity
  16. ...and 7 more sections

Figures (4)

  • Figure 1: Example of Structural Invariance in a Chemical Reaction. Green highlighting indicates molecular scaffolds that remain unchanged during the reaction, while modified regions represent reactive sites.
  • Figure 2: Overview of C-SMILES and Copy Mechanism based on Transformer.
  • Figure 3: SMILES Alignment (left) and Cross-Attention score (right).
  • Figure 4: Retrosynthesis Prediction under Copy Mechanism. Highlight in green when atoms/bonds are directly copied from Products.