[Paper Review] Retrosynthesis prediction enhanced by in-silico reaction data augmentation

Xu Zhang, Yiming Mo+2 more|arXiv (Cornell University)|Jan 31, 2024

Genomics and Phylogenetic Studies4 citations

TL;DR

This paper introduces RetroWISE, a self-boosting framework that enhances retrosynthesis prediction by generating in-silico reaction data from unpaired molecular data using a base model trained on real paired data. By iteratively generating high-quality synthetic reactions and fine-tuning the model, RetroWISE achieves state-of-the-art performance, improving top-1 accuracy by 8.6% on the USPTO-50K dataset and consistently boosting rare transformation prediction.

ABSTRACT

Recent advances in machine learning (ML) have expedited retrosynthesis research by assisting chemists to design experiments more efficiently. However, all ML-based methods consume substantial amounts of paired training data (i.e., chemical reaction: product-reactant(s) pair), which is costly to obtain. Moreover, companies view reaction data as a valuable asset and restrict the accessibility to researchers. These issues prevent the creation of more powerful retrosynthesis models due to their data-driven nature. As a response, we exploit easy-to-access unpaired data (i.e., one component of product-reactant(s) pair) for generating in-silico paired data to facilitate model training. Specifically, we present RetroWISE, a self-boosting framework that employs a base model inferred from real paired data to perform in-silico reaction generation and augmentation using unpaired data, ultimately leading to a superior model. On three benchmark datasets, RetroWISE achieves the best overall performance against state-of-the-art models (e.g., +8.6% top-1 accuracy on the USPTO-50K test dataset). Moreover, it consistently improves the prediction accuracy of rare transformations. These results show that Retro- WISE overcomes the training bottleneck by in-silico reactions, thereby paving the way toward more effective ML-based retrosynthesis models.

Motivation & Objective

• To address the scarcity and inaccessibility of high-quality paired reaction data (product-reactant pairs) in retrosynthesis modeling.
• To overcome data-driven limitations in machine learning-based retrosynthesis by generating synthetic, in-silico reaction data from easily accessible unpaired data (e.g., isolated products or reactants).
• To develop a self-boosting framework that iteratively improves model performance by using generated reactions to retrain and refine the model.
• To enhance prediction accuracy, especially for rare or low-frequency transformations, which are underrepresented in existing datasets.
• To improve computational efficiency and scalability for real-world deployment in multistep retrosynthesis planning.

Proposed method

• A base model is first trained on real paired reaction data (e.g., from USPTO) to predict reactants from products.
• The trained base model generates in-silico reaction pairs from unpaired data (e.g., isolated products or reactants) collected from public sources or web scraping.
• The generated in-silico reactions are used to augment the original training data, enabling retraining of a more robust and generalizable retrosynthesis model.
• The framework employs iterative training: each new model generates higher-quality synthetic reactions, which are used to further improve the model in a self-boosting loop.
• The method uses a transformer-based architecture with ~44.5 million parameters, optimized for computational and memory efficiency during inference.
• SMILES augmentation via Root-aligned SMILES (R-SMILES) is applied to improve alignment between product and reactant representations, reducing the one-to-many ambiguity in sequence generation.

Experimental results

Research questions

• RQ1 Can in-silico reaction data augmentation significantly improve the performance of machine learning-based retrosynthesis models when real paired data is limited?
• RQ2 Can a self-boosting framework that iteratively generates and re-trains on synthetic reactions outperform existing state-of-the-art models?
• RQ3 Does the proposed method improve prediction accuracy for rare or low-frequency transformations that are underrepresented in training data?
• RQ4 How does the use of unpaired data (e.g., isolated products or reactants) compare to traditional paired data in terms of model generalization and efficiency?
• RQ5 To what extent does iterative refinement of the model through synthetic data generation enhance prediction quality and robustness?

Key findings

• RetroWISE achieves a top-1 accuracy of 95.1% on the USPTO-50K test set, representing an 8.6% improvement over the baseline.
• The model consistently improves prediction accuracy for rare transformations, demonstrating better generalization on low-frequency reaction types.
• Iterative training increases top-1 accuracy by 1.1% (from 95.1% to 96.1%) when using 50 iterations, showing the effectiveness of self-boosting refinement.
• The framework maintains high computational efficiency, with inference time per product ranging from 8.03 ms to 115.07 ms on USPTO-50K using a single GPU.
• The model is more lightweight than prior transformer-based methods (44.5M parameters vs. 75.4M in RetroPrime), enabling easier deployment.
• Despite high molecular similarity (Tc ≥ 0.95), some predictions still show errors—such as incorrect stereochemistry or spurious reagents—highlighting persistent challenges in model generalization.