Source-linked AI summary
A Graph to Graphs Framework for Retrosynthesis Prediction
Chence Shi, Minkai Xu, Hongyu Guo, Ming Zhang, Jian Tang
TL;DR
Retrosynthesis prediction requires finding reactants for a target molecule, while existing template-based methods are computationally expensive and limited by template coverage. G2Gs addresses this with template-free graph-to-graphs translation through synthon splitting and variational reactant-graph generation, outperforming template-free baselines by up to 63% in top-1 accuracy and approaching template-based performance.
Problem
Retrosynthesis prediction seeks reactants for a target molecule, while template-based approaches are computationally expensive and suffer from limited coverage.
Method
G2Gs identifies reaction centers to split a target molecular graph into synthons, then translates each synthon into reactant graphs with variational graph translation.
Results
G2Gs outperformed existing template-free approaches by up to 63% in top-1 accuracy and approached state-of-the-art template-based performance.
Takeaways & Limitations
G2Gs does not require domain knowledge and scales well to large datasets, making it attractive in practice.
Takeaways & Limitations
The current work focuses on standard single-outcome reactions and identifies end-to-end training and multi-step retrosynthesis as future extensions.
Abstract
from arXiv · showhide
A fundamental problem in computational chemistry is to find a set of reactants to synthesize a target molecule, a.k.a. retrosynthesis prediction. Existing state-of-the-art methods rely on matching the target molecule with a large set of reaction templates, which are very computationally expensive and also suffer from the problem of coverage. In this paper, we propose a novel template-free approach called G2Gs by transforming a target molecular graph into a set of reactant molecular graphs. G2Gs first splits the target molecular graph into a set of synthons by identifying the reaction centers, and then translates the synthons to the final reactant graphs via a variational graph translation framework. Experimental results show that G2Gs significantly outperforms existing template-free approaches by up to 63% in terms of the top-1 accuracy and achieves a performance close to that of state-of-the-art template based approaches, but does not require domain knowledge and is much more scalable.
1. Introduction
Retrosynthesis prediction seeks reactants for a target molecule, but existing approaches face large transformation spaces and limitations from template matching. G2Gs formulates the task as template-free graph-to-graphs translation and reports strong benchmark performance.
- Retrosynthesis prediction seeks a set of reactants that can synthesize a target molecule, supporting synthesis planning and drug discovery.
- Existing methods include template-based models that match targets against large reaction-template collections and template-free models.
- G2Gs represents molecules as graphs and translates a target graph into reactant graphs without reaction templates.
- G2Gs identifies reaction centers to split synthons, then uses variational graph translation to generate final reactant graphs.
- Up to 63% higher top-1 accuracy than existing template-free baselines was achieved on USPTO-50k, while performance approached state-of-the-art template-based methods.
2. Related Work
Prior retrosynthesis methods rely on reaction templates or sequence representations, while G2Gs operates directly on molecular graphs without templates. It adapts graph generation to retrosynthesis by modeling diverse reactants and reaction conditions with latent variables.
- Template-based methods: Template-based methods face a crucial challenge: selecting chemically feasible reactions from hundreds of qualified templates matched to a target molecule.These templates may be hand-crafted by experts or extracted automatically from large chemical databases.
- Template-free methods: Template-free sequence-to-sequence approaches use SMILES representations but can ignore rich chemical contexts contained in molecular graphs.G2Gs instead operates directly on graph structures and avoids reaction templates and computationally expensive graph isomorphism.
- G2Gs: G2Gs generates 100% chemically valid predictions with high accuracy while excluding reaction templates and computationally expensive graph isomorphism.This passage states the framework’s reported validity and computational advantages over the contrasted approaches.
- G2Gs: G2Gs differs from related reaction-center learning by using a distinct reaction-center definition for the retrosynthesis prediction task.The comparison concerns both the operational definition of reaction center and the targeted task.
- G2Gs: Unlike general molecular graph generation, G2Gs targets retrosynthesis and introduces low-dimensional latent variables to model diverse reactants and reaction conditions.The retrosynthesis setting is intrinsically multimodal because a product can correspond to reactants under varying conditions such as reagents and temperature.
3. The Graph to Graphs Framework
G2Gs formulates retrosynthesis as graph-to-graphs translation: it identifies reaction centers, splits the product into synthons, and generates reactants through graph transformations. Its template-free design combines reaction-center prediction with variational translation and scales independently of the number of reaction rules.
- Reaction Center Identification: G2Gs estimates reactivity for product-graph atom pairs and selects high-scoring pairs above a threshold as reaction centers.The model formulates reaction-center identification as binary link prediction using atom-indexed reactant and product graphs.
- Reaction Center Identification: Top-k reaction-center selection can produce more diverse synthesis routes, but increases inference time.The default procedure selects the highest-scoring atom pair above a threshold.
- Synthon Extraction: Disconnecting reaction-center bonds partitions the product into synthons, which are translated individually into final reactant graphs.If no reaction center is identified, the whole product is treated as a synthon.
- Variational Graph Translation: Each synthon-to-reactant translation is modeled as a conditional graph generator using sequential graph-transformation actions.The translation procedure generates nodes and edges through transformations applied to the synthon graph.
- Variational Graph Translation: A latent vector captures uncertainty because the same synthon can correspond to different reactants, enabling diverse graph translations.The latent code is introduced to address the multi-modality of reactant prediction.
- Scalability Analysis: G2Gs bypasses reaction templates, and its model size scales linearly with the maximum molecule atom count while remaining invariant to the number of rules and reactions.Both framework modules use molecular graph embeddings rather than deployed reaction templates.
4. Empirical Studies
On USPTO-50k, G2Gs is evaluated against template-free and template-based baselines using top-k exact-match accuracy. It shows strong performance, with gains over template-free methods, competitiveness with GLN, and evidence that its modules support diverse and potentially valid reactant generation.
- Experiment Setup: USPTO-50k contains 50k atom-mapped reactions across 10 reaction types, split into 80% training and equally sized validation and test sets.
- Experiment Setup: G2Gs is compared with five baselines: two template-free sequence models and three template-based methods, including GLN.
- Experiment Setup: The evaluation uses top-k exact-match accuracy for k = 1, 3, 5, and 10, matching canonical SMILES predictions with ground truth.
- Predictive Performance: Up to 63% relative improvement in top-1 exact-match accuracy is achieved over template-free baselines when the reaction class is known, and up to 29% when it is unknown.
- Predictive Performance: G2Gs approaches or outperforms GLN, exceeding it on top-3 and top-5 exact-match accuracy when the reaction class is given.
- Predictive Performance: The reaction-center module can pinpoint centers in most cases, while the graph-translation module achieves high accuracy translating synthon graphs into reactant graphs.
- Case Studies: Different latent vectors generate diverse molecule structures from the same synthon, and visualization cases include successful ground-truth matches and a potentially valid route missed by exact matching.
5. Conclusion and Outlook
G2Gs reformulates retrosynthesis as template-free graph-to-graphs translation and uses variational graph translation to capture uncertainty and encourage diverse predictions. It outperformed the state-of-the-art template-free counterpart by up to 63% in top-1 accuracy, approached template-based performance, and scales to large datasets, while future work targets end-to-end and multi-step retrosynthesis.
- G2Gs reformulates retrosynthesis as a graph-to-graphs translation task and uses a variational graph translation module to capture uncertainty and encourage diversity.
- 63%: G2Gs outperformed the state-of-the-art template-free counterpart by up to 63% in top-1 accuracy and approached state-of-the-art template-based performance.
- G2Gs excludes domain knowledge and scales well to large datasets, making it attractive in practice.
- Future work will extend G2Gs toward end-to-end training and multi-step retrosynthesis tasks.