Source-linked AI summary
Predicting Organic Reaction Outcomes with Weisfeiler-Lehman Network
Wengong Jin, Connor W. Coley, Regina Barzilay, Tommi Jaakkola
TL;DR
Predicting organic reaction products is difficult because reactions may involve hundreds of atoms, making exhaustive transformation search intractable, while template methods have coverage and efficiency limitations. The paper identifies reaction centers, enumerates chemically feasible candidate products, and ranks them with Weisfeiler-Lehman-based models. It reports higher accuracy than a leading template-based approach, 140-times-faster execution, and performance exceeding domain experts.
Problem
Organic reaction prediction requires selecting products from an intractable space of graph transformations, while template-based methods are limited by coverage and scalability.
Method
The method learns reaction-center atom and bond changes, enumerates chemically feasible candidate products, and ranks them using Weisfeiler-Lehman models including a Difference Network.
Results
83.9% and 77.9% accuracy were achieved on two USPTO-derived datasets, outperforming the baseline by 10% while running 140 times faster.
Takeaways & Limitations
The template-free framework supports reaction prediction at a scale suitable for much larger reaction databases.
Takeaways & Limitations
Template-based comparison is constrained by baseline scalability, limiting direct comparison to USPTO-15K and restricting the baseline to examples it could process.
Abstract
from arXiv · showhide
The prediction of organic reaction outcomes is a fundamental problem in computational chemistry. Since a reaction may involve hundreds of atoms, fully exploring the space of possible transformations is intractable. The current solution utilizes reaction templates to limit the space, but it suffers from coverage and efficiency issues. In this paper, we propose a template-free approach to efficiently explore the space of product molecules by first pinpointing the reaction center -- the set of nodes and edges where graph edits occur. Since only a small number of atoms contribute to reaction center, we can directly enumerate candidate products. The generated candidates are scored by a Weisfeiler-Lehman Difference Network that models high-order interactions between changes occurring at nodes across the molecule. Our framework outperforms the top-performing template-based approach with a 10\% margin, while running orders of magnitude faster. Finally, we demonstrate that the model accuracy rivals the performance of domain experts.
1 Introduction
Organic reaction prediction is difficult because complex reactions involve many possible graph transformations, while template-based methods face coverage and scalability limits. The paper proposes a template-free approach that identifies small reaction centers, enumerates candidates, and ranks them with Weisfeiler-Lehman-based models.
- Complex organic reactions are difficult to predict, making experimentation time consuming, expensive, and dependent on experienced chemists.
- Reactions can involve more than 100 atoms, making exhaustive exploration of graph edits intractable.
- Template-based approaches face coverage and scalability problems because many templates and costly graph matching are required.
- The proposed template-free method identifies a small reaction center, then enumerates chemically constrained candidate products and ranks them.
- The models use Weisfeiler-Lehman representations, attention to identify reaction centers, and a Difference Network to rank candidate transformations.
- 83.9% and 77.9% accuracy were achieved on two USPTO-derived datasets, outperforming the baseline by 10% and running 140 times faster.
2 Related Work
Prior reaction-outcome prediction work includes template-based candidate generation, template-free mechanistic approaches, and neural molecular-graph representations. The paper’s approach identifies atom interactions in the reaction center, enumerates chemically feasible bond configurations, and scores the resulting candidates.
- Template-based Approach: Template-based methods specify reaction templates in different ways and select final products from multiple generated candidates.Earlier work used hand-encoded templates, while later methods learned larger template sets from data and ranked candidates with neural models.
- Template-free Approach: Template-free reaction prediction has also been explored through mechanistic-level modeling of elementary steps.The paper distinguishes its graph-level prediction of overall reactant-to-product transformations from approaches requiring multiple mechanistic predictions.
- Template-free Approach: The paper’s workflow predicts pairwise atom interactions, selects the top K atom pairs, enumerates chemically feasible bond configurations, and scores the resulting candidate outcomes.Each bond configuration generates a candidate reaction outcome, which another model scores to identify the true product.
- Molecular Graph Neural Networks: Molecular graph representation is a central issue in reaction modeling, alongside established Morgan-fingerprint representations.Neural molecular representations aggregate features from neighboring graph nodes as an alternative to fixed hashing.
3 Overview
The approach predicts reaction centers on molecular graphs, enumerates constrained product candidates, and ranks them using WL-based representations that capture local, distal, and higher-order chemical effects.
- Reaction-center identification: Reaction centers are selected as the highest-scoring atom pairs, after which possible bond configurations within them are enumerated into candidate products.Valence and other constraints help eliminate invalid configurations despite the candidate set's exponential dependence on K.
- Reaction-center identification: The model predicts reactivity labels for atom pairs from learned representations incorporating surrounding chemical context.The labels indicate whether an atom pair's relation changes between reactants and products.
- Weisfeiler-Lehman representations: The WLN iteratively incorporates neighboring atom and bond features to produce context-dependent, isomorphism-invariant atom representations.These representations capture local chemical environments and can incorporate distal effects through a global attention model.
- Weisfeiler-Lehman representations: Reactivity prediction uses local atom representations or a global attention model that selectively incorporates chemically relevant atoms outside the reaction center.Sigmoid attention permits multiple atoms to be relevant to a particular atom.
- Training and limitations: The reaction-center labels are predicted independently because scoring O(N^2) atom pairs makes higher-order dependencies between different pairs computationally prohibitive.The authors report that independent prediction nevertheless yields sufficiently good performance.
- Candidate ranking: Candidate ranking compares atom-mapped reactants and products through difference vectors, with WLDN modeling higher-order interactions on a difference graph.The difference graph focuses computation near the reaction center and preserves neighbor dependencies between difference vectors.
4 Experiments
Experiments on USPTO datasets evaluate reaction-center coverage, candidate generation, product ranking, rare reactions, and human performance. The template-free approach achieves high coverage with fewer candidates and substantially lower runtime, while its global model and WLDN ranking improve performance.
- Experimental setup: The experiments use USPTO and USPTO-15K datasets with separate training, development, and testing splits.USPTO contains 480K reactions split 400K/40K/40K; USPTO-15K follows a 10.5K/1.5K/3K split.
- Reaction-center identification: Above 90% reaction-center coverage is achieved with K = 8 atom pairs.Coverage is measured by whether all true reaction-center pairs are predicted, ensuring the recorded product appears among generated candidates.
- Reaction-center identification: The global model consistently outperforms the local model, including when a disconnected reagent determines reactivity.In the illustrated amide reduction, the global model ranks bond 4-23 as most susceptible, whereas the local model excludes it from its top ten.
- Candidate generation: For K = 6, the model reaches 89.8% coverage with 60.1 candidates, versus 90.1% coverage with 482 candidates from 5006 templates.The template-based baseline uses templates extracted from training data with a minimum of five precedent reactions.
- Efficiency: The template-free method applies in less than 50 ms per test instance versus 7 seconds for 5006 templates, about 140 times faster.The baseline’s cost arises because each template application requires solving a subgraph isomorphism problem.
- Candidate ranking and human evaluation: The WLDN model outperforms the template-based baseline and WLN in candidate ranking, while performance remains reasonable for rare transformations.Human evaluation on 80 reactions gives the model 69.1% accuracy versus 48.2% average human accuracy and 72.0% for the best individual.
5 Conclusion
The paper concludes that a template-free framework predicts reaction outcomes by identifying reaction centers, enumerating bond changes, and ranking candidates with Weisfeiler-Lehman models. Compared with template-based methods, it runs 140 times faster and can scale to much larger reaction databases.
- The framework predicts a small reaction center before enumerating possible bond-configuration changes to generate candidate products.Its reaction-center identifier and candidate-ranking model use Weisfeiler-Lehman networks and variants.
- 140 times faster than the template-based approach, enabling application to much larger reaction databases.
- The approach is presented as a fully data-driven direction for chemical reaction prediction.
A Human Evaluation Setup
The human evaluation used 80 USPTO test reactions spanning eight template-popularity groups and involved ten chemists. The model achieved expert-chemist-level top 1 accuracy.
- Ten chemists predicted products from reactants across all evaluation groups.Eight were experienced organic chemists, while two were chemical-engineering graduate students with less formal training.
- The evaluation covered 80 randomly selected USPTO test reactions across eight groups defined by reaction-template popularity.Each group contained ten instances selected randomly from the USPTO test set.
- The model performed at the expert chemist level in top 1 accuracy.