Source-linked AI summary
RetroXpert: Decompose Retrosynthesis Prediction like a Chemist
Chaochao Yan, Qianggang Ding, Peilin Zhao, Shuangjia Zheng, Jinyu Yang, Yang Yu, Junzhou Huang
TL;DR
Retrosynthesis involves a large search space and route choices that require expertise. RetroXpert decomposes prediction into reaction-center identification and reactant generation through intermediate synthons. On USPTO-50K, it reports higher Top-1 accuracy than the cited state-of-the-art results, while its current approach focuses on finding the best reactant set.
Problem
Retrosynthesis has a large search space, incomplete mechanistic understanding, and route selection that depends on expert judgment and multiple feasibility factors.
Source: §1 · p. 1Method
RetroXpert identifies reaction centers with EGAT to form synthons, then uses RGN to generate the associated reactants.
Source: §1 · p. 2Results
70.4% and 65.5% Top-1 accuracy with and without reaction type, respectively, exceed the cited state-of-the-art results of 63.2% and 52.6% on USPTO-50K.
Source: §1 · p. 2Takeaways & Limitations
Intermediate synthons determine reactant generation order, and augmenting RGN training with unsuccessfully predicted synthons is reported to improve its robustness.
Source: §1 · p. 2RetroXpert is designed to find the best reactant set, while template-based methods perform better at Top-20 and Top-50 prediction by enumerating different templates.
Source: §3.2 · p. 8
Abstract
from arXiv · showhide
Retrosynthesis is the process of recursively decomposing target molecules into available building blocks. It plays an important role in solving problems in organic synthesis planning. To automate or assist in the retrosynthesis analysis, various retrosynthesis prediction algorithms have been proposed. However, most of them are cumbersome and lack interpretability about their predictions. In this paper, we devise a novel template-free algorithm for automatic retrosynthetic expansion inspired by how chemists approach retrosynthesis prediction. Our method disassembles retrosynthesis into two steps: i) identify the potential reaction center of the target molecule through a novel graph neural network and generate intermediate synthons, and ii) generate the reactants associated with synthons via a robust reactant generation model. While outperforming the state-of-the-art baselines by a significant margin, our model also provides chemically reasonable interpretation.
1 Introduction
Retrosynthesis is difficult because its search space is large and route selection depends on expert judgment and multiple feasibility factors. RetroXpert proposes a two-step, synthon-based approach and reports higher Top-1 accuracy than the cited state-of-the-art baseline.
1 Introduction
- Retrosynthesis faces a huge search space, incomplete mechanistic understanding, and subjective route selection requiring expertise and experience.Route feasibility can depend on reactant availability, reaction conditions, yield, and toxic byproducts.Source: §1 · p. 1
- Template-based methods are limited to the chemical space covered by their template databases, constraining discovery of novel reactions.Manual encoding is impractical given the growth in reaction counts.Source: §1 · p. 2
- Template-free sequence models lack interpretability and chemistry knowledge, while G2Gs permits at most one bond disconnection and generates multiple reactants independently [28].The paper notes that independent reactant generation ignores relationships among multiple reactants.Source: §1 · p. 2
- RetroXpert uses EGAT to identify reaction centers and form synthons, then uses RGN to generate associated reactants; unsuccessful predicted synthons augment RGN training.The intermediate synthons determine reactant generation order, and the augmentation is reported to improve RGN robustness.Source: §1 · p. 2
- 70.4% and 65.5% Top-1 accuracy with and without reaction type, respectively, exceed the cited state-of-the-art results of 63.2% and 52.6%.The comparison is on USPTO-50K.Source: §1 · p. 2
2 Methodology
RetroXpert decomposes retrosynthesis into reaction-center identification that creates synthons, followed by generation of the associated reactants. EGAT predicts bond disconnections, while the reactant generation network uses synthons to organize reactant prediction.
2 Methodology
- Retrosynthesis is decomposed into identifying bonds to disconnect and generating the associated reactants from the resulting synthons.Each synthon is a substructure of a reactant, making the reactant count and substructures known before generation.Source: §2 · p. 3, Figure 1 · p. 3
2.1 EGAT for reaction center identification
- EGAT predicts each bond’s disconnection probability and adds a graph-level prediction of the number of disconnections to help distinguish multiple reaction centers.EGAT also learns edge embeddings, which it uses to predict bond-disconnection probabilities.Source: §2.1 · pp. 3–4
- EGAT incorporates atom and bond features, with optional semi-template indicators that encode product-side subgraph patterns from training reactions.The method remains template-free because it does not plan retrosynthesis using full reaction templates and EGAT works without semi-templates, with slight performance degradation.Source: §2.1 · p. 5
2.2 Reactant generation network
- The reactant generation network formulates synthons-to-reactants prediction as sequence-to-sequence modeling, using reaction types when available, the product, and synthons as input.The target reactants are arranged according to the synthons; synthons are converted to SMILES even when they may be chemically invalid.Source: §2.2 · p. 5, Figure 3 · p. 5
- Reversely arranged synthons and reactants augment training, while encoder-decoder and masked self-attention address input coverage and dependence on previously generated reactants.These design choices target permutation invariance, use of all source information, and dependence on earlier outputs.Source: §2.2 · pp. 5–6
- Training on unsuccessfully predicted synthons is intended to let the reactant generator recover desired reactants when EGAT predicts reaction centers incorrectly.The authors identify difficulty distinguishing coexisting reaction centers as EGAT’s major bottleneck.Source: §2.2 · p. 6
3 Experiments
Experiments evaluate RetroXpert on USPTO-50K and USPTO-full, including reaction-center and reactant-generation results. On USPTO-50K, it improves Top-1 accuracy over the cited baseline, while template-based methods do better at Top-20 and Top-50.
3 Experiments
3.1 Reaction center identification results
- 99.2% (with reaction type) and 86.4% (without) of test molecules had their disconnection-bond count predicted correctly by Aux.Source: §3.1 · p. 7
- 86.0% (w/) and 64.9% (wo/) were EGAT disconnection accuracies using the auxiliary-task prior, versus 74.4% and 51.5% for Main alone.Source: §3.1 · p. 7
3.2 Reactant prediction results
- With ground-truth synthons, augmentation boosts RGN accuracy when reaction type is given but generally lowers it without type; these results are an upper bound.Source: §3.2 · p. 7, Table 2 · p. 8
- 70.4% Top-1 accuracy with reaction types exceeded 63.2%, while 65.6% without reaction types exceeded 52.6%.Source: §3.2 · p. 8
- Template-based GLN and RetroSim outperform RetroXpert at Top-20 and Top-50 by enumerating multiple reaction templates.Source: §3.2 · p. 8
- About 9/10 of the roughly 10% Top-2 versus Top-1 gap comprised Top-1 predictions experienced chemists judged reasonable.Source: §3.2 · p. 8
4 Large scale experiments
On USPTO-full, RetroXpert significantly outperforms SCROP and GLN. The authors report this result as evidence that the method scales to a larger real-world dataset.
4 Large scale experiments
- On USPTO-full, RetroXpert significantly outperformed SCROP and GLN, which the authors say demonstrates scalability to a large real-world dataset.Source: §4 · p. 8
- USPTO-full has 760K training reactions; reversing synthons produced 1.35M training examples and 1.65M total RGN training examples after adding unsuccessful synthons.Source: §4 · p. 8
- The USPTO-full setup retained 3,788 semi-templates appearing at least 10 times, with Nmax=5 covering 99.87% of training data.Source: §4 · p. 8
5 Prediction visualization
The visualizations show EGAT identifying reaction centers with auxiliary-task guidance and RGN recovering desired reactants despite incorrect predicted synthons. Table 3 compares retrosynthesis results with existing methods.
5 Prediction visualization
- Even when predicted synthons differ from ground truth, RGN still generates the desired reactants.Figure 5 reports a Top-1 prediction matching the ground truth.Source: §5 · p. 8, Figure 5 · p. 9
- Auxiliary-task guidance lets EGAT identify the true reaction center when candidate bonds and surrounding structures look similar.Figure 4 marks EGAT’s predicted reaction center and disconnection probability in pink, with the ground-truth disconnection shown as a blue cross.Source: §5 · p. 8, Figure 4 · p. 9
- Table 3 compares retrosynthesis results with existing methods; NeuralSym results are copied from [10], while specified SCROP and GLN runs use USPTO-full.Source: Table 3 · p. 9
6 Discussion
The discussion identifies a limitation in how retrosynthesis is evaluated: multiple syntheses may be valid, while the current metric considers only the recorded reaction.
6 Discussion
- Multiple valid syntheses may exist, but the current metric evaluates only the given reaction, motivating better evaluation metrics.Source: §6 · p. 10
Broader Impact
RetroXpert is presented as a template-free, two-step retrosynthesis method inspired by chemists and described as scalable to large real-world datasets.
Broader Impact
- RetroXpert is a template-free, chemist-inspired two-step method that the authors describe as scalable to large real-world datasets.Source: Broader Impact · p. 10
A Dataset information
The appendix reports per-category results on USPTO-50K alongside dataset statistics and model-feature details. It also describes the Transformer architecture and its parameter settings.
A Dataset information
- USPTO-50K has 10 recognized reaction types with an extremely unbalanced distribution; Tables 5 and 6 report training-reaction disconnection-bond statistics for USPTO-50K and USPTO-full.Source: A Dataset information · p. 13, Table 4 · p. 13, Table 5 · p. 13
B Atom and bond features
- Table 7 lists EGAT atom features, including retrosynthesis-specific extensions, and Table 8 lists one-hot bond features.Table 7 notes that atomic mass is scaled as a real-valued feature and that the semi-template size is 654 for USPTO-50K.Source: Table 7 · p. 13, Table 8 · p. 14
C Transformer
- The Transformer encoder uses stacked self-attention and feed-forward layers; the decoder adds cross-attention and masked self-attention to combine source and generated-target information.Positional encoding supplies a position-dependent signal to token embeddings; Figure 6 omits residual connections and layer normalization for simplicity.Source: C Transformer · p. 14, Figure 6 · p. 14
C.1 Parameters setting
- The Transformer uses four encoder and decoder layers of size 256, eight attention heads, and batches of 4096 tokens with gradients accumulated over four batches.Label smoothing is set to 0 because a nonzero value deteriorates discrimination.Source: §C.1 · p. 15
D.1 Per category performance
- RetroXpert performs well across reaction categories and is comparable to or better than RetroSim and GLN in 9 of 10 categories; a similar conclusion holds when reaction type is unknown.For rare classes 4 and 10, it performs much better than GLN; for common types 1 and 2, the type prior helps little.Source: Figure 7 · p. 15, Figure 8 · p. 15, §D.1 · p. 16
D.2 Ablation study of atom features
The ablations show that semi-templates help EGAT identify reaction centers but are not required. Training and test augmentation have little influence individually, while using both significantly boosts retrosynthesis performance.
D.2 Ablation study of atom features
- Removing semi-templates causes a slight EGAT performance drop, although the feature is helpful rather than required for finding reaction centers.Table 9 reports the atom-features ablation; the prose describes the change without giving a numerical value.Source: Table 9 · p. 16, §D.2 · p. 16
D.3 Ablation study of augmentation in RGN
- Using both training and test augmentation significantly boosts retrosynthesis performance, while either augmentation alone has only a tiny influence.The ablation compares settings with and without reaction type.Source: Table 10 · p. 16, §D.3 · p. 16
- Training augmentation adds EGAT-missed synthons to RGN training, while test augmentation feeds predicted synthons regardless of EGAT success, avoiding an EGAT-imposed performance cap.Without test augmentation, reactants are predicted only for products whose reaction centers EGAT identifies successfully.Source: §D.3 · p. 16
E Top-1 and Top-2 predictions
Chemist review found that many apparent Top-1 errors were reasonable reactants; Figure 9 illustrates cases where both Top-1 and Top-2 predictions are reasonable. These include alternative reagents and transformations across several reaction classes.
E Top-1 and Top-2 predictions
- 9 in 10 apparent Top-1 errors whose Top-2 matched the ground truth were judged reasonable and valid by experienced chemists.About 10% of Top-1 predictions were counted as wrong while their Top-2 predictions matched the ground truth; Figure 9 shows examples where both predictions are reasonable.Source: E Top-1 and Top-2 predictions · p. 17, Figure 9 · p. 17
- The reviewed alternatives include metal-catalyzed cross-couplings, N- and O-alkylations, ester saponification, esterification, and deprotection.These are the major reaction groups identified among cases where both Top-1 and Top-2 can be considered correct.Source: E Top-1 and Top-2 predictions · p. 17
- Examples include different protecting groups yielding the same alcohol and both methyl and ethyl esters yielding the corresponding acid under certain conditions.The paper also notes alternative Suzuki-coupling reagents and real reactants among its Top-1 predictions.Source: E Top-1 and Top-2 predictions · p. 17