acceptodds
Under review as a conference paper at ICLR 2027

MGSEKformer: Multi-Granularity Graph-Sequence Enhanced Knowledge Transformer for Single-Step Retrosynthesis Prediction

Abstract

Single-step retrosynthesis prediction serves as a fundamental building block for multi-step synthesis route planning in modern drug discovery and organic material synthesis. To improve both efficiency and accuracy, researchers have explored a variety of deep learning frameworks, most of which focus on molecular feature extraction at a single granularity—either the atomic or functional group level. However, these approaches have not sufficiently addressed the differences and interactions between these two distinct granularities in terms of their impact on chemical reactions, which limits further improvements in predictive performance. To this end, we propose the Multi-Granularity Graph-Sequence Enhanced Knowledge Transformer framework (MGSEKformer) for single-step retrosynthesis. Specifically, we employ two PAGformers to extract local structural priors and long-range dependencies at the atomic and functional group levels, respectively, and use cross-attention and gating units to model their interactions and aggregate complementary information adaptively. Furthermore, building upon the standard feedforward transformation, we design a knowledge-attention branch to replace the fixed-weight fusion based on simple concatenation in a conventional Kformer, enabling adaptive injection of variable-length chemical prior knowledge and overcoming the reliance on predefined reaction templates. Comparative experiments against 13 baseline models show that the MGSEKformer achieves optimal multi-candidate inference performance. Ablation studies further demonstrate that functional-group feature encoding and cross-granularity interactive fusion substantially improve the ranking accuracy of high-confidence candidates, effectively supporting practical retrosynthetic route design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-step Retrosynthesis Prediction, Multi-granularity Graph-sequence Enhancing, Position-Aligned Graph-injected Transformers, Variable-length Knowledge Injection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_J7P9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies template-free single-step retrosynthesis and argues that atom-level and functional-group-level features affect bond changes differently. MGSEKformer encodes an atom-level SMILES/graph pair and a pooled-subgraph pair produced by a pretrained MESPool module with parallel graph-injected Transformers, injects ElementKG element knowledge through an attention branch in the FFN, fuses the two granularities with bidirectional cross-attention and gating, and decodes reactant SMILES. It reports results on USPTO-FULL against 13 baselines, together with USPTO-50K results. The motivation is chemically reasonable, but the comparison protocol and the component evidence do not yet support the claimed breadth of improvement.

Critical strengths

Chemically grounded multi-granularity motivation. The paper explains why granularity matters with concrete chemistry, such as the acidic site of β-dicarbonyl compounds and positional activation effects, which is more specific than the usual multi-scale helps argument. Explicit alignment and gated cross-granularity fusion give a coherent way to combine local structure with sequence context.

Critical weaknesses

(i) The paper builds a new deduplicated random 8:1:1 split of USPTO-FULL, while the baseline numbers appear to be taken from papers that used the original split, so the small margins over RXNGraphormer may compare different test sets. Training uses R-SMILES augmentation, yet R-SMILES itself is not compared.

(ii) USPTO-50K is a subset of USPTO-FULL, overlap between its test reactions and the FULL training data needs to be ruled out. On USPTO-50K the model is below several baselines at every k, which is not discussed and conflicts with the claim of optimal multi-candidate performance in the abstract. I think the conclusion should be scoped to strong multi-candidate results in the reported USPTO-FULL setting.

Minor

The main text points to a theoretical derivation in the appendix, which contains architecture details only.

The paper's topic is fair enough for a chemical journal, but these papers are usually hard to accept for ICLR, unless you are solving a new direction in chemistry with AI-adapted modules. Currently, this does not affect my scoring, and I will recheck other reviewers' comments.

Questions for authors

See weaknesss

Reviewer_e9b7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MGSEKformer predicts single-step reactants from a product molecule. It builds atom-level and pooled coarse-grained product representations, processes them with parallel graph-aware Transformers, exchanges information through cross-attention and gating, injects element knowledge through an attention branch, and generates reactant SMILES with a Transformer decoder (Sections 3–4). On USPTO-FULL, it reports Top-1/3/5/10 exact-match rates of 47.3/63.4/68.6/73.7% (Table 1). Its Top-3/5/10 scores are the highest in that table, while Top-1 is 0.1 point below RXNGraphormer. On USPTO-50K, it reports 47.0/66.3/72.3/77.8%, below several listed baselines. The ablation improves Top-1 from 46.0% to 47.3% over an atom-only baseline, but Top-10 decreases from 74.1% to 73.7% (Table 2). A ten-seed plot reports variation for the full model (Appendix A.5).

Critical strengths
  1. The model combines atom-level structure, a coarser pooled view, and element information in a single reactant-generation pipeline. On USPTO-FULL, its Top-3/5/10 scores exceed the listed alternatives (Table 1), making the overall design worth studying.
  2. The paper includes a component ablation and trains the full model with ten seeds (Table 2; Appendix A.5). These checks provide more information than a single leaderboard score, even though the ablation gains are small and the seed results are only plotted.
Critical weaknesses
  1. The comparison protocol is not established clearly enough to support a strong performance claim. Section 4 describes a deduplicated random 8:1:1 split and 5× R-SMILES augmentation, but does not state whether the 13 baselines were rerun on exactly these data with the same preprocessing and candidate evaluation. Table 1 also shows that on USPTO-50K, MGSEKformer trails LocalRetro by 6.4, 11.2, 13.6, and 14.6 points at Top-1/3/5/10. The paper's broad claims about leading multi-candidate performance need to be restricted to the reported USPTO-FULL comparison and qualified by protocol comparability.
  2. The evidence does not yet validate the proposed chemical interpretation. MESPool is trained using reconstruction and bond-type priors, while functional-group identification is illustrated with one molecule and no quantitative group-label evaluation (Appendix A.2). In Table 2, adding all components raises Top-1 by 1.3 points but lowers Top-10 by 0.4 points relative to the atom-only baseline. Exact-match Top-k measures recovery of recorded reactants; it does not by itself show that ElementKG improves chemical validity or expands feasible reaction routes. Figure 2 is a single qualitative example and even treats a proposed two-step alternative as evidence for single-step retrosynthesis.
Questions for authors
  1. Were all 13 baselines rerun on the same deduplicated USPTO-FULL and USPTO-50K splits, with the same preprocessing and Top-k evaluation? If not, please identify the source and protocol for each result and provide matched comparisons with the strongest baselines.
  2. How accurately does MESPool recover independently defined functional groups across the test sets? Can you compare it with a simple rule-based grouping and report the effect of its pooling threshold?
  3. Can you provide a direct measure of SMILES validity, reaction feasibility, or candidate diversity to support the ElementKG and route-coverage claims? How do you explain the lower Top-10 ablation result and the weaker USPTO-50K performance?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.