Retrieved reactions help retrosynthesis beyond training coverage
Reaction precedents extend retrosynthetic prediction beyond training coverage · arXiv:2610.40249
In plain words
To plan a chemical reaction backward, a model looks at a target molecule and suggests the molecules that could make it. This paper tests whether showing the model recorded reactions while it makes a prediction can help when its training examples do not cover the needed chemistry.
With the trained model held fixed, adding records to its lookup collection raised exact precursor-set matches for reactions whose transformations had been withheld from training. The gains appeared on both historical and 2025 test reactions.
The research
When a molecule needs a transformation missing from a model’s training, can reaction records supplied at prediction time help it identify the right starting materials? The question matters because a model’s training data may omit needed transformations or predate relevant chemistry.
The authors add five retrieved reaction examples to a target molecule’s input, letting a fixed model use records absent from its task-specific training.
To test this, the authors trained DianShi-RetroStepLM-RAG with restricted reaction records, then held its checkpoint fixed and switched only the records it could retrieve.
On reactions whose transformation was excluded from both training and restricted retrieval, mean exact-match accuracy rose from 13.0% to 58.0% on the 2001–2024 test set and from 4.9% to 23.1% on the 2025 set, averaged across three alternative withheld-template assignments.
A separate time-based test also improved when a pre-2017-trained checkpoint gained access to later records: exact matches rose on both 2001–2024 and 2025 reactions.
Together, these fixed-checkpoint comparisons show that retrieved records can extend predictions beyond the training coverage tested here.
The retrieval expansion changed corpus size, recency and composition together, so these experiments do not show which property drove the gains. Exact matches to recorded precursors also do not establish that a proposed reaction can be carried out successfully.
What the evidence shows
- The coverage question
Retrosynthesis asks which precursor molecules could produce a target molecule. A model may struggle when its training data lacks the needed transformation or predates relevant chemistry.
The paper tests whether the model can use reaction records supplied at prediction time to help fill those coverage gaps.
- Add examples at prediction
The method, DianShi-RetroStepLM-RAG, predicts a precursor set after receiving the target molecule and five retrieved reaction examples.
To isolate the role of available records, the authors trained with restricted data, kept the resulting model checkpoint fixed, and changed the collection used to retrieve examples at prediction time.
- Withheld chemistry improves
For the transformation test, the authors withheld reaction templates from both training and the matching retrieval collection, using three alternative template assignments. With restricted retrieval, mean exact-match accuracy was 13.0% on 2001–2024 reactions and 4.9% on 2025 reactions. Giving the same checkpoint access to the full collection raised these means to 58.0% and 23.1%.
- The time gap matters
A separate test trained the model on records from before 2017, then expanded the records it could retrieve. Exact precursor-set matches rose from 58.6% to 70.7% on 2001–2024 reactions and from 28.7% to 36.5% on 2025 reactions. On USPTO-50K, a set of patent reactions from 1976–2016, the score barely changed: 73.5% to 73.3%.
- A broader model check
The authors also tested 14 general-purpose models without task-specific updates. Across three reaction test sets, examples from the DianShi collection produced the highest accuracy among four evaluated conditions in all 42 model–test-set combinations, and improved every model over its no-example condition.
The collections differed in several ways, so this comparison does not identify which difference explains their ranking.
Deep
- Original paper authors
- Jiaxing Sun, Xuli Shen, Jiang Wu, Ruiyi Ding, Ruikang Zhang
- Submitted to arXiv
- License
- CC BY 4.0