HiRes: Retrieval-Augmented Model for Reaction Condition Recommendation
HiRes, short for Hierarchical Reaction Representations, is a system that helps recommend chemical reaction conditions by leveraging retrieval techniques. It combines several advanced components like a graph encoder and transformation-aware cross-attention, along with multi-stream reaction fusion and a k-NN retrieval layer, which allows for accurate predictions based on reliable data. This model has excelled in the USPTO-Condition benchmarks, achieving impressive top-1 accuracies of 0.929 for Catalyst, 0.534 for Solvent, and 0.530 for Reagent. Notably, it matches the best recorded performance for Catalyst and outperforms models like REACON in the Solvent and Reagent categories. A paired bootstrap analysis confirms that its retrieval approach significantly boosts overall effectiveness.
Key facts
- HiRes is a retrieval-augmented condition recommendation system.
- It uses a graph encoder, transformation-aware cross-attention, multi-stream reaction fusion, and k-NN retrieval.
- Achieves Catalyst Acc@1 of 0.929, Solvent Acc@1 of 0.534, Reagent Acc@1 of 0.530.
- Ties best baseline on Catalyst, outperforms REACON on Solvent and Reagent.
- Paired bootstrap analysis validates retrieval integration.
- Designed for reaction condition recommendation after retrosynthetic disconnection.
- Learned reaction space serves as classifier feature and inspectable precedent memory.
- Published on arXiv with ID 2605.21420.
Entities
Institutions
- arXiv