Retrosynthesis Prediction with Local Template Retrieval

Retrosynthesis, which predicts the reactants of a given target molecule, is an essential task for drug discovery. In recent years, machine learing based retrosynthesis methods have achieved promising results. this work, we introduce RetroKNN, local reaction template retrieval method to further boost performance template-based systems with non-parametric retrieval. We first build atom-template store and bond-template that contains templates in training data, then retrieve from these k-nearest-neighbor (KNN) search during inference. The retrieved are combined neural network predictions as final output. Furthermore, propose lightweight adapter adjust weights when combing KNN conditioned on hidden representation templates. conduct comprehensive experiments two widely used benchmarks, USPTO-50K USPTO-MIT. Especially top-1 accuracy, improved 7.1% dataset 12.0% USPTO-MIT dataset.These results demonstrate effectiveness our method.