Tensor databases empower AI for science: A case study on retrosynthetic analysis

Retrosynthetic analysis is highly significant in chemistry, biology, and materials science, providing essential support for the rational design, synthesis, and optimization of compounds across diverse Artificial Intelligence for Science (AI4S) applications. Retrosynthetic analysis focuses on exploring pathways from products to reactants, and this is typically conducted using deep learning-based generative models. However, existing retrosynthetic analysis often overlooks how reaction conditions significantly impact chemical reactions. This causes existing work to lack unified models that can provide full-cycle services for retrosynthetic analysis, and also greatly limits the overall prediction accuracy of retrosynthetic analysis. These two issues cause users to depend on various independent models and tools, leading to high labor time and cost overhead. To solve these issues, we define the boundary conditions of chemical reactions based on the Evaluatology theory and propose BigTensorDB, the first tensor database which integrates storage, prediction generation, search, and analysis functions. BigTensorDB designs the tensor schema for efficiently storing all the key information related to chemical reactions, including reaction conditions. BigTensorDB supports a full-cycle retrosynthetic analysis pipeline. It begins with predicting generation reaction paths, searching for approximate real reactions based on the tensor schema, and concludes with feasibility analysis, which enhances the interpretability of prediction results. BigTensorDB can effectively reduce usage costs and improve efficiency for users during the full-cycle retrosynthetic analysis process. Meanwhile, it provides a potential solution to the low accuracy issue, encouraging researchers to focus on improving full-cycle accuracy. • We design and implement the first tensor database system. • We design a tensor format for storing chemical reactions. This format efficiently stores all key information related to chemical reactions, including reactants, products, and reaction conditions (such as solvents and reagents). We provide users with services for storing and retrieving chemical reactions. • We integrate multiple retrosynthetic analysis prediction models including reactants prediction and reaction conditions prediction. We also integrate SMILES embedding models. These integrations offer full-cycle retrosynthetic analysis services to users. Our work reduces usage costs and improves the pipeline’s efficiency. • We provide search and analysis functions. Through these interfaces, we re-rank and analyze the final prediction results. At the same time, we provide real reaction equations similar to the predicted results for user reference. These works can enhance the accuracy and interpretability of the final outcomes.