English 
Japanese 
Published on: July 9, 2026
Industrial and commercial synthesis is necessary step to go through in process chemistry. Aiming for and accumulation of reaction success is one of the key factors of synthesis of the target molecule. If there is a global model for the prediction of reactivity with high accuracy, it would be undoubtedly useful tools for synthesis.
Here is an intensive investigation on developing and testing global reactivity model by groups of Janssen Pharmaceutica N.V.1) They call their model Bert-Enriched-Embedding (BEE). They basically follow the code and approach for prediction of chemical reaction yields by deep learning2) and polished it up by adapting it to a binary classification model, using Bidirectional Encoder Representations from Transformers (BERT)3) and SMILES.
There are open database like USPTO4) and commercial ones including Reaxys5) and Pistachio.6) At first sight, it is not so hard to gather a dataset for deep learning-based approach for building up a global reactivity model. However, you will soon realize that the isolated reaction yield reported in a paper depends on a plethora of factors and they are too many to mention. Substrate purity, reagent lot, reaction vessel, the exact temperature of “room temperature”, “technique” of researcher are the representative examples. Those “dirty”, or ununiform, data complicates the model and it definitely hampers the construction of global reactivity model.
The Janssen’s group intended to build up a tool for chemists. With the dataset it is not so easy to select the best reagents and substrates for production of the desired compound. Instead, they made a question to solve much easier: “Which is the better reagent and/or starting material?” Even with this simplification, chemists can raise the success rate if the model is useful for universal reactivity prediction.
Here we don’t get into the validation in detail because the authors used their internal data. Thus we won’t be able to check the result by ourselves. All in all, they checked the model in comparison with the state-of-the-art one in a chronological way and then evaluated their own model in a prospective manner. Both in silico and synthesis teams collaborate with each other to furnish a reliable model for industrial application in synthesis.
It is worth reading this article for two reasons. First, their global reactivity model, BEE, is designed and developed for highly practical perspective. One would say that it is the best if the model generates the optimal set of reagents and conditions for the reaction of interest. But their idea is to help decision making of chemists. Their approach is not directly practical for process chemistry because the threshold is set to less than 5% or not. But in this way, potential candidate reaction could be found and usual reaction optimization would be enough for process chemists. This model opened the possibility of higher-level optimization of reaction and synthesis in a global way.
Secondly, it tells us that collaboration is essential for this type of technology and software development. Intense engagements of scientist in cheminfomatics and chemistry enabled them to develop a practical tool on their own. You can easily imagine that the design of the model came from the understanding of the needs on the ground level. Mutual communication and collaboration are the key for useful technology development.
Global reactivity model is widely applicable so long as it’s designed for chemists. The Janssen’s research and outcome is stimulating in that internal collaboration across department allowed them to reach the practical tool development.
You can say that it’s possible by intercompany collaboration. We really appreciate our collaborators for intense engagement to reach THE goal. We would love to collaborate with you in any sense from drug discovery to tool development. Conversation provides us an opportunity for brainstorming and we love to have a discussion with you.