Graph neural network-based artificial intelligence model... KAIST research team accurately predicts organic chemical reactions like chemists
In order to know the result of a chemical reaction from the captured image of the cover paper of Nature Machine Intelligence September issue, it is necessary to confirm it through experimentation, but using artificial intelligence that predicts chemical reactions with high accuracy can potentially replace these experiments or use substances in future digital laboratories. It is expected to reduce the time and cost required for development.
Organic chemists look at reactants and predict the results of organic chemical reactions to synthesize molecules with desired properties, such as drugs or organic light emitting diodes (OLEDs). However, it is generally time-consuming and expensive to directly confirm the product of a chemical reaction through an experiment.
In addition, organic chemistry reactions can produce different products from the same reactant, so even an experienced organic chemist cannot accurately predict all chemical reactions.
In order to overcome this limitation, research on predicting organic reactions using artificial intelligence (AI) is actively taking place. Most research focuses on how to use a language translation model to think of reactants and products in two different languages and translate them from one language to another.
Although this method has high prediction accuracy, it is difficult to interpret that the artificial intelligence understands the chemistry and predicts the product, making it difficult to trust the results predicted by the model.
In addition, KAIST (President Lee Kwang-hyeong) developed artificial intelligence that thinks like a chemist by Professor Youseong Jung's research team, in which Shuan Chen, Ph.D. The artificial intelligence developed by the research team accurately predicts the outcome of organic reactions.
Established a chemical reaction prediction model, a generative model that simultaneously learns the composition information and structural information of inorganic compounds, and learns the existing material database. It is possible to discover new hidden and promising new substances that were not possible.
Professor Jeong's team designed a model based on chemical intuition, and not only was able to chemically explain the results predicted by the model, but also achieved very good prediction accuracy in public databases.
Professor Jung's team got the idea from how chemists predict the outcome of a reaction. Chemists identify reaction centers and apply chemical reaction rules to predict possible products. Following this process, chemical reaction rules were derived from a public chemical reaction database.
In order to predict the chemical reactivity of molecules based on chemical reaction rules, we developed a Graph Neural Network (GNN) model that treats molecules as graphs. Putting the reactants into the model identifies the chemical reaction rules and reaction centers and successfully predicts the products.
Professor Jung's team succeeded in predicting organic reactions with more than 90% accuracy using USPTO data, which is widely used in chemical reactions. The developed model can refer to 'prediction uncertainty' that provides high reliability to the model in actual use.
For example, the accuracy of a model that is considered low uncertainty increases to 98.6%. The model was found to be more accurate than a small group of synthetic experts at predicting a series of randomly sampled organic reactions.
With the success of this study, the research team proved that the strategy of designing a neural network in the same way as a chemist thinks is more rational and shows better performance than the existing method that uses a model that has performed well in other fields.
The research team expects that this research will dramatically speed up the molecular design process, and it is expected to have practical applications in the development of new compounds. Professor Yooseong Jung's team is currently preparing to apply for a patent on the research results.
On the other hand, the results of this study, in which KAIST Biochemical Engineering Ph.D. Shuan Chen participated as the first author, were published in the international academic journal Nature Machine Intelligence, 'Generalized Template-Based Graph Neural Network (A) for Accurate Organic Reactivity Prediction'. Generalized-template-based graph neural network for accurate organic reactivity prediction- see )' was selected and published as the cover paper of the September issue.
Comments