New AI learning model leads the revolution of intelligent chemistry

One of the basic problems in organic chemistry is to predict the products that will be formed in chemical reactions, including the desired main products and various by-products that affect the yield Although for simple reactions, the product can be determined clearly, but for many complex organic reactions, this is not a simple thing Moreover, for the strict drug synthesis process, it is necessary to understand each by-product, even if the content is only one thousandth At present, the experiment is still the most important way to analyze the reaction results This is a time-consuming, expensive process that requires the guidance of an experienced chemist For the design and synthesis of specific target molecules, the experimental part is the largest short plate It may not take a few days to design a route, but the workload of verifying this route increases with the increase of reaction steps, which is usually counted in weeks, which is a heavy burden for the research and development of new drug process And often a route can be opened but the yield is too low to have to re plan the route If the main products and by-products can be predicted in advance and the conversion ratio of each product can be predicted, the existing work will be changed dramatically An example of a reaction model (source: arXiv) prior to the most advanced solution based on reaction templates The reaction template specifies the molecular subgraph pattern that can be applied and the corresponding graph transformation Since multiple templates can match a set of reactants, another model is trained to filter candidate products using standard surveillance methods The main disadvantages of this method are coverage and scalability A large number of templates are required to ensure that at least one can refactor the right product At present, these templates are made manually by experts or generated from response database by heuristic algorithm In addition to coverage, application templates involve chart matching, which makes checking a large number of templates too expensive Therefore, the current methods are limited to small data sets with limited types of responses Overview of new AI methods (source: arXiv) research team of Connor W Coley from Massachusetts Institute of Technology (MIT) worked with IBM to develop a template free program for self-learning to predict organic chemical reaction products This new AI system, similar to Google's artificial neural network, can provide a new set of patent reactions that have never been encountered after 24-hour learning without any organic chemistry knowledge, with an accuracy of 80.3% The team said this meant that the AI exceeded MIT's previous record (6.3%) for similar forecasting programs An example of reaction prediction (source: arXiv) this new AI system, instead of producing candidate products by reaction template, first predicts a group of atoms / bonds in the reaction center, and then produces candidate products by listing all possible bond configuration changes in the group Compared with the template based method, the framework runs 140 times faster and can be extended to a larger reaction database Reaction center identifier and candidate ranking model are learned from weisfeiler Lehman network and its variants This AI system can't let people know what's actually happening in chemistry, but it's similar to the "black box" model The program does not contain any reaction parameters, such as temperature or solvent, because these details are often not translated into a format that allows machine digestion So far, there is no test to verify the actual operation of the prediction program But the results are encouraging enough Imagine that if we can expand the database and let AI learn more existing chemical knowledge, then in the near future, it is likely to bring a revolution in the chemical industry, especially in new drug research and development, and the efficiency of new drug research and development is likely to be greatly accelerated Its application prospect is limitless Paper link: https://arxiv.org/abs/1709.04555