This achievement will accelerate the development of new drugs.
Experts from the MIT Artificial Intelligence Laboratory have developed a neural network that is able to compose molecules with one hundred percent chemical certainty. In the future, the algorithm will make it easier for chemists.
Previously, there was already a system that automatically built a diagram of molecules. It was called SMILES. In it, each atom and bond had its own set of symbols. The result was a very long code that, within the SMILES logic, seemed infallible, but made no sense in the field of chemical laws. Now researchers have come up with a different system.
The new algorithm works directly with molecular graphs that represent the structural formula of a chemical compound. First, it "encodes" the incoming molecule, breaks up molecular graphs into clusters, each of which represents a specific building block. These clusters are created automatically thanks to machine learning. They are compiled into a "tree structure" that corresponds to the original graph.
During "decoding", the molecule goes from simple to complex. First, artificial intelligence creates a skeleton of a tree structure, and then complements and complicates it with connected clusters. Because of this, the reconstructed molecular graph becomes an exact copy of the original one.
The system was trained on 250 thousand molecular graphs. As a result, she achieved one hundred percent chemical accuracy in her models, while the previous SMILES system showed only 43 percent. The authors argue that the algorithm can also compose molecules based on given properties.
In March, employees at Beijing-based Baidu created a robotic system that reproduces human speech, faithfully replicating its unique characteristics.