Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules

Abstract

We report a method to convert discrete representations of molecules to and\nfrom a multidimensional continuous representation. This model allows us to\ngenerate new molecules for efficient exploration and optimization through\nopen-ended spaces of chemical compounds. A deep neural network was trained on\nhundreds of thousands of existing chemical structures to construct three\ncoupled functions: an encoder, a decoder and a predictor. The encoder converts\nthe discrete representation of a molecule into a real-valued continuous vector,\nand the decoder converts these continuous vectors back to discrete molecular\nrepresentations. The predictor estimates chemical properties from the latent\ncontinuous vector representation of the molecule. Continuous representations\nallow us to automatically generate novel chemical structures by performing\nsimple operations in the latent space, such as decoding random vectors,\nperturbing known chemical structures, or interpolating between molecules.\nContinuous representations also allow the use of powerful gradient-based\noptimization to efficiently guide the search for optimized functional\ncompounds. We demonstrate our method in the domain of drug-like molecules and\nalso in the set of molecules with fewer that nine heavy atoms.\n

Affiliated Institutions

• University of New Brunswick CA
• University of Toronto CA
• Harvard University US
• University of Cambridge GB
• Canadian Institute for Advanced Research CA
• Princeton University US
• Google (United States) US

Related Publications