Abstract

Molecular generation using AI approaches have gained wide recognition, thanks to the availability of numerous approaches (LSTM, VAE, GAN…).¹ Unfortunately those algorithms suffer from producing complex to unfeasible molecules in terms of synthetic feasibility.² Recently new AI approaches involving chemical reactions have been described.³ Mixing an initial molecular structure with commercial starting materials in the context of a reaction is a natural way to define a policy for generating new compounds. This method ensures synthetic accessibility of the generated molecules as the synthetic scheme is inherently obtained during the design process.

In this work we built a library of molecules starting from a defined fragment possessing two exit vectors where commercial starting materials can react with.  The reaction prediction was performed with a template-based neural network coupled with an applicability domain estimator. The goal was to find hits for the PIM-1 protein,⁴ an isozyme of PIM kinase found in many cancers which inhibition is a promising approach to stop cell-growing and reproduction.

The library was generated under the constraints of drug likeness metrics and structure-based scoring. The best scoring compounds were profiled by medicinal chemist and some of them were found similar to known PIM-1 inhibitors. According to our knowledge it is the first report that uses a generative AI incorporating synthetic feasibility by design, under structure-based constraints with a single fragment as starting point. This clearly demonstrates the tremendous potential of such approach to easily generate valuable new starting points in the context of a hit discovery program.