Synthesis is at the heart of organic chemistry. It is the process of constructing a molecule through a series of reactions. Efficient molecular synthesis is significant, especially in areas like medicinal chemistry. However, it can be challenging to plan a synthesis for a complex molecule.
Retrosynthesis is an iterative process of breaking down a target molecule into a series of simpler molecules which are readily available. It provides a systematic way to plan a chemical synthesis. But, as the space of possible chemical reactions is vast, it is hard for chemists to make the right disconnections while optimizing for multiple factors such as the cost of precursors and efficiency of the reaction.
At its heart, the challenge of making a single retrosynthesis prediction can be recast as a Machine Translation task. Consequently, Natural Language Processing techniques can be leveraged to tackle this using Deep Learning.
This project is focused on extending the DeepChem Library to support Retrosynthesis models, specifically to make single step predictions. This will be achieved by implementing a Molecular Transformer model and enhancing support for Reaction datasets.