Abstract:
Deep learning is being widely used for de novo generation of molecules. Molecules can
be represented in the form of string of characters, SMILES representation, which allows the implementation of transformer architectures. In this work, we propose a transformer decoder based network for the generation of molecules with high validity, uniqueness and novelty. The proposed model is capable of conditional generation where the condition can be based on a scaffold or/and multiple physicochemical properties. Moreover, we show that saliency maps can be used to make the generative process interpretable.
Description:
In this thesis, we propose LigGPT, a transformer decoder model for conditional molecular
generation. The main component of the model is the masked self-attention mechanism that
allows it to learn the SMILES grammar and long range dependencies very well. LigGPT
has comparable validity and uniqueness scores to other models on the MOSES dataset. It
outperforms other models in terms of novelty and internal diversity on the MOSES dataset.
The model performs better than other models on the GuacaMol dataset. Using saliency maps
we show that the generative process of model is interpretable. LigGPT is more efficient than
the famous character based recurrent neural network as is evident by training on only ten
percent of data.
Apart from unconditional generation, we show the ability of LigGPT to generate molecules
based on properties. Moreover, it can also be trained to retain the scaffold structure while
generating molecules having desired values of certain properties. This can have tremendous
applications in any sector which involves the creation of novel molecules. We even demon-
strate LigGPT’s usage in one shot lead optimization. Consequently, LigGPT is a strong
model and has the capability of making a positive impact on real world application for
molecular generation.
