The chemical space is enormous and it is not feasible to first synthesize molecules and look if they have the desired properties. So, we propose transformer decoder based conditional generative model that can efficiently generate molecules having the desired physico-chemical properties. We call it LigGPT. We first show state-of-the-art results on unconditional molecule generation, followed by conditional generation results based on QED, logP, SAS and TPSA. We show that our model is capable of controlling multiple properties at the same time and moreover, it can also maintain the underlying scaffold structure. LigGPT is also capable of performing one-shot property optimization on a given start molecule. Such model can act as catalyst in the drug discovery process and has enormous other applications in drug discovery.
Viraj Bagal
Rishal Aggarwal
Vinod P.K.
Deva Priyakumar U.
Generative networks, optimization.
Algorithm.
Technology designed and implemented.
AI in Healthcare.