jcatapang
Can anyone share on this thread any similar attempts to this submission? We used deep learning (autoencoders) to generate our drug designs.
Can anyone share on this thread any similar attempts to this submission? We used deep learning (autoencoders) to generate our drug designs.
Thank you for your observation. Additionally, we actually have seven drugs but I only submitted the ones for the Mpro (3 drugs). The other drugs are ACE2 receptor blockers, which are found in the paper.
First one looks ok apart from the misplaced double bond (might be some form of submission mistake). Second one looks “exotic” to me.
@pgz Regarding your point on the misplaced double bond, is it possible that this arises because @jcatapang’s team used deep learning to generate the drug designs?
1 reply
@Zhang-He, you are correct. We can’t control that behavior as the deep learning algorithm is the one that designs all the structures. It can be part of our future work to try and incorporate these rules into the algorithm somehow, so that the designs would conform to what’s stabler. As of the moment, the code doesn’t take into account stability. Apologies.
Speaking of AI-generated drugs, you guys might want to check out IBM’s exploration tool. COVID-19 Bluemix It includes drug attributes like Target Affinity (AFF), Drug-likeliness (QED), Synthetic Accessibility (SA), Solubility (LogP), Novelty (NOV), Selectivity (SEL), Toxicity (TOX), and Molecular Weight (MolW).
1 reply
Thanks for this @jcatapang. This is amazing!
Just here to share. This AI-based research for COVID Moonshot has also birthed a new commercial software I’ve developed. Deep Drug Search has four main components: ligand generation via deep learning, ligand-protein docking, ADME-Tox prediction via machine learning, and a molecular viewer. Cheers 