- Konstantin Furs
- Hanna Karpenka
- Yury Kornoushenko
- Nikolaev Grigory
- Ivan Bosko
- Nikita Shuldov
- Artsemi Yushkevich
Combining modern computer modeling with generative machine learning models have played an ever-increasing role in the design of novel drug compounds. The use of it will significantly decrease the time for developing new drug candidates against SARS-CoV-2 and substantially reduce the average cost of bringing a new drug to market.
What it does
Computer-aided design of new bioactive compounds can be performed by virtual screening of molecular databases allowing one to find molecules with the required structural and pharmacophoric features. An essential role in structure-based drug discovery belongs to molecular docking that is widely used to predict the ligand conformation and its position and orientation within the binding site of target proteins, assess the binding affinity and investigate the interaction profile of drug candidates. The latest developments of semiempirical quantum mechanical and density functional theory methods as well as applications of explicit quantum mechanical calculations to structure-based drug design in the context of identification and optimization of drug candidates show the growing importance of quantum chemistry in the study of protein-ligand interaction. The use of machine learning methods makes it possible to find new compounds!
How we built it
We conduct a virtual screening with the Pharmit web service, using a pharmacophore model based on a protease bound to an X77 inhibitor (6w63) and pharmacophore model based on a spike protein bound to an ACE2 receptor (6M0J). As the third target, we use the form of the main protease unrelated to ligands. The best compounds selected during the virtual screening are subjected to molecular docking with previously indicated targets using the qvina software package with "exhautiveness" value of 20. Taking free binding energy obtained from previous step and converting structures to molecular fingerprints, we create 3 datasets, one for each of previously chosen targets. Then we train an adversarial autoencoder model using these datasets, and try to generate new fingerprints with a preset property of good binding energy. As part of the filtering process for new generated fingerprints we use our own deep regression model to predict binding energy from fingerprints, filtering “bad” compounds. Our submission will consist of best results obtained from docking, plus compounds derived from generated fingerprints, additionally validated by docking. After the completion of these stages, we plan to conduct molecular dynamics modeling on our best compounds with the help of the Amber software package.
Accomplishments that we're proud of
We pround of accomplishments of our team in computer modeling of new drugs against HIV-1 and oncology link .
What's next for Drug discovery for COVID-19
The further of this project proposes to use the predicted mimetic candidates as the scaffolds for computer-based generation of their analogs with improved biological activity and drug-like properties followed by synthesis and detailed biochemical assays.
Try It out
al, covid, drug, drug-design, ml