ELLIS against Covid-19

Screening for SARS-CoV inhibitors

ellis Guenter Klambauer 01 April 2020 - 01 April 2020
01 April 2020 • 14:55 - 13:10

Guenter Klambauer (Johannes Kepler University Linz)

Dr. Klambauer talked about the drug discovery efforts at the ELLIS group he leads in Linz. The drug discovery approach is to find small molecules that can inhibit some of the machineries of the virus that are needed for its replication and function. COVID-19 virus has a handful of proteins that can be targeted by a molecule found using a drug discovery approach. Machine learning can contribute to the drug discovery process in various stages such as compound design and virtual screening. In compound design, methods like variational autoencoder, generative adversarial networks, reinforcement learning, and recurrent neural networks have been used to build a molecule from representations of atoms in a proper space. Virtual screening refers to using computers to search through a database of available molecules to find the promising ones for inhibiting the virus function. Drug discovery can normally take many years to find an FDA approved drug especially due to the time needed for screening through compounds and preclinical trials. Machine learning can facilitate these steps and find those molecules which have a better chance to pass clinical trials with less side effects. There is also a drug discovery challenge organized by JEDI called “a billion molecules against COVID-19” to find a list of the lead compounds by screening a billion molecules and narrowing the search down to find an immediately usable molecule which is already FDA approved.

Video:

Question & Answers

Link to the recording of the live Questions & Discussion session for this talk. 

  • Q: Can machine learning help predict how existing drug can help treat COVID19? If yes, how to choose positive and negative standard for learning? 

    • A: Yes, can help. We use a rigorous procedure called Cluster-Cross-Validation (Mayr, 2018) to evaluate predictive quality.

  • Q: Why not mixing the virtual screening approach with drug discovery? e.g., generate a new compound and then try to match it with something already existent?

    • A: This is actually done in RL approaches, where a target prediction model is used as reward function (see e.g. Popova et al, 2018)

  • Q: Can we directly find compounds that affect specific viruses, given genomic information? E.g, give the compounds given RNA of covid19 that can destroy the lipid membrane. For example, predict the protein folding via protein folding prediction, and then use it to find the chemical compounds. Is this possible at the moment? If so, any research you can point to?

    • A: Yes, good point. Maybe check "proteochemometric modelling" -- this is the area of chemoinformatics that tries to learn across proteins.

  • Q: Did hydroxy-chloroquine come up in your drug screen, as this seems to be somewhat effective for treating COVID patients?

    • A: No, was predicted negative, but this means only that it does not act through protease-inhibition.

  • Q: Have you ever try to screen for inhibitors of corona viral RDRP (RNA dependent RNA polymerase?

    • A: Good idea! We have models for that... data is lying around and not used at the moment due to time constraints.

  • Q: Have you used drugs that have been shown to work in some way, such as hydroxychloroquine, to approximate the discovery of targets for action or drugs with similar characteristics in your drug prediction models?

    • A: Yes, we see some drugs in the top of the list that have known antiviral activity. Hydrochlorine appears inactive but that only means it does not work via protease inhibition

  • Q: Did your drug prediction experiments yield Remdesivir as a possible drug for the COVID-19?

    • A: No, but this only means it does not work via protease-inhibition.

  • Q: Are the drug prediction models (such as the target prediction) biased towards cancer drugs?

    • A: Those particular models not (at least not directly), but maybe DrugBank.