Mobile Menu

Challenges for ML algorithms in COVID-19 drug screening

AI Cures, the newly-conceived contribution of MIT J-Clinic to COVID-19 research, aims to develop machine learning methods to discover antiviral molecules against the novel virus and other emerging pathogens. To this end, they summarise what screening data they have collected so far to train ML models to predict antiviral properties against SARS-CoV-2. The blog post also critically assesses the quality and utility of the data available.

Antiviral activity, they explain, is measured by the concentration of the molecule required to reduce viral activity by half, IC50. However, an effacious, toxic drug would be as good as useless, the other metric to be considered is CC50, the concentration of drug that kills half of human cells. The two inversely related values were the most commonly used scores in published papers.

Out of over 200 reported active compounds, 32 were identified as being active by more than one study, but only a single active compound so far was common to all – Remdesivir. Some studies found the much-hyped drug molecules Chloroquine and Hydroxychloroquine to be inactive.

The activity assessments of other candidate molecules were found to be highly contradictory and the author, having consulted colleagues, suggests several explanations for why this might be. It’s likely that while variable experimental techniques would have contributed to the divergence of results, it makes the initial findings for Remdesivir all the more promising.

The researchers next wanted to compare the screening results to the compounds currently in COVID-19 clinical trials. The surprising results highlighted that many of the drugs currently being trialled have yet to be studied experimentally and what’s more some have been shown to be inactive. The concerned author queries: “does [this] mean that clinical trials involve drugs that are not sufficiently tested? Are screening results not predictive of compound efficacy in patients?”

About AI Cures:

The COVID-19 pandemic highlights the acute need to develop fast, on-demand therapeutics against pathogens and health threats. Traditional approaches to drug development are expensive and too slow to react to pandemics like COVID-19. AI tools, on the other hand, have the potential to accelerate and transform this effort, enabling rapid, large scale search and identification of effective candidates for therapeutics. To translate this potential into success, we believe we must bring together researchers in computational and life sciences. While these communities are today closer than ever, we still sometimes speak different languages (coded or otherwise). The goal of this website is to lower the barrier for people from varied backgrounds to get involved and contribute to our shared, global goal.

  • For computational experts, we provide datasets, ways to frame ML problems, suggestions of possible ML tools to try, along with discussions of technical challenges.
  • For life science researchers, we offer the opportunity to bring in desperately needed domain expertise, posing questions for AI researchers to solve, and data to look at. Data specific to COVID-19 is still limited today but the situation is going to change substantially in the coming weeks and months.

More on these topics

Covid-19 / drug discovery / Machine Learning

Share this article