APBN New Site

APBN Developing Site

10 Minutes to Screen 4,000 Drugs

Drug discovery is slow, expensive and cursed with high failure rates. As the COVID-19 outbreak screams for a cure to stop it, Deargen Inc. shows the world how to exploit artificial intelligence (AI) to expedite the process and increase chances of success.

by Shaun Tan Yi Jie

While COVID-19 continues to spread unabated, the number of people who have succumbed to the disease is well in the tens of thousands, multiple times that of the death toll registered by SARS in 2003. A vaccine or cure is paramount to stem the tide, but drug discovery is extremely slow and difficult: the SARS pandemic came and went without a treatment being found. Now, AI promises not to let history repeat itself.


E Pluribus Unum: Atazanavir

Researchers at Deargen Inc., a drug discovery company in South Korea, have published an article on the bioRxiv preprint server where they report using an AI model to screen for commercially available antiviral drugs that may act on COVID-19.

“We included all FDA-approved drugs […] including antiviral and other drugs, and approximately 4,000 drugs are included,” said Dr. Bo Ram Beck, lead author of the study. He told APBN that they chose to evaluate non-antiviral drugs as well because “we may find other drugs not approved for viral infection like anticancer drugs which may be useful to develop novel drugs to treat the coronavirus infection.” However, these drugs were not documented in the paper as “at least in vitro experiments are required to suggest antiviral effects of non-antiviral drugs.”

Dr. Beck’s team used their pre-trained deep learning AI model, Molecule Transformer-Drug Target Interaction (MT-DTI), to identify commercially available drugs that could attack COVID-19 proteins. “[T]he AI model predicts affinity between drugs and target proteins – how strongly the small molecule of interest will bind to the target protein,” he described. The model’s prediction is based on “available public data of pairs of drugs and target proteins with affinity values, dissociation constant Kd, and/or inhibitory concentration 50 (IC50)” gathered by the team.

Their results showed atazanavir, an antiretroviral medication used to treat and prevent the human immunodeficiency virus (HIV), to be most promising. Atazanavir was predicted to have binding affinity to multiple components of COVID-19, including 3C-like proteinase (Kd = 94.94 nM), RNA-dependent RNA polymerase (21.83 nM), helicase (25.92 nM), 3’ – to – 5’ exonuclease (82.36 nM), 2’-O-ribose methyltransferase (390 nM), and endoRNAse (50.32 nM), suggesting that all subunits of the virus replication complex may be inhibited simultaneously by atazanavir.


Baricitinib, Remdesivir: Pretenders to the Throne?

In contrast, British start-up Benevolent AI churned out baricitinib, a drug typically used to treat moderate and severe rheumatoid arthritis. When asked why the results are different, Dr. Beck replied that they were not familiar with Benevolent AI’s technology, “but my opinion is that approaches of two groups were unique to each other.” He listed three criteria for their work:

  1. FDA-approved antiviral drugs may be easier to accelerate to clinical trials as ‘experimental therapeutic options’ compared to drugs developed for other disease indication such as cancers;
  2. Toxicity profiles of other drugs like anticancer drugs might be a risk factor;
  3. Direct inhibition of the viral replication process may be the most effective mode of action to treat the infection post-entry.

“Therefore, we used viral replication-associated proteins as target proteins, and antivirals developed to inhibit viral proteinase like atazanavir or […] RNA/DNA replication like ganciclovir,” he rationalized.

Interestingly, Dr. Beck revealed exclusively to APBN that they also used their model to run calculations for remdesivir – a drug once unsuccessful in treating Ebola – now touted as one of the most promising candidates against COVID-19. The results, which were not included in the publication as the drug is not FDA-approved, actually showed remdesivir to have stronger binding affinity than atazanavir to four of six components of the virus: 3C-like proteinase (Kd = 113.13 nM), RNA-dependent RNA polymerase (20.17 nM), helicase (6.48 nM), 3’ – to – 5’ exonuclease (45.20 nM), 2’-O-ribose methyltransferase (134.39 nM), and endoRNAse (70.27 nM). This is a further boost to its burgeoning reputation, although he cautioned that remdesivir “should be scientifically investigated further in the future since our results are predictions through AI model.”


10 Minutes to Solve a Crisis

How did Deargen decide to participate in this crisis? Dr. Beck recalled that after the genome of the virus was released, “one of our members popped the question: ‘what can we do for the [COVID-19] outbreak?’”, which kickstarted the project. “The AI model was originally made to perform drug reposition and hit identification for our drug discovery pipelines, which means the model was built since the beginning of Deargen Inc., and it was ready to use,” he disclosed. “After ideation, the actual prediction process only took approximately 10 minutes including data input and raw prediction results.”

Dr. Beck provided an exclusive action log of the project to APBN:

  • 2020.01.28: Idea discussion – What can we do for the COVID-19 situation?
    • NCBI, COVID-19 genome sequences? Yes
    • NCBI, COVID-19 protein annotation? Yes
    • Deargen’s database has virus dataset? Yes
  • 2020.01.29: AI prediction and analysis + preprint manuscript initiation
    • Raw prediction results screened: Antivirals only, FDA approved, marketed
  • 2020.01.31: Preprint completed
  • 2020.02.02: Preprint approved & uploaded

Dr. Beck added that realization of the project was possible at this stage of the pandemic because their Bidirectional Encoder Representations from Transformers (BERT) framework-based model uses simplified molecular-input line-entry system (SMILES). “SMILES string is a 1D string to represent a chemical structure, thus atoms close to each other in 2D or 3D structure can be remotely expressed in SMILES in many cases,” he explained. This means it is possible to include target proteins that do not have experimentally confirmed 3D crystal structures in the computation, such as COVID-19 proteins. “Therefore, applying BERT can solve this issue.”

Dr. Beck is now “looking forward to develop our novel AI model to predict protein-protein interaction”, in order to design antibodies that can neutralize the virus. According to him, this is another possible approach to fight this war against the World Health Organization-dubbed “public enemy no. 1”. [APBN]

About the interviewee

Dr. Bo Ram Beck is Leader of the Drug Development Laboratory at Deargen Inc., an AI-driven drug development biotech venture based in South Korea. His fields of interest are immunology and microbiology.