AlphaFold - The AI Leader in Protein Structure Prediction

In this final installment of our AlphaFold series, we will tell the story of how DeepMind, a research developer in the AI community, leveraged its AlphaFold system to play a vital role in the fight against COVID-19.

How AlphaFold joined the fight against COVID-19

AlphaFold, an AI system developed by DeepMInd, surprised the world when it crushed the competition at CASP13 in the summer of 2018 (1). Two years later, AlphaFold is still the leader in protein structure prediction and facing its biggest test yet, helping to combat a deadly pandemic that has claimed over 2.5 million lives (2).

As COVID-19 swept the globe in the spring of 2020, the scientific community got to work. Epidemiologists, virologists, and structural biologists were joined by computer scientists in a concerted effort that has resulted in thousands of published papers and an influx of new data. The AlphaFold team wanted to contribute to the pandemic response by learning more about the structures of lesser known COVID-19 proteins (3). They hoped the information would improve understanding of how the virus functions and offer a starting point for new studies to build off (3). Solving the structure of viral proteins associated with COVID-19 gives research teams and drug companies an even larger arsenal of therapeutic options to try. In fact, one of the structures predicted by AlphaFold was later used in a study that looked for cavities in COVID-19 proteins that could be a good target for new drugs (4).

Promising results and key support from peers

The usual method for solving protein structures by experimental methods can take months or years. Fortunately, a team based at the University of Texas at Austin was able to quickly elucidate the structure of the infamous spike protein and published their results in Science, March 2020 (5), which served as the comparative basis for AlphaFold’s AI prediction of the spike protein; the result was accurate (3). There was debate over whether COVID-19 structures determined by AlphaFold should be released before peer review and while the system was still under development; however, promising results and support from leaders at the Francis Crick Institute encouraged the team at DeepMind to make structures freely available to the public (6).

Understudied but not insignificant

One of the reasons that AlphaFold is more powerful than other computational structure predictors is that it doesn’t need a template protein with a similar shape (3). Using this so-called “free-modeling” approach, AlphaFold took on the structures of understudied COVID-19 proteins, which lacked available template structures. They quickly predicted structures for six COVID-19 proteins (SARS-CoV-2 membrane protein, Nsp2, Nsp4, Nsp6, Papain-like proteinase (C terminal domain), and ORF3a) and released the first results in March 2020 (3). Confidence scores were added to different areas of the protein to show where the system was the most likely to be correct. Since the initial release, the structure of ORF3a was experimentally solved by a team at UC Berkeley (7), and the structure for Nsp6 has been used in studies looking for promising drug candidates (3,7). The protein structures were updated by the latest version of AlphaFold in August 2020 (7, 3).

And the fight goes on…

A key takeaway of this experience is the ability for an AI system to make accurate conclusions about protein structure without using a previously determined template (3). This opens up a new frontier in molecular bioscience and protein engineering. As part of their commitment to fighting the pandemic, DeepMind is updating and publishing new structures on their website. AlphaFold proved its mettle against COVID-19 and continues to play an instrumental role encouraging a more holistic understanding of the COVID-19 virus.

References

  1. Senior, A.W., Evans, R., Jumper, J. et al. Improved protein structure prediction using potentials from deep learning. Nature 577, 706–710 (2020). https://doi.org/10.1038/s41586-019-1923-7
  2. Mapping the Coronavirus Outbreak Across the World. Bloomberg . com. (2021). Retrieved from https://www.bloomberg.com/graphics/2020-coronavirus-cases-world-map/
  3. Jumper, J., Tunyasuvanakool, K., Kohli, P., et al. Computational predictions of protein structures associated with COVID-19, Version 3. DeepMind website (2020). Retrieved from https://deepmind.com/research/open-source/computational-predictions-of-protein-structures-associated-with-COVID-19
  4. Gervasoni, S., Vistoli, G., Talarico, C. et al. A Comprehensive Mapping of the Druggable Cavities within the SARS-CoV-2 Therapeutically Relevant Proteins by Combining Pocket and Docking Searches as Implemented in Pockets 2.0. Int. J. Mol. Sci. 21(14), 5152 (2020). https://doi.org/10.3390/ijms21145152
  5. Wrapp, D., Wang, N., Corbett, K. S., et al. Cryo-EM structure of the 2019-nCoV spike in the prefusion conformation. Science 367, 1260-1263 (2020). https://doi.org/10.1126/science.abb2507
  6. Crick scientists support DeepMind’s effort to make COVID-19 data available. The Francis Crick Institute. (2020). Retrieved from https://www.crick.ac.uk/news/2020-03-05_crick-scientists-support-deepminds-effort-to-make-covid-19-data-available
  7. Kern, D. M., Sorum, B., Hoel, C. M. et al. Cryo-EM structure of the SARS-CoV-2 3a ion channel in lipid nanodiscs. bioRxiv. (2020). https://doi.org/10.1101/2020.06.17.156554
  8. Pandey, P., Prasad, K., Prakash, A., Kumar, V., Insights into the biased activity of dextomethorphan and haloperidol towards SARS-CoV-2 NSP6: in silico binding mechanistic analysis. J. Mol. Med. 98, 1659 - 1673 (2020). https://doi.org/10.1007/s00109-020-01980-1