The Nobel Awards Season just ended, with the “Oscars of Science” awarded to some of the world’s brightest minds. The entire science world was watching, and just like with the Oscars, there was an element of suspense, drama, envy, celebration, and happiness. Most of the Nobel Laureates are also phenomenal speakers and communicators with decades of teaching experience, and thousands of people across the world are glued to their monitors to hear their inspiring stories. The Nobel Prizes are awarded in Physics, Chemistry, Physiology or Medicine, Literature, Peace, and Economic Sciences. Unfortunately, there is no Nobel Prize for Computer Science, Mathematics, or Engineering. So, it seems like it would be nearly impossible for an AI scientist to be awarded a Nobel. But what if that AI scientist used AI to make a significant advance in Physics, Chemistry, Physiology or Medicine, Literature, Peace, and Economic Sciences? Well, I know of one – Demis Hassabis of DeepMind. And I certainly do think that he deserves one. So is it possible?
The Rise of AlphaFold as a Standard Laboratory Tool And Applications in Drug Discovery
Serendipitously, on the 20th of September, 2023, I was having coffee with Dr. Michael Levitt. In 2013, Michael, along with Martin Karplus and Arieh Warshel, shared a Nobel Prize in Chemistry for “the development of multiscale models for complex chemical systems.” Their work laid the foundation for computer simulations that combine classical and quantum mechanical physics, which has been applied to various areas of chemistry and biochemistry, including the study of protein folding. Back in January, after many months in review, together with Dr. Alán Aspuru-Guzik and scientists from my company, Insilico Medicine, including me, Dr. Levitt published a paper titled “AlphaFold accelerates artificial intelligence powered drug discovery: efficient discovery of a novel CDK20 small molecule inhibitor” demonstrating that it is possible to use the AlphaFold-derived protein structure as a starting point to design and synthesize new molecules for hepatocellular carcinoma (HCC). In that paper, the team wanted to prove that AlphaFold may be used for drug discovery as one of the many elements in a comprehensive discovery pipeline. Two other generative platforms were used in this experiment, PandaOmics to identify new targets that are likely to work in HCC but did not have an experimental crystal structure, and Chemistry42, which can take a protein structure or a template molecule and design novel molecules with the desired properties. To cut the story short, this work demonstrated that the AlphaFold structure, with some modifications, could be used to identify new hit molecules that are very far from being called drugs but can effectively kill liver cancer cells. Of course, this was just a demo experiment to show that it is possible. In a real drug discovery program, where you are preparing to bet hundreds of millions of dollars and six to ten years of your life, you are very likely to spend several thousand dollars and a few weeks to get real experimental crystal structures instead of using a predicted structure. But we wanted to prove the point that AlphaFold could be used as a building block in a drug discovery pipeline. We also wanted to show that all the work, including two rounds of chemical synthesis under thirty days each, took around fifty days to complete.
Most likely, this is what Demis Hassabis referred to in his tweet “AlphaFold has been used to accelerate the design of a potential drug for liver cancer – by helping researchers identify a target in just 30 days. This is the first successful application of our AI to hit identification in drug discovery”. Unfortunately, there was no reference to the paper but we have not seen another hit for liver cancer out of AlphaFold. But I was very happy when Demis confirmed that it was indeed our study.
Another study using AlphaFold proteins for drug discovery titled “Discovery of novel and selective SIK2 inhibitors by the application of AlphaFold structures and generative models” did not use AI to identify and prioritize the target but used Chemistry42 to generate potent experimentally-validated hits. Besides drug discovery, AlphaFold was used by thousands of scientists globally for many other basic research tasks. Especially tasks where the crystal structure may be useful to test a hypothesis, but there is no budget to get the real crystal.
The AlphaFold paper published in July 2021 became very popular, and at the time of this writing, it was cited more than fourteen thousand times by Google Scholar (over nine thousand times counted by Nature) and got a whopping thirty-five hundred Altmetric points (~150th most popular paper in history of Nature). AlphaFold became the third highest-cited paper co-authored by Demis Hassabis, behind the famous Go and Atari Games papers.
During our walk with Dr Michael Levitt, I asked, “I think that AlphaFold got extremely popular. And now, Demis Hassabis founded Isomorphic Labs to go after drug discovery to focus on small molecules. And he formed the science advisory board comprised entirely of the Nobel Laureates, so it should be a slam dunk. Do you think AlphaFold as a consortium may get the Nobel Prize? It will be the first Nobel awarded to AI”.
Dr. Levitt responded without hesitation: “At a recent Nobel Foundation event in Seoul, I was asked on stage about AI someday winning the prize. I said that one of the hidden purpose of Nobel prizes was to create role models. Thus, it needs to go to people.”
I argued that there were many people working on the project with multiple advisors and contributors and this could turn out to be a consortium award. But the first and the last authors were John Jumper and Hassabis.
“Whether it goes to Jumper and Hassabis needs a careful study of the field. Chris Sander, who pioneered the idea of correlated amino acid mutations in an early paper of 1977, and John Moult, who pioneered the idea of blind prediction competition (CASP) in 1998, are also deserving”.
The day after our discussion, Demis Hassabis and his AlphaFold co-author, John Jumper, shared the Lasker Basic Medical Research Award. The Lasker Award usually precedes the Nobel Prize.
I also reached out to Demis to comment. And he mentioned that that Alphafold “took years of research and painstaking effort from many people” to create. “We’re thrilled to see the ways in which it’s being used by scientists around the world to impact everything from drug discovery to plastic eating enzymes,” he said.
And the impact of AlphaFold on the scientific community is indeed huge. Since AlphaFold’s 2021 release, more than half a million scientists have accessed its database, Hassabis said in the news briefing according to Science News.
The Important Role of PR and Media Management
Since its inception, DeepMind has been a well-rounded champion in the world of AI. They were great at everything. From pioneering deep reinforcement learning, to working with the most helpful and sophisticated investors, to academic publishing and working closely with the editors at top journals, to public relations and sensational performances – they kept very high standards without compromises. From the very early days, they managed to build a phenomenal PR and communications team, which serves as a very effective engine to popularize their work. In 2016, they did a compelling presentation of AlphaGo capabilities by televising a live competition with the world’s greatest Go master. The competition was broadcast all over the world, and in Korea, China, and Japan, millions of people were glued to the screen watching the game. It later resulted in the award-winning AlphaGo Movie, which has been viewed over 34 million times. This feat inspired a generation and drove the hype in AI, which in turn fuelled even more advances that are enjoying today. The release of AlphaFold was also accompanied by a popular internally-produced documentary video, which also showed the emotional component of its development.
Sophistication in AI and growing popularity enabled DeepMind to publish multiple research papers in Nature – many purely algorithmic papers without additional laboratory validation. A feat that is very difficult to achieve. Experimental validation may take years and the resulting structures or molecules may still be of lower quality than the structures professional medicinal and computational chemists working in biotechnology are used to. Adding experimental validation always complicates the peer-review process as industrial experimental scientists rarely pay attention to AI and want to see experimental evidence that would cost tens or even hundreds of millions to generate. Publishing a purely algorithmic paper in Nature is extremely difficult and before DeepMind many considered it impossible. However, walking proximity to Nature and a solid record of success certainly make the process easier. And every such paper generates massive attention, inspiring many more people to pursue careers in science.
There have been only a few precedents where Nobel Prizes were awarded to individuals working for commercial companies. But this year’s Nobel Prize was awarded to Hungarian scientist Katalin Kariko, a BioNTech veteran. And DeepMind has a smooth process for conceiving, implementing, publishing, and promoting academic and commercial research and making it available to the broad community that surpasses most academic institutions by a broad margin.
Prediction – The Nobel Prize in Chemistry Within A Decade
Unsurprisingly, this year, Jumper and Hassabis were the favorite among the suggested laureates according to Chemistry Views.
AlphaFold is one of the most popular AI projects on the planet. However, since the time of its launch, it has not yet resulted in the creation of a drug that could go into human clinical trials. It became a very useful tool in many labs all over the world. It helped thousands of scientists perform their computational or even laboratory experiments and may be considered for a Nobel Prize by itself. However, it is very likely that in a decade or sooner, with enough budget and luck, it will result in a therapeutic to treat a rare or even a broad disease. And that achievement would certainly help AlphaFold’s creators cross the finish line. It is very likely that one or more of the academic labs or companies working in the field including Google’s most recent startup Isomorphic Labs, may reach this important milestone.
The prize before any drug designed using the predicted structure completes or even enters human clinical trials. AlphaFold became a very popular tool in the academic community and scientists often use the predicted structures in their academic work and to augment their academic publications. To make this article more balanced, I should mention that many of the predicted structures are not perfect with multiple studies explaining the limitations. One of the top voices in drug discovery and “AI realist”, medicinal chemistry veteran Dr. Derek Lowe writing for Science In The Pipeline summarized some of these in his August 2023 post titled “Docking With AlphaFold Structures: Oops”. Therefore, scientists should exercise caution when over relying on the structures predicted using AlphaFold or any other computational algorithm when publishing their papers. The avalanche of these papers may possibly lead to erroneous results in the long term. Whenever funding is available, it is a good idea to perform an experiment to validate the results. Especially in drug discovery, where such experiments are only a start of a very expensive, long, risky, and difficult journey.
I also asked Dr. Roger Kornberg, the 2006 Nobel Laureate in Chemistry, about the possibility of the Nobel Prize awarded for AlphaFold. Dr. Kornberg responded that “while AlphaFold is derivative of a generation of protein crystallography, it represents a major advance in chemical science. Justification in terms of drug development is not necessary for the highest recognition. So if Hassibis is the clear originator and driving force, then the Nobel to him alone or with an essential collaborator(s) will likely follow”.
Demis Hassabis is a legend in many fields and many scientists and entrepreneurs around the world inspired by his work – including the author of this article – and we would celebrate such an achievement. If a Nobel Prize for AI can be ever be awarded, Demis Hassabis is certainly the most deserving human.