McGill’s Office of Research and Innovation invited two recent recipients of the Canada Gairdner International Award to present their research to the McGill community. Demis Hassabis, CEO and co-founder of Google DeepMind, and Lynne Maquat, J. Lowell Orbison Endowed Chair and professor in the University of Rochester’s Department of Biochemistry and Biophysics, spoke to students and faculty about recent biomedical advancements. 

Hassabis, a 2023 recipient, gave the first lecture, focusing mostly on his work with DeepMind’s AlphaFold project, which uses artificial intelligence (AI) to understand how proteins fold. His lecture began with a brief history of DeepMind and the achievements that preceded AlphaFold. Hassabis co-founded DeepMind—now owned by Google—in 2010. Since the company’s inception, AI research and capabilities have exploded. 

Before AlphaFold, DeepMind made AI history with AlphaGo, a project in which the company’s AI system beat a master Go player in a highly publicized game that had a one-million-dollar bet riding on it. During his presentation, Hassabis explained that the sheer number of possible combinations of moves in the game made it so difficult. 

“One easy way to see how complex Go is is that there are 10 to the power 170 possible positions in Go, which is far more than there are atoms in the universe,” Hassabis said. “So we had to come up with systems that were much cleverer that learned about the structure of Go and learned heuristics about Go so that we could make the search tractable.”

DeepMind initially introduced AlphaGo to ‘strong amateur’ Go players to learn the game. Subsequently, they honed the model’s skills by playing it against increasingly better versions of itself in a process called reinforcement learning. Hassabis explained that the projects to build AI that could master complex games were a means to develop technologies that scientists could then apply to more socially-relevant problems.

“Games were just a means to an end,” Hassabis said. “We wanted to develop these ideas, but we wanted them to be very general, so that eventually, once they got powerful and sophisticated enough, we could transfer them to work on real-world challenges. And specifically, my passion was to apply them to scientific problems.”

The next step for Hassabis was to solve the famous protein folding problem: How a protein’s amino-acid sequence determines its 3D structure. Proteins are essential building blocks in biology and their structures can determine their functions. Understanding how even a single protein folds from just its amino-acid sequence, however, can take years of work. In 2020, AlphaFold 2 won the Critical Assessment of Structure Prediction (CASP)—a contest in which teams of researchers compete to predict protein structures. Not only did DeepMind’s model win, but its result fell within an atomic threshold of accuracy, leading the CASP organizers to designate the problem as solved. Today, the AlphaFold database contains over 200 million protein structures predicted by the model. 

In 2021, Hassabis founded Isomorphic Labs, a start-up and sister company of DeepMind, focused on using AI for drug discovery. 

To conclude his lecture, Hassabis turned to ethical questions surrounding AI. While he sees immense potential for AI to benefit scientific discovery and society, he cautioned against a “move fast and break things” attitude. 

“This is too important to work in that way,” Hassabis explained. “I think we should instead use the scientific method to try and plan ahead of time and do controlled experiments and get a better understanding of what [an AI model] is before we deploy it around the world. So I think transformative technology like [Artificial General Intelligence] requires exceptional care, and what we’re trying to do at Google DeepMind is to be both bold and responsible with the technology.”

Once Hassabis ended his talk, Maquat, a 2015 Gairdner laureate, took the stage to present her pioneering discovery of nonsense-mediated messenger ribonucleic acid (mRNA) decay (NMD). This cellular mechanism degrades abnormal mRNA to regulate cell function in both healthy and pathological conditions in humans.

NMD operates through two primary pathways involving exons—the coding regions of genes located either upstream, closer to the gene’s beginning, or downstream, closer to the gene’s end. The first pathway is reserved for newly synthesized mRNAs that end at a premature termination codon (PTC): A sequence of three RNA building blocks. This process also involves exon junction complexes (EJC)—protein complexes deposited upstream of exon-exon junctions in newly spliced mRNAs. If NMD recognizes a PTC while some EJCs remain downstream, the mRNA undergoes degradation at both the 5′ and 3′ ends. This pathway predominantly targets newly made mRNAs.

The second pathway is distinct, terminating translation upstream of a 3′ untranslated region (UTR). Unlike the first pathway, this one also applies to mRNAs in the steady state in addition to newly made mRNAs. 

Maquat’s lab first described NMD in the context of a condition called mRNA-deficient beta zero thalassemias in 1981.

“What we found were unexpected links between RNA metabolism, the nucleus, and the cytoplasm. These links, we showed, are [EJCs],” Maquat said. “The newly made mRNA is poised and ready to undergo NMD should a PTC be recognized in a downstream EJC.”

Following this observation, Maquat and her colleagues wondered how the cell distinguishes between translation termination codons that trigger NMD and those that do not.

“We came up with a surprising result and that was that the answer was where the introns resided in the pre-mRNA, which was confusing to a lot of people because the introns are gone from […] the fully spliced mRNA that is then targeted for decay,” Maquat explained. “And so we proposed that splicing in the nucleus must deposit a mark on newly made mRNAs that persist until the first round of translation.”

The biggest surprise arose when the researchers discovered that NMD targets newly made mRNAs that maintain their association with the nucleus. This challenged conventional wisdom and revealed a new facet of this quality control mechanism.

“We were able to prove that NMD targets newly made mRNAs on the cytoplasmic side of the nuclear envelope,” Maquat explained during her presentation.

In addition to sharing her groundbreaking discovery of NMD and its crucial role in maintaining the integrity of gene expression and proper cellular functioning, Maquat also shared valuable advice for dealing with the pushback that she received when other scientists in her entourage may have doubted the significance of her research.

“I have to say it was really scary for me but I couldn’t think of another explanation for the data. And I think when one gets data that are controversial, it’s really important not to overinterpret the data,” Maquat said.

She also described the specific role of NMD that eliminates mutated mRNAs. This can result in dominantly inherited diseases in which a single abnormal gene from one parent is sufficient for the disease to manifest.

In their exploration of other NMD factors, Maquat’s lab started studying Fragile X syndrome (FXS). 

“[FXS] is the most common single gene cause of intellectual disability and autism. It affects one in 4000 males [and] one in 6000 to 8000 females,” Maquat explained. 

FXS is caused by the loss of Fragile X mental retardation protein (FMRP), an RNA-binding protein and translational repressor. Maquat’s lab has shown that in FXS, the absence of FMRP causes global NMD hyperactivation. This leads to inefficient neural differentiation—a process that allows unspecialized cells to turn into neurons—and synapse maturation during which the connections between neurons become more efficient.

After the lectures, Hassabis and Maquat sat down for a Q&A period. As AI is increasingly being applied to biomedical research, Maquat mentioned the potential intersection between Hassabis’ research and her own in response to what she considers one of the most challenging problems that science hopes to solve.

“I think the hard problem is figuring out the networks, and there’s competition within the networks, and they’re gonna change during cell differentiation and development,” Maquat said, addressing Hassabis. “So that is a very difficult problem, and you’re approaching it fortunately, as well as ourselves, and hopefully the two will connect.”