Sensationalist headlines are multiplying on the capacity of artificial intelligence to revolutionize scientific discoveries. Microsoft announced on January 9 that a team of researchers used artificial intelligence and high-performance computing, approximately 5,000 classical computers, to analyze 33 million possible battery materials in eighty hours. A task that, according to the team, would have taken twenty years previously. Specifically, the model traveled through the periodic table and deposited new atoms in various locations through a substitution process. About a hundred of the most promising elements were transferred to the Pacific Northwest National Laboratory, which specializes in battery production, which built a prototype using an entirely new material using 70% less lithium. But this prototype, as promising as it is, is not close to being commercialized.
A few weeks earlier, DeepMind, Google’s AI unit, had published an article with great fanfare in Nature relating to the discovery of 2.2 million new crystals, the equivalent of nearly eight hundred years of knowledge, according to them. The Graph Networks for Materials Exploration (Gnome) model, a deep learning tool, reportedly made 2.2 million predictions, of which the most stable 380,000 would provide promising candidates for experimental synthesis. DeepMind even synthesized 41 of these new compounds. A preliminary study having analyzed the latter nevertheless raises questions about the scope of the discovery. These synthesized compounds could actually be already known materials, but with atoms arranged in a more random way.
All these clues suggest that AI models still have difficulty modeling disorder. This preliminary study has not yet been peer-reviewed, but it highlights the need for the scientific method to accurately gauge the findings of artificial intelligence. The same goes for innovative medicines. The hopes resting on AI are immense.
New drugs thanks to AI?
Demis Hassabis, co-founder of DeepMind, says his subsidiary Isomorphic Labs has the ambition to more than halve the time needed to research new molecules, from five to two years. This announcement comes days after Isomorphic Labs’ first two pharmaceutical partnerships, with Lilly and Novartis, worth a combined $3 billion. These agreements follow several others in the industry: Exscientia, based in Oxford, works with giants Sanofi and Bristol Myers Squibb among others; Owkin and Aqemia also collaborate with the first, while Insitro has joined forces with the second.
Isomorphic Labs’ platform builds on the scientific advances made by DeepMind’s AlphaFold technology, an AI capable of predicting the structure of almost any existing protein from the sequencing of their DNA. These proteins can fold in countless ways, with their final shape determining how they function and how they interact with other things.
AlphaFold is impressive, having predicted the shape of over 200 million proteins. But this tour de force must comply with the reality of biology. In 2022, a American research team had studied the ability of AlphaFold to model the interactions of essential proteins of the bacteria Escherichia coli with antibacterial compounds. It appeared disappointing, even if this first study also pointed out ways to improve it. Artificial intelligence provides a new generation of approximate tools that allow us to solve problems that previously seemed intractable, and that’s deeply exciting. But be careful not to overestimate the capabilities of this technology. The advent of the future revolutionary material, or drug, will still largely depend on human intelligence and intuition.
.