Researchers at Google’s DeepMind AI team have used AI to create advanced sorting algorithms, which although not specifically designed for drug discovery, could potentially benefit the field.
Published in Nature, DeepMind’s latest work demonstrates the use of deep reinforcement learning to create more efficient routines for sorting and hashing. These algorithms find use in various computational tasks, especially in computationally heavy processes such as drug discovery and simulations.
The researchers from DeepMind created an AI focused on code generation. To this end, they adapted the AlphaGo AI, a system known for defeating a human champion in the game of Go in 2016. The researchers created the AI system, known as AlphaDev, after staging a “game” approach, in which the AI treated a set of computer instructions like game moves. The AI then learned to “win” by sorting lists of three and five items as efficiently as possible. The resulting algorithm for the latter was 70% more efficient than current benchmarks.
While AlphaDev offers promise, DeepMind’s AlphaFold project has already begun to offer tangible benefits to drug discovery. Last year, we highlighted several ways the technology was finding use in life sciences.
AlphaFold’s impact on drug discovery and research
In an April conversation with Springer Nature, Pushmeet Kohli, the head of AI for Science at DeepMind, highlighted the potential of AlphaFold in protein folding to aid drug discovery. He stated, “That’s what we are seeing. It is supporting antibiotic resistance research. It is accelerating drug discovery.”
In recent years, AlphaFold has “emerged as the go-to tool in biology,” according to Kohli. He further explained that it enables “unrestricted access to every structure for both academic and commercial organizations, encompassing nearly 200 million proteins, which is essentially the entire protein universe. This has facilitated reaching every segment of the research community.”
Exploring potential drug discovery ramifications of AlphaDev
This Illustration shows what a sorting algorithm does. [Image courtesy of DeepMind]
Given the inherent complexity and logistical challenges involved, drug discovery efficiency tends to resist advances. In fact, a report in the Future Journal of Pharmaceutical Sciences noted that cost and timelines have substantially increased over the past three decades.
But the field of AI in drug discovery continues to advance. In 2022, MIT researchers announced EquiBind, a geometric deep-learning model that speeds the process of identifying drug candidates. In particular, EquiBind yielded a 1,200-fold increase in speed compared to QuickVina2-W, a prominent computational molecular docking model.
DeepMind’s AlphaFold also promises to help researchers answer common questions. As Kohli explained, “From a sequence you have the structure, but what other proteins this protein will bind to? Which small molecules or drug would this protein interact with? What would happen if there’s a mutation in this protein?” The company’s AI can help unravel such questions by predicting protein structures based on their amino acid sequence. “There is a lot of work that is happening on taking sort of the structure produced by AlphaFold, and then trying to predict the function of that protein,” Kohli noted.
On a related note, EquiBind demonstrated its prowess in a technique known as “blind docking,” allowing it to directly predict the optimal ligand-to-protein binding configuration without knowing the protein’s target pocket beforehand. This capacity could accelerate the process of identifying viable drug-like molecules for specific protein targets.
Other recent AI advances in drug discovery
In 2021, MIT also announced a machine learning system known as DeepBAR that slashed the time taken to calculate the binding affinities between drug candidates and their targets.
Such breakthroughs promise to help the industry chip away at the decade-plus timeframes often involved in commercializing new drugs. The pace of implementing AI into drug discovery processes, however, may be gradual, as UBS recently noted in a report on generative AI in particular.
In terms of the AlphaDev algorithm, its performance improvements in specific computational tasks like sorting are unlikely to directly translate to broader tasks such as bioinformatics, genomics and molecular modeling, which involve a variety of algorithms and models. “I don’t think this problem of predicting the function has been solved. It’s not solved at all, completely, but it is a very active area of research,” Kohli added.
Filed Under: Data science, Drug Discovery, Drug Discovery and Development, machine learning and AI
