Chinese scientists make ‘impossible’ AI breakthrough in drug research
- The team says its machine learning model can accurately predict the complex molecular changes that determine a drug’s effectiveness
- Structural biologists, including China’s Yan Ning, regard the task as beyond the capabilities of artificial intelligence
The breakthrough could speed up the preclinical drug development process, according to the researchers, led by Professor Li Ziqing with Westlake University’s school of engineering. Part of their work was published by peer-reviewed journal Advanced Science in October.
Protein molecules – the body’s building blocks – are made up of long chains of amino acids. The three-dimensional structure of the chains causes reversible changes to the protein, depending on the required biological function, and these alternative structures are referred to as conformations.
Structural biologist Yan – who returned home last month after resigning from Princeton University – fuelled debate over AI’s limitations in the field in one of her first public appearances since her return.
Speaking at a forum hosted by the Southern University of Science and Technology in Shenzhen on November 27, Yan highlighted the technology’s limitations in understanding protein conformation, an important yet difficult aspect in pharmaceutical development.
But, according to Li and his team, their AI model ProtMD overcomes the problem and can accurately predict which conformations proteins will form in different physiological environments.
ProtMD’s algorithm can calculate the movements of a protein at the atomic level, generating data based on molecular dynamics. With its different calculation logic, ProtMD is more generalised than other methods – including AlphaFold – but its performance level is the same for unknown structure pairs, the researchers said.
“This model cannot only predict the protein’s upcoming conformation according to its previous state, it can also predict the conformational change after interaction with the drug molecules, thus giving us the chance to evaluate the drug’s effect,” Li said.
The ability to more accurately predict drug-protein affinity in advance of the lengthy clinical trial process would save pharmaceutical developers some of their biggest costs and enshrine AI’s effectiveness in the design of new drugs.
But while traditional computer simulations have accelerated the discovery stage, they have been incapable of the fast virtual screening needed to make accurate predictions of how drug-like molecules interact with target proteins.
DeepMind AI breakthrough helps solve how diseases invade cells
Both computational and structural biologists have made many attempts to resolve the problem but have been stymied by the highly dynamic and time-dependent process that takes place when a receptor changes to accommodate a small molecule.
“This study is the first step in using AI methods to analyse the dynamic conformation of proteins,” Li said.
“The lightweight version of ProtMD has outperformed the state-of-the-art model, and its industrial-grade version can further improve the efficiency of drug affinity prediction and virtual screening,” he said.
China-US divide extends to AI, both in research and real life: study
Dr Niu Zhangming, co-author of the paper and CEO of drug development company MindRank AI, said ProtMD provides “a new direction for the development of machine learning models around protein”.
“This method achieves a breakthrough in the underlying principle and processes, the highest prediction accuracy among similar methods, Niu said.
