Newly developed AI-powered techniques could dramatically speed up the discovery of drugs and advanced materials, allowing scientists to design molecules that target chemically useful properties in minutes instead of years.
New artificial intelligence technologies can greatly accelerate the discovery of new drugs and materials, potentially cutting years off processes that traditionally required time-consuming trial and error.
Research from the University of Florida and New York University outlines a method that can generate promising molecular candidates about 10 times faster than existing AI approaches, without sacrificing accuracy or chemical realism. This research highlights how generative AI is beginning to reshape the foundations of chemistry and materials science.
From educated guesses to guided design
All modern medicine and advanced materials begin with a hypothesis, a proposed arrangement of atoms that might kill bacteria, store energy, or interact efficiently with light. With billions of small molecules to choose from, identifying the right one has long been an almost impossible problem.
Rather than modifying known compounds, AI models can invent entirely new molecular structures based on desired properties.
Recent advances in generative AI have helped narrow that search. Rather than modifying known compounds, AI models can invent entirely new molecular structures based on desired properties. A newly developed method called PropMolFlow (property-guided molecular flow) takes this approach even further by significantly increasing the rate of generation of viable candidates.
“For most of the history of science, the discovery of matter often preceded its understanding. Useful compounds were discovered by chance, and scientists figured out why they worked,” says Stefano Martiniani, assistant professor of physics, chemistry, mathematics, and neuroscience at New York University and author of the paper. “Generative AI offers the possibility to reverse this: specify properties and then find structure. PropMolFlow represents another step in bringing that vision to life.”
Design molecules backwards
Researchers describe molecular design as an “inverse problem.” Scientists are rarely interested in molecules themselves. Instead, you need something that performs a specific function.
“Chemists usually don’t look for ‘molecules,'” Martiniani explains. “Instead, they want molecules that do something specific: interact strongly with light for optical applications, or have a specific electronic structure that determines how they absorb energy or conduct electricity.”
PropMolFlow is built on earlier AI models inspired by image generation tools such as DALL-E, which was first adapted for molecular design in 2022.
PropMolFlow is built on earlier AI models inspired by image generation tools such as DALL-E, which was first adapted for molecular design in 2022. Previous approaches have improved accuracy and chemical validity, but often require thousands of computational steps. PropMolFlow achieves similar or better results in approximately 100 steps.
“For a field where computational speed directly translates into discovery speed, this represents a meaningful advance,” added Mingji Liu, assistant professor in the Department of Chemistry at the University of Florida and one of the paper’s authors. “This work does not replace what came before, but rather demonstrates that the next generation of molecular generators can be significantly faster while maintaining the precision that makes these tools useful.”
New AI models can design molecules with specified properties 10 times faster than previous methods, potentially speeding up the process of creating drugs and materials. This figure shows how the system converts random noise into a complete molecular structure based on the properties of the target. (Image courtesy of University of Florida and New York University)
Accuracy without shortcuts
Speed alone is not enough when the molecules produced break basic chemical rules. The team then tested PropMolFlow against the established model and found that it produced chemically valid structures over 90% of the time.
Speed alone is not enough when the molecules produced break basic chemical rules.
“This is important because many previous approaches produced structures that looked superficially plausible but violated fundamental chemical rules,” Martiniani says.
To avoid the risk of the AI grading homework on its own, the researchers verified their results using density functional theory, a physics-based method that does not rely on machine learning.
“This type of validation provides the confidence needed for the generated molecules to be seriously considered in real-world applications,” Liu says.
Implications for discovery
Researchers believe the combination of speed, precision, and rigorous validation has the potential to revolutionize early-stage molecular discovery.
“Candidates targeting thousands of chemically valid properties can be generated in minutes instead of hours, allowing researchers to iterate faster,” Martiniani explains.
Although extending the approach to larger and more complex molecules remains a challenge, the principles behind PropMolFlow provide a “template for more ambitious applications” in drug development and advanced materials research, Liu said.
