Designing useful molecules has long been a process of educated guessing. Chemists typically start with a known structure, modify it, test its properties, and repeat the cycle until something promising emerges. That approach has delivered many of the drugs and materials used today, but it is slow and constrained by what is already known. A new computational method developed by researchers at New York University and the University of Florida aims to invert this workflow by starting from desired properties and working backward to the molecular structures that could produce them.
Zeng, C., Jin, J., Ambrose, C., Karypis, G., Transtrum, M., Tadmor, E. B., Hennig, R. G., Roitberg, A., Martiniani, S., & Liu, M. (2026). PropMolFlow: property-guided molecule generation with geometry-complete flow matching. Nature Computational Science. https://doi.org/10.1038/s43588-025-00946-y
The method, called PropMolFlow, is described in a recent study published in Nature Computational Science. Led by Stefano Martiniani at New York University and Mingjie Liu at the University of Florida, the work builds on recent advances in generative artificial intelligence for chemistry. Rather than searching large libraries of existing molecules or making incremental changes to known compounds, the model generates entirely new molecular candidates that are guided directly by target properties such as electronic structure, light absorption, or binding behavior.
Stefano Martiniani at New York University stated,
“With the ability to generate thousands of chemically valid, property-targeted candidates in minutes rather than hours, researchers can iterate faster: generate candidates, filter computationally, validate the best ones with physics or experiments, and feed results back to improve the next round.”
At the core of the approach is the idea that molecule design is an inverse problem. Researchers are rarely interested in molecules for their own sake. They are looking for specific functions, whether that means inhibiting a biological pathway, storing energy in a battery, or absorbing light in a solar cell. PropMolFlow encodes these functional goals into the generation process itself, allowing the algorithm to propose structures that are statistically more likely to meet the specified criteria.
Compared with earlier generative models, PropMolFlow emphasizes efficiency as well as accuracy. The researchers report that the method can produce viable molecular candidates roughly ten times faster than comparable approaches. This improvement comes from a more direct transformation of random noise into chemically valid structures, reducing the number of computational steps needed to arrive at a complete molecule. In practical terms, what once required thousands of iterations can now be achieved in a few hundred.
Speed alone, however, is not sufficient if the generated molecules violate basic chemical rules. To address this, the team evaluated the structural validity of the outputs, examining whether bonding patterns and geometries were chemically reasonable. In these tests, the model produced valid structures in the large majority of cases, outperforming several baseline methods. It also maintained competitive accuracy in matching the target properties used to guide generation.
A recurring concern in AI-driven molecular design is how results are validated. When one neural network proposes a molecule and another predicts its properties, there is a risk that both systems share the same biases. To reduce this problem, the researchers compared the AI’s predictions with calculations from density functional theory, a physics-based method grounded in quantum mechanics. The close agreement between the two provided an independent check that the generated molecules were not only fast to compute but also physically meaningful.
The study involved contributions from researchers at the University of Minnesota and Brigham Young University, reflecting the growing interdisciplinary nature of computational chemistry. The authors note that while their demonstrations focused on relatively small molecules, the underlying principles are applicable to more complex systems. Extending these methods to larger drugs or advanced materials remains an active area of research, but the framework offers a template for doing so.
From an engineering perspective, the implications are tied to iteration speed. Faster molecular generation allows researchers to explore broader design spaces, filter candidates computationally, and reserve expensive experiments for the most promising options. This is particularly relevant in fields such as drug discovery and materials engineering, where development timelines are often measured in years and costs escalate quickly.
PropMolFlow does not replace existing tools but adds another layer to the molecular design pipeline. By integrating property targeting, chemical validity, and independent physical validation, the method demonstrates how generative AI can move closer to practical deployment in laboratory and industrial settings. As computational models continue to influence how molecules are designed, approaches that combine speed with careful validation are likely to shape the next phase of discovery.
Adrian graduated with a Masters Degree (1st Class Honours) in Chemical Engineering from Chester University along with Harris. His master’s research aimed to develop a standardadised clean water oxygenation transfer procedure to test bubble diffusers that are currently used in the wastewater industry commercial market. He has also undergone placments in both US and China primarely focused within the R&D department and is an associate member of the Institute of Chemical Engineers (IChemE).
