Meta Launches OMol25 and UMA: Powerful New AI Tools for Accelerating Molecular Research

Meta has unveiled two groundbreaking AI tools for molecular research: OMol25, the largest open dataset for AI-driven chemistry, and UMA, a universal AI model designed to predict chemical properties of molecules and materials. OMol25 comprises over 100 million high-precision molecular calculations, making it substantially larger than any prior open dataset in the field. The creation of this dataset required an impressive 6 billion hours of supercomputing time. It encompasses a diverse array of molecules, from small organic compounds and biomolecules (such as protein fragments and DNA segments) to metal complexes and electrolytes. The dataset also includes detailed information on charged and spin states, multiple conformations (spatial arrangements), and chemical reactions. By providing energy and force values alongside additional details like charge distributions and orbitals, OMol25 aims to help AI models understand how molecules behave in various scenarios. This comprehensive dataset is expected to significantly aid research in drug discovery, battery materials, and catalyst development. OMol25 is freely available on Hugging Face. Simultaneously, Meta is introducing UMA (Universal Model for Atoms), a new AI model trained on OMol25 and other datasets. UMA can predict chemical properties at the atomic level, doing so much faster than traditional methods. Unlike previous models, which required separate training for each specific application, UMA is versatile enough to handle a wide range of tasks, including molecular simulations for drug discovery and materials research for batteries and catalysts. The model employs modern graph neural networks and a "Mixture of Linear Experts" architecture, combining speed and accuracy. Benchmark tests have shown that UMA can achieve results previously attainable only with specialized, fine-tuned models. With UMA, simulations and calculations that once took days can now be completed in mere seconds, enabling researchers to efficiently screen thousands of new molecules before synthesizing them in the lab. UMA models are also available on Hugging Face. To further enhance AI-driven molecular simulations, Meta has developed a new technique called "Adjoint Sampling." Traditional AI models often require extensive training data to generate new molecular structures. Adjoint Sampling, however, allows AI models to learn and propose novel structures even in the absence of real-world examples. This method leverages concepts from stochastic control theory and diffusion processes, which are particularly effective for molecular simulations. Early tests have demonstrated that Adjoint Sampling can produce molecular conformations that not only match but frequently outperform those generated by classical software, especially for molecules with numerous flexible components. The model, code, and additional information for Adjoint Sampling are available on both Hugging Face and GitHub. While these advancements represent significant progress, Meta acknowledges that several challenges remain. For instance, areas such as polymers, certain metals, and complex protonation states are not yet fully addressed. Additionally, the AI models need to improve their ability to predict charges, spins, and long-range interactions accurately. Despite these ongoing issues, the introduction of OMol25 and UMA marks a substantial step forward in the field of computational chemistry, offering researchers powerful new tools to explore and innovate in the realm of molecular and materials science.