Advances in artificial intelligence have reshaped how chemists predict reactions, but turning predictions into reliable laboratory procedures remains a persistent challenge. A research team at Yale University, led by Victor Batista, has introduced a platform aimed at addressing this gap by organizing chemical knowledge in a way that mirrors how chemists actually work at the bench.
Li, H., Sarkar, S., Lu, W., Loftus, P. O., Qiu, T., Shee, Y., Cuomo, A. E., Webster, J.-P., Kelly, H. R., Manee, V., Sreekumar, S., Buono, F. G., Crabtree, R. H., Newhouse, T. R., & Batista, V. S. (2026). Collective intelligence for AI-assisted chemical synthesis. Nature. https://doi.org/10.1038/s41586-026-10131-4
The platform, known as MOSAIC, compiles millions of published chemistry protocols into a structured system that can generate step-by-step experimental procedures for synthesizing molecules, including compounds that have not previously been reported. Rather than relying on a single large language model, MOSAIC is built around thousands of smaller, specialized AI agents, each representing expertise in a specific area of chemical synthesis.
Yale University, led by Victor Batista stated,
“Chemistry has accumulated millions of reaction protocols, but making practical use of that knowledge remains a bottleneck. MOSAIC is designed to transform that information overload into actionable laboratory procedures.”
Batista, a professor of chemistry in Yale’s Faculty of Arts and Sciences and director of the Center for Quantum Dynamics on Modular Quantum Devices, describes the problem MOSAIC addresses as one of access rather than availability. Chemistry literature contains an enormous volume of validated experimental methods, but locating the most relevant protocol for a specific synthetic goal can take days or weeks. MOSAIC is designed to reduce that friction by translating dispersed knowledge into actionable laboratory guidance.
The system was developed in collaboration with researchers from Boehringer Ingelheim Pharmaceuticals’ U.S. research division, reflecting growing industry interest in AI tools that extend beyond prediction into experimental execution. Timothy Newhouse, a Yale chemistry professor and co-corresponding author on the study, explains that synthetic chemistry already operates much like cooking: researchers follow tested procedures, adapt conditions, and learn from prior attempts. MOSAIC formalizes this process by allowing chemists to query a broad, curated set of “recipes” drawn from across the literature.
One distinguishing feature of MOSAIC is its modular design. Instead of averaging knowledge into a single model, the platform consults 2,498 independent AI “experts,” each trained on a defined subset of chemical reactions or techniques. This structure allows the system to combine insights from multiple domains when proposing a synthesis route, improving performance across a wider range of chemical spaces.
In benchmarking studies, MOSAIC was shown to outperform several commercial AI chemistry tools on tasks involving protocol generation and compound diversity. The platform was used to guide the successful synthesis of more than 35 previously unreported compounds, spanning applications in pharmaceuticals, catalysts, materials science, and consumer chemistry.
Another practical feature is MOSAIC’s ability to provide uncertainty estimates alongside its proposed procedures. These estimates indicate how closely a given request aligns with the experience of the underlying expert models, offering chemists a way to judge risk and prioritize experiments rather than treating AI outputs as equally reliable.
The developers have released MOSAIC as an open-source framework, making it compatible with future AI models and adaptable to new data sources. This decision reflects a broader shift in computational chemistry toward tools that support reproducibility and shared infrastructure, rather than closed systems optimized for narrow tasks.
According to first authors Haote Li and Sumon Sarkar, the long-term goal is to move AI from a predictive role into one that actively supports experimental decision-making. In practice, MOSAIC functions less like an oracle and more like a navigation system, helping researchers move through complex chemical spaces using established knowledge, while signaling where uncertainty remains.
As chemistry continues to evolve from printed handbooks to digital databases and now to AI-assisted platforms, tools like MOSAIC suggest a future where the bottleneck in drug and materials development may shift from finding information to efficiently applying it in the laboratory.
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).
