A new machine-learning model called CheMeleon, developed by researchers at MIT and BASF, is designed to predict how molecules will behave even when little experimental data is available for a specific task. The model's approach could streamline the process of discovering new drugs and materials, which traditionally relies on costly and time-consuming laboratory experiments. The research behind CheMeleon appeared in the September 14 issue of the Journal of Chemical Information and Modeling, with an earlier version available online since 2025.
CheMeleon operates in two stages. First, it undergoes pre-training on a large dataset of approximately one million molecules sourced from PubChem, a chemical database maintained by the U.S. National Institutes of Health. During this stage, the model learns to predict about 1,600 "molecular descriptors", numerical characteristics like molecular weight or the number of oxygen atoms, which chemists have developed over decades. This pre-training allows CheMeleon to build a general understanding of chemical relationships without relying on noisy experimental results.
After this initial training, users can fine-tune CheMeleon with a limited amount of task-specific data to predict various properties, including solubility, viscosity, bioactivity, or reactivity with specific proteins or enzymes. Akshat Shirish Zalte, an MIT graduate student involved in the research, noted that this contrasts with classical models where experts manually define how descriptors relate to a property. Instead, CheMeleon learns these relationships during its pre-training phase, similar to how large language models learn from vast text datasets.
The researchers evaluated CheMeleon on 58 benchmark datasets relevant to small-molecule drug discovery. They reported that CheMeleon outperformed other established models, including traditional methods and existing foundation models, across these tests. William Green, an MIT professor and co-author, stated that nine of the top 20 entries in an OpenADMET competition earlier this year utilized CheMeleon in their workflows for predicting properties of unsynthesized molecules. This indicates the model's adoption in drug discovery efforts.
Several pharmaceutical companies have shown interest in CheMeleon, with at least one company reportedly instructing its staff to use the model as a starting point for property modeling. Professor Green emphasized that CheMeleon "drastically reduces the number of molecules that need to be tested experimentally," which can lead to significant cost and time savings in the discovery process. However, the exact reduction in laboratory work has not yet been quantified.
CheMeleon is open-source and available through ChemProp, an MIT-developed machine-learning package. This allows other researchers to download and fine-tune the model for their specific applications. While CheMeleon helps prioritize molecules for further investigation, it does not eliminate the need for experimental validation of molecular behavior. The model's role is to narrow the field of candidates, making the experimental phase more efficient.