MIT researchers say chemistry AI beats comparison models with little task-specific data

An AI model trained on easily calculated molecular measurements is helping researchers choose which compounds deserve expensive laboratory work. In a November 6 account, MIT detailed CheMeleon, developed with BASF researchers, and its reported advantage over comparison models across 58 tests. Its approach targets a difficult starting point: predicting a useful property when few task-specific measurements exist.

The model is not a new release. The research appeared in the September 14 issue of the Journal of Chemical Information and Modeling, and an earlier version had been online since last year. MIT’s new account describes both the method and its spread into pharmaceutical workflows, including use before the paper completed peer review.

Starting where experimental data run thin

Finding a molecule with the right behavior remains largely a trial-and-error process. Researchers can use calculations and machine learning to narrow the candidates worth making and testing. But a newly discovered cellular target for a potential drug, for example, may have little data showing how strongly different molecules bind to it.

MIT graduate student Akshat Shirish Zalte says many deep-learning models need more data than traditional methods. CheMeleon’s response is to learn a general chemical starting point first, rather than requiring each new prediction task to supply enough examples to teach the model everything from scratch.

The team began with about one million molecules from PubChem, the freely available chemical database maintained by the U.S. National Institutes of Health. For those molecules, it calculated roughly 1,600 descriptors: numerical characteristics such as molecular weight and the number of oxygen atoms. Predicting those descriptors became the model’s initial training task.

That differs from drawing training information from published chemistry literature, Zalte explains. It also differs from a classical expert-driven model, where specialists manually supply knowledge about which molecular characteristics relate to a property. CheMeleon instead learns relationships between molecular structure and the descriptor catalog during pre-training.

A reusable model meets outside tests

After that initial training, users can adapt CheMeleon to predict solubility, viscosity, bioactivity or reactivity with a particular protein or enzyme. The researchers compared it with other foundation models and classical models built around expert-designed principles. They report that CheMeleon came out ahead across 58 test cases.

An OpenADMET competition earlier this year provided another sign of uptake. According to MIT professor William Green, nine of the top 20 entries used CheMeleon in their process. The contest asked participants to predict properties of new molecules that had not previously been synthesized; it was not described as a comparison of CheMeleon alone against every other entry.

Green says the results prompted pharmaceutical-company inquiries while the paper was still awaiting peer review. At least one unnamed company subsequently instructed staff to start with CheMeleon whenever modeling a property for initial decisions about which molecules to pursue, he says.

Open access, with experimental savings still unquantified

CheMeleon is open source and available through ChemProp, the machine-learning package developed and maintained by Green’s MIT group. Users can download it and fine-tune it for their own purposes. The research team includes MIT scientists and three researchers from BASF SE in Germany.

Green says the approach sharply reduces how many molecules need experimental testing to predict new molecules’ properties accurately. That is the practical promise: fewer costly synthesis-and-testing steps. MIT’s account does not quantify those experimental savings, leaving their size in pharmaceutical workflows unresolved even as companies reportedly adopt the model.