MIT team releases CheMeleon model for molecular prediction

New method handles scarce data

The new approach outperforms existing methods when few data points exist for specific technologically relevant properties. System creators include Graduate student Akshat Shirish Zalte, recent graduate Jackson Burns Ph.D., and Charlles Abreu, professor William Green, as well as three BASF researchers. Earlier this year, their results appeared in the Journal of Chemical Information and Modeling. The team built CheMeleon using ChemProp, an open-source machine-learning package also maintained by the Green group at MIT.

Pharma companies adopt the tool

Several pharmaceutical companies have already found regular use for the open-source model in their workflows. An earlier contest organized by OpenADMET saw nine of the top 20 entries utilize CheMeleon to predict properties of unsynthesized molecules. Professor Green noted that one pharma company sent a message instructing all staff to start with CheMeleon for initial modeling decisions. Anyone can download the software through ChemProp and fine-tune it for their own specific purposes.

Model learns from descriptors

The system pretrains on detailed data about a million different molecules to learn structural correlations. It predicts properties ranging from solubility and viscosity to bioactivity or reactivity with certain proteins. The training set includes about 1,600 descriptors developed by chemists over decades for this purpose. This minuscule fraction of known chemistry allows the model to guide synthesis efforts effectively.