MIT reports that CheMeleon led 58 molecular property prediction tests

A team at MIT developed CheMeleon, a machine-learning model designed to predict molecular properties when little data is available for a particular task. MIT reports that, in a comparison of 58 test cases, the model ranked first or achieved results statistically indistinguishable from the best according to the criteria used. Its strategy is to pretrain a neural network to predict calculated molecular descriptors and then fine-tune it for specific properties.

Predicting how a compound will behave could help researchers select which compounds to synthesize and test. According to MIT, molecular discovery remains largely a process of trial and error: there is no way to perfectly anticipate the variety of properties and responses of a compound. CheMeleon aims to help narrow the set of candidates that require experimental research, not eliminate the need for those tests.

Learning from descriptors

CheMeleon was pretrained on information from one million molecules in PubChem, an open-access chemical database maintained by the U.S. National Institutes of Health. At this stage, instead of learning from experimental measurements of a specific property, the model predicts values from a catalog of approximately 1,600 calculated molecular descriptors. MIT cites molecular weight and the number of oxygen atoms as examples.

The idea is for the network to learn patterns between a molecule’s structure and these descriptors. The model can then be fine-tuned to predict particular properties, such as solubility, viscosity, bioactivity, or reactivity with a protein or enzyme. This means that, when tackling a new task, it would not have to learn all the relationships between structure and property from scratch.

The authors’ preprint describes CheMeleon as a directed message-passing neural network (D-MPNN), pretrained to predict Mordred descriptors calculated for one million molecules. It presents these descriptors as deterministic, low-noise training targets and notes that the method combines chemical information encoded in descriptors with representations learned by a neural network. This approach explains the methodology, but does not by itself demonstrate performance in applications other than those evaluated.

Results on benchmark datasets

MIT reports that the team compared the method with other foundation models and with classical models based on principles designed by specialists. The preprint describes an evaluation on 58 benchmark datasets from Polaris and MoleculeACE. The authors report win rates of 75% on Polaris tasks and 97% on MoleculeACE tests. In the preprint’s summary, a win includes the best result or one statistically indistinguishable from the best under the evaluation criteria used.

These figures summarize results on the datasets and under the criteria described by the authors. They are not a general accuracy rate for every chemical problem, nor do they show that the model outperforms other methods for all properties or usage conditions. The MIT article and the preprint describe the same work: the preprint provides methodological and evaluation details, but does not independently corroborate the article.

Applications and required resources

MIT suggests that the predictions could guide which molecules are worth studying experimentally, in a process that can be slow and costly. The article also reports that several pharmaceutical companies have used CheMeleon and that nine of the twenty highest-ranked submissions in a competition organized by OpenADMET had used the model. These are examples of use and interest reported by the institutional source; by themselves, they do not constitute an independent evaluation of accuracy in specific products or programs.

The authors’ repository documents how to fine-tune CheMeleon with Chemprop 2.2.0 or later and how to generate molecular fingerprints without fine-tuning. It also indicates that reproducing the study requires considerable computing resources: calculating features for one million SMILES strings took approximately 500 CPU hours, and pretraining CheMeleon took around 1,000 GPU hours. These resources apply to the reproduction described in the repository, not to the cost of each individual prediction or fine-tuning run.

Overall, CheMeleon combines pretraining with calculated descriptors and subsequent fine-tuning for specific molecular properties. The reported results support its performance in the evaluations examined by the authors; its usefulness in a particular application depends on the data and relevant checks.