← Back to Home

Predicting Molecular Properties with Limited Experimental Data? MIT Explores Drug Screening with CheMeleon

By first learning computable molecular features and then adapting the model with limited experimental data, CheMeleon aims to improve prediction in early drug discovery. Results from 58 benchmark datasets offer support, but demonstrating an improvement in actual R&D success rates remains some way off.

By SURL BioNews

When researchers have just identified a protein that may be linked to a disease, they often do not yet know which molecules are worth the time to synthesize and test. This is when predictive tools are most needed, yet experimental data are scarcest. CheMeleon, introduced by the Massachusetts Institute of Technology (MIT) on November 6, addresses this challenge: the model first learns from the basic features of large numbers of molecules, then helps assess candidate compounds’ biological activity, toxicity, or solubility.

The work, a collaboration between researchers at MIT and BASF, was published in the *Journal of Chemical Information and Modeling* on September 14. The team evaluated the model on 58 benchmark datasets related to small-molecule drug discovery. According to the journal abstract, statistical comparisons support CheMeleon’s superiority over traditional methods, such as random forests using molecular fingerprints or descriptors as inputs, as well as existing foundation models. However, this represents an overall advantage under the conditions tested and should not be interpreted as an ability to predict every property accurately.

CheMeleon starts with “molecular descriptors,” which chemists have long used: features that can be calculated from a structure, such as molecular weight and the number of oxygen atoms. According to MIT, the researchers pretrained the model using approximately 1 million molecules and approximately 1,600 descriptors, allowing it to learn the relationships between molecular structures and these features. These training labels can be obtained computationally, reducing pretraining’s reliance on expensive experimental measurements.

After this initial learning stage, the model, which has approximately 10 million parameters, is fine-tuned with task-specific data. In drug discovery, its purpose is to help prioritize experiments: when only a small number of activity measurements are available for a new target, or solubility data for candidate molecules are limited, researchers hope to use it to narrow the scope of the next round of testing. Whether it can maintain its advantage across different chemical structures and experimental conditions still requires evaluation on a case-by-case basis.

There are also signs of the model’s use in a competition. Citing the research team, MIT reported that 9 of the top 20 entries in the OpenADMET competition used CheMeleon in their workflows. This shows that the model can be incorporated into other teams’ prediction work, but the rankings reflect complete competition entries; this number alone cannot isolate CheMeleon’s own contribution.

The method was made public before its journal publication: the preprint first appeared online in June 2025 and was revised in February 2026. The authors’ public code repository provides the code needed to reproduce the research and links to training data and model weights. Users can fine-tune the model through ChemProp or directly generate molecular feature representations for other analyses. These public resources give external teams an opportunity to examine the results and test how well the model applies to their own data.

The evidence so far remains focused on molecular property prediction. Benchmark and competition performance have not yet demonstrated how much the tool can shorten R&D timelines, reduce costs, or increase the likelihood that drug candidates will enter clinical trials. Toxicity predictions also cannot be treated directly as conclusions about safety in humans. For R&D teams, the next key question is whether molecules selected based on the model’s recommendations can consistently deliver better results in subsequent experiments.

References

  1. MIT News
  2. PubMed / Journal of Chemical Information and Modeling
  3. arXiv
  4. Jackson Burns / GitHub