Learn a useful starting point
Predicting a molecule’s behaviour is difficult when a laboratory has only a small set of measurements. CheMeleon tackles that problem by learning a reusable starting point before it is trained on the particular property a researcher wants to predict.
MIT highlighted the work on November 6. This is an explainer tied to that account, not a new-paper announcement: the open manuscript was first submitted on June 18, 2025, and revised on February 9, 2026.
Calculated features first, experimental targets later
The authors pre-train the model on calculated features of one million PubChem molecules. These descriptors can capture structural details, such as atom counts, without requiring a fresh laboratory measurement for every training target.
The trained representation is then adapted using examples labelled for the task of interest, such as solubility or biological activity. The idea is to avoid learning both molecular structure and the target relationship from a small dataset at the same time.
That does not mean experimental data have become unnecessary. Calculated descriptors supply the initial learning task; measurements for the intended prediction still serve a different purpose.
What a 75% win rate means
Table 1 of the open manuscript reports 21 wins across 28 Polaris benchmarks, or 75%. A win includes either the best result or a result statistically indistinguishable from it under the authors’ comparison procedure.
It is not 75% accuracy across molecules, nor evidence that CheMeleon uniquely defeated every alternative on those tasks. The paper notes that several models can share a win, and some tasks do not clearly separate them. These are the authors’ results, not an independent replication.
Small structural changes can remain difficult
The discussion also warns that improved overall performance does not reliably identify activity cliffs: cases where a small structural change produces a large change in biological activity.
MIT describes the model as openly available through Chemprop. For a new project, the important follow-up is whether the chosen test resembles the intended molecules and property. A model that helps prioritise experiments has not thereby established that a compound is safe, effective or ready for use.
Sources
- MIT News: predicting molecular behaviour from scant dataNovember 6 institutional account supplies the current-events peg and accessible context. Adoption and performance statements are researcher/institution claims, not independently reproduced results.
- arXiv: CheMeleon manuscript submission recordPrimary author list and submission history: June 18, 2025 initial posting, February 9, 2026 revision. Explicitly not a new November manuscript.
- CheMeleon: open manuscript, version 2Primary methods, Table 1, model-comparison procedure and discussion. 21/28 Polaris wins include statistically indistinguishable best performance; activity-cliff limitations remain. The open version is not claimed to be identical to the inaccessible final journal text.
