Foundational Chemillion model predicts molecule properties from scarce samples and reduces drug tests
Listen to this article
Read by Anchor
A joint research team from the Massachusetts Institute of Technology and the German company BASF succeeded in developing an open-source artificial-intelligence model called Chemillion, which enables the prediction of chemical molecules’ properties and behaviour based on extremely small amounts of experimental data. The approach aims to overcome one of the heaviest obstacles facing the pharmaceutical and materials industries, where the synthesis and laboratory testing of new compounds remain costly, time-consuming trial-and-error processes, and where there is no prior certainty about how a compound will interact with cellular targets or how toxic it will be, nor can its solubility and melting point be measured accurately before actual production.
The researchers trained the foundational model on one million molecules extracted from the public database PubChem belonging to the U.S. National Institutes of Health. Unlike conventional deep models that rely on the text of published papers and require massive databases of prior experiments, the team focused on making the model learn chemical principles by predicting about 1,600 classic molecular descriptors devised by chemists over decades, including molecular weight, number of oxygen atoms, and the geometric shape of the molecule in terms of size and branching.Pre-training on classical structural principles gave the model a general chemical understanding that can later be fine-tuned with only a few data points for the target property.
Published comparative test results in the journal *Chemistry & Modeling* show Chemillion outperforming other foundational models and traditional systems run by specialists in 58 out of 58 clinical and chemical tests. The open-source ChemPro package allows researchers and companies to download the model and re-configure it to estimate multiple properties, including solubility, viscosity, biological activity, toxicity rates, and the degree of binding to newly discovered cellular receptors, sharply reducing the number of compounds that need experimental wet-lab testing.
The model began to make a real impact on pharmaceutical production lines before its scientific review was complete, as an international competition organized by the OpenAdmit platform to predict properties of novel, previously unmanufactured molecules showed that nine of the top twenty teams used Chemillion in their work. This success prompted global pharmaceutical companies to reach out to the team, and one major drug company adopted an internal policy requiring its research teams to use the model as a unified starting point when modelling any molecular property, thereby avoiding wasteful chemical tests.
Relying on pre-defined physical descriptors frees drug developers from the monopoly of large data sets that only giant central laboratories possess.This path gives pharmaceutical research teams, biotechnological labs, and petrochemical science groups in the Gulf, Egypt, and the Levant direct competitive advantage, because work in our region often encounters the high cost of initial laboratory tests and the scarcity of locally available disease or target compound samples. An open model capable of accurately inferring behaviour from limited samples cuts foundational experiment budgets and shifts evaluation tasks from expensive wet-lab testing to digital simulation that precedes any investment decision in chemical or pharmaceutical manufacturing.
The model available through the ChemPro platform opens a practical window for academic institutions and start-ups to reduce reliance on complex computational infrastructure, providing researchers with a flexible tool that gains expertise over time without needing to re-engineer weights from scratch for each new compound.