Monroe combined with TabPFN led most often in an independent comparison of AI models for predicting drug molecule properties

Monroe combined with TabPFN led most often in an independent comparison of AI models for predicting drug molecule properties

On August 29, cheminformatician Pat Walters published an independent comparison of seven methods using two sets of ADMET measurements. These properties determine how a potential drug molecule behaves in the body. On both datasets, the combination of Monroe and TabPFN ranked among the best methods most often in the statistical comparison.

During early molecule screening, these predictions help researchers decide which compounds should proceed to further experiments. Walters evaluated the methods on identical tasks. Monroe converts a molecule’s structure into a numerical representation. TabPFN compares that representation with examples from a table of measurements and predicts the property of a new molecule.

For each dataset, some molecules were set aside in advance for final testing. The remaining data underwent five rounds of validation, with the molecules divided into five groups each time. Structurally similar molecules stayed in the same group, so the evaluation measured performance on entire chemical families. All seven approaches used the same splits, and their differences were assessed statistically.

On the first dataset, ExpansionRx, Monroe + TabPFN was either the best method or statistically indistinguishable from the best in 18 of 27 endpoint and metric combinations. CheMeleon achieved this result in 12. Mol-JEPA + TabICL and MEGA-CL had no first-place results and no statistical ties with the leader. In the Mol-JEPA preprint, its authors reported lower model error on some small datasets, although TabICLv2 was more accurate more often on public temporal splits. The new comparison tests that claim alongside other methods using shared splits.

“A second dataset is necessary because a comparison performed once is a hypothesis.”

The second collection of measurements, Biogen ADME, was added to test the conclusion from the first comparison. On this dataset, Monroe + TabPFN led in all 18 combinations. CheMeleon’s advantage on some tissue binding tasks in the first dataset did not recur. Selecting a model requires setting aside the test data in advance, comparing several approaches on the same molecules, and saving the predictions for later verification.

Claude Code assembled the computations from Walters’s instructions. The project log contains 22 substantive human prompts. These added a ChemProp control, a second data collection, and a check for possible leakage of Mol-JEPA labels during pretraining. The repository contains the code, comparison rules, and saved predictions, allowing readers to trace the decisions and verify the calculations.