Machine Learning for Molecular Activity Prediction: Quantification of Task Relations and Hardness for Small Organic Molecules and Proteins

Poster presentation at the EUROPIN Summer School in Drug Design on quantifying the relationships and hardness between molecular activity prediction tasks for small organic molecules and proteins, motivating principled task selection in transfer learning settings.

A map of medicinal-chemistry task space: UMAP of 5,094 FS-Mol task centroids coloured by protein family. Arrows point from three highlighted test tasks to their nearest training tasks under the optimal-transport (OTDD) distance, the basis for principled source-task selection.

Authors: Hosein Fooladi and Johannes Kirchmair.