Quantifying the Hardness of Bioactivity Prediction Tasks for Transfer Learning

Poster presentation at the Chemoinformatics Strasbourg Summer School (CS3-2024) introducing a framework to quantify how “hard” a bioactivity prediction task is, with the goal of guiding source-task selection for transfer learning under chronic data scarcity in drug discovery. See the associated publication.

The gain from meta-learning (ΔAUPRC over a baseline) decreases with the proposed external-plus-internal task hardness (Pearson r = −0.72), so hardness estimated before training predicts how much a new bioactivity task will benefit from knowledge transfer.

Authors: Hosein Fooladi, Steffen Hirte, and Johannes Kirchmair.