Adaptive Invariant Risk Minimization for Molecule Property Prediction

The current race to identify promising repurposing drug candidates for COVID-19 highlights a significant limitation of existing property prediction algorithms. Since their accuracy tightly depends on the access to large amounts of in-domain training data, these algorithms provide limited utility today when high-throughput screening is lacking. At the time of writing, there are only 48 drugs and 880 fragments with measured in-vitro SARS-CoV-2 activity in the public domain [6]. This scenario is not unique to the current pandemic, but is also common for initial stages of novel therapeutic development where limited screening data has to inform on-going experimental exploration.