Removing Biases from Molecular Representations via Information Maximization

Explore a 53-minute conference talk on removing biases from molecular representations through information maximization. Dive into the challenges of high-throughput drug screening and learn about InfoCORE, an innovative approach to tackle batch effects in large-scale experiments. Discover how this method establishes a variational lower bound on conditional mutual information and adaptively reweighs samples to equalize batch distributions. Examine the superior performance of InfoCORE in molecular property prediction and molecule-phenotype retrieval through extensive experimental results. Gain insights into how this versatile framework resolves general distribution shifts and addresses data fairness issues by minimizing correlation with spurious features and removing sensitive attributes. The talk covers background information, details of the InfoCORE method, experimental results, and concludes with a Q&A session.