Removing Biases from Molecular Representations via Information Maximization