Data repositories from historical pharmaceutical research often reflect biased sampling that limits machine learning model generalization