Evaluating molecular representations in machine learning models for drug response prediction and interpretability