Text this: Learning from low-rank multimodal representations for predicting disease-drug associations