Fig. 9
From: How to approach machine learning-based prediction of drug/compound–target interactions
Classification-based test performance results of protein family-specific PCM models (each using a different representation type as an input feature vector) in terms of MCC scores for (a) random-split, (b) dissimilar-compound-split, and (c) fully-dissimilar-split datasets. The models are ranked according to decreasing performance on the fully-dissimilar-split dataset (i.e., the large-scale analysis)