Fig. 1

The training scheme for the side chain–side chain interactions. Every prediction is minimized (1) and rescored (2). Each prediction is classified as either TP, FP, FN, or TN (3). For each of these classes, an average contact map is created. Contact maps are shown for an artificial example containing only three residues. To train the potential, the side chain–side chain interaction a/b is selected because it is more frequent in TP and FN predictions than in FP and TN (4). The side chain–side chain interaction a/c, on the other hand, is selected because it is more frequent in FP and TN than in TP and FN predictions (4). The a/b interaction is strengthened by decreasing its energy, while the a/c interaction is disfavoured by increasing its energy (5). This leads to the new scoring function \(E_{86}^{\text {trained}}\), with which the predictions are rescored. Steps (3) to (6) are iterated 30 times on the training dataset