Figure 1

Two protein classification problems. (Left) In the SCOP database, we simulate the remote homology detection problem by holding out a test family (shown in dark gray) from a superfamily and using the other families as positive training data (shown in light gray). The task is to correctly predict the superfamily or fold membership of the held-out sequences. (Right) We simulate the fold recognition problem by holding out a test superfamily (dark gray) from a fold and using the other superfamilies as training data (light gray). The task is to correctly recognize the fold of the held-out sequences.