Fig. 1

Effect of homology threshold to predict 13 sub-structures. The accuracy Q₁₃ for classifying proteins into 13 sub-nuclear compartments using the homology-based inference with PSI-BLAST (based on 3522 experimentally annotated proteins) varied with the E-value thresholds (darker gray bars on the left). For proteins for which a protein with experimentally known nuclear sub-structure annotation was more sequence similar than the threshold, performance depended on the threshold (black line). The highest accuracy Q₁₃ = 68% was reached at E-value ≤10^− 50(black arrow). However, if forcing predictions for all proteins, Q₁₃ dropped to 38% compared to random (27%). The performance of machine learning-based profile kernel SVMs on the same set was Q₁₃ = 59% (gray horizontal line). The lighter gray bars mark the combination of homology inference and machine learning. The optimal threshold for the combination was E-value ≤10^− 20. One standard error marked on each bar and on the black line and through the dotted lines for ML