Table 1.

Accuracy and improvement of neural network predictors of natural disordered regions (PONDRs^®)

Name	Training Set	# Disordered Residues	Accuracies %
			Ordera	Disorderb
XL1	7 X-ray	502	71	47
VL1	7 NMR, 8 X-ray	1,366	83	45
XL-XT	VL1 plus XTc		71	59
VL2	53 X-ray, 35 NMR, 52 CD	17,978	76	65
VL3d	54 X-ray, 40 NMR, 58 CD	22,434	84	59
VSL1e	230 long DRf 983 short DR Ordered regions	25,958 9,632 354,169	83	79
VSL2g	230 long DR 983 short DR Ordered regions	25,958 9,632 354,169	81	82

^aO_PDB_S25

^bCombined dis_X-ray, dis_NMR and dis_CD

^cXT is a joint name for the N-terminus (XN), and the C-terminus (XC) predictors, which were trained using x-ray crystallographic data, where the terminal disordered regions were 5 or more amino acids in length.

^dBesides the addition of a few more chains, substantial cleaning of the training databases was carried out between VL2 and VL3. Several incorrectly labeled chains were identified and fixed and order/disorder boundaries were adjusted in a few other proteins

^eThe VSL1 predictor combines two predictors optimized for long (>30 residues) and short (≤30 residues) disordered regions, respectively, using weights generated by a third meta-predictor. The attributes used include amino acid frequencies, sequence complexity, ratio of net charge / hydrophobicity, averaged flexibility, and averaged PSI-BLAST profiles calculated over symmetric input windows.

^fDisordered region

^gVSL2 is a slightly improved version of VSL1 predictor. The training data for VSL2 were slightly different: 8 ambiguous sequences were removed; His-tags were not used in training, short DR of 1-3 residues were not used in training. Also, linear SVM instead of logistic regression was used for VSL2 version (Kang Peng, personal communication).

Note: Both VSL1 and VSL2 take advantage of length dependencies