Table 3.

Performance of shAIC relative to other methods evaluated using the control set of 38 proteins and cross-validation of shAIC.

	Part A. Comparison with existing programs					Part B. shAIC: overall and segregated
	ShiftX	Sparta	CamShift	SHIFTS	Sparta+	ShAIC	Sheet	Helix	Coil
Correlation coefficient squared, Rtert2, for tertiary chemical shift a
N	0.520	0.564	0.468	0.282	0.661	0.653	0.721	0.600	0.653
C’	0.243	0.370	0.336	0.089	0.478	0.462	0.528	0.476	0.382
Cα	0.443	0.522	0.395	0.328	0.621	0.594	0.577	0.621	0.602
Cβ	0.350	0.455	0.399	0.172	0.533	0.477	0.573	0.396	0.479
HN	0.370	0.389	0.405	0.172	0.523	0.451	0.521	0.468	0.437
Hα	0.533	0.475	0.548	0.422	0.614	0.543	0.573	0.447	0.450
Correlation coefficient squared, Rsec2, for secondary chemical shift a
N	0.577	0.616	0.531	0.333	0.701	0.694
C’	0.555	0.643	0.624	0.340	0.704	0.696
Cα	0.741	0.784	0.724	0.661	0.828	0.815
Cβ	0.556	0.632	0.591	0.381	0.687	0.649
HN	0.435	0.455	0.463	0.249	0.573	0.508
Hα	0.721	0.695	0.733	0.647	0.772	0.733

	Rmsd/ppm (control set)						Rmsd/ppm shAIC (training set)
							Derivationb	Cross-validationc
N	2.827	2.694	2.905	4.313	2.356	2.343	2.391	2.561
C’	1.253	1.105	1.128	1.756	1.004	1.016	1.074	1.172
Cα	1.144	1.044	1.175	1.337	0.926	0.961	0.946	1.035
Cβ	1.219	1.099	1.157	1.525	1.012	1.071	1.138	1.336
HN	0.559	0.546	0.528	0.630	0.466	0.503	0.465	0.501
Hα	0.283	0.318	0.276	0.319	0.262	0.276	0.262	0.280
90% confidence intervals d /ppm (control set)
N	4.476	4360	4.674	7.005	3.675	3.769
C’	1.988	1.749	1.838	2.841	1.631	1.626
Cα	1.822	1.639	1.896	2.153	1.409	1.462
Cβ	1.993	1.761	1.865	2.425	1.651	1.750
HN	0.884	0.872	0.827	0.958	0.725	0.798
Hα	0.451	0.528	0.454	0.520	0.393	0.421

^aSquared correlation coefficients (coefficient of determination) for observed vs. predicted tertiary or secondary chemical shift is described in the text. Only chemical shift values for which all programs provided a prediction were included in the analysis (e.g., Sparta does not provide predictions for terminal residues). Outliers were removed from the analysis based on the criteria that, for all methods, the error was larger than five times the standard deviation, with the standard deviation estimated from rmsds in the training set broken down into residue and secondary structure type.

^brmsds between predicted and observed chemical shift in the training set of 681 proteins after derivation of all parameters.

^crmsds between predicted and observed chemical shift in the training set of 681 proteins using cross-validation. The set was divided into 10 equal subsets and for each subset the 9 other sets were used to derive the parameters, which in turn was used to predict the shift for the first set.

^d90% of the predictions have an error less than this threshold.