Total number of training and testing examples for chemical shift prediction for each atom type. The training set is comprised of the combination of the SPARTA+ training set and the training and testing set for SHIFTX+, and removing all redundant chains. We have developed a new test set comprised of 200 high-resolution proteins with chemical shifts available from RefDB; the test data eliminates duplicate chains, and residues with no deposited chemical shift values. The LH-Test set refers to the subset of the total set of test proteins with only low sequence homology to other proteins such that sequence or structural homology cannot be exploited. We also created two curated test sets which additionally exclude paramagmetic proteins, some Hα chemical shifts that have calculated ring current effect exceeding 1.5 ppm, and “outliers” detected by the PANAV program (13). Further information is provided in Methods and ESI.

	# of PDBs	H	Hα	C′	Cα	Cβ	N
Train	647	72 894	56 149	58 228	79 611	70 621	74 896
Test	200	19 120	11 727	8231	13 140	10 139	15 374
Test (curated)	200	18 494	11 240	7861	12 533	9883	14 610
LH-Test	100	8634	4979	3332	5685	4278	6576
LH-Test (curated)	100	8606	4950	3331	5251	4201	6480