Fig. 6: Analysis on model design and ablation study. | Nature Communications

Fig. 6: Analysis on model design and ablation study.

From: ProRefiner: an entropy-based refining strategy for inverse protein folding with global graph attention

Fig. 6

a Hydrogen bonds between two target residues in blue (HIS 9, ILE 70 on 2KCD) and their most attended residues in orange. b A salt bridge and a hydrogen bonds between ASP 70 of T4-lysozyme and its most attended residues. c. A disulfide bond between CYS 99 and CYS 94 of human Ero1-alpha (Q96HE7). d Models' median recovery rate on CATH dataset (n = 1120 structures), with respect to different percentages of residues masked on sequences generated by base model ESM-IF1. e Runtime and GPU memory usage of ProRefiner and the model without pseudo edge feature, denoted by ProRefiner - PsFeat. Data is obtained through 16 independent runs and the plots show the mean with the 95% confidence interval of the mean, estimated from 1000 bootstrap samples. f Recovery rates and nssr scores of ProRefiner and ProRefiner - PsFeat on CATH (n = 1120), EnzBench (n = 51) and BR_EnzBench (n = 320). ESM-IF1 is used as the base model for inference on CATH. The inner box plots show the first quartile, median and the third quartile. Whiskers in box plots extend to the most extreme data point that lies within 1.5 times IQR from the nearest quartile. Source data are provided as a Source Data file.

Back to article page