Fig. 4: Utility of IgFold for antibody structure prediction.
From: Fast, accurate antibody structure prediction from deep learning on massive set of natural antibodies
Box plots have center at median, bounds indicating interquartile range (IQR), whisker length of 1.5 × IQR, and poitns outside of 1.5 × IQR range shown as outliers. Source data are provided as a Source Data file. a Paired antibody structure prediction benchmark results (n = 197 structure predictions) for IgFold without templates, IgFold given the FV structure without the CDR H3 loop (IgFold[Fv-H3]), and IgFold given the complete Fv structure (IgFold[Fv]). b Per-target comparison of CDR H3 loop structure prediction for IgFold and IgFold[Fv-H3], with each point representing the RMSDH3 for both methods on a single benchmark target. c Nanobody structure prediction benchmark results (n = 71 structure predictions) for IgFold without templates, IgFold given the FV structure without the CDR3 loop (IgFold[Fv-CDR3]), and IgFold given the complete Fv structure (IgFold[Fv]). d Per-target comparison of CDR3 loop structure prediction for IgFold and IgFold[Fv-CDR], with each point representing the RMSDCDR3 for both methods on a single benchmark target. e Runtime comparison of evaluated methods on the paired antibody structure prediction benchmark (n = 197 structure predictions). ABlooper runtimes are calculated given an IgFold-predicted framework, and thus represent an underestimation of actual runtime (f) Runtime comparison of evaluated methods on the nanobody structure prediction benchmark (n = 71 structure predictions). g Distribution of predicted RMSD and CDR H3 loop lengths for 1.3 million predicted human paired antibody structures.
