Skip to main content
. 2018 May 22;46(Web Server issue):W396–W401. doi: 10.1093/nar/gky432

Figure 2.

Figure 2.

Predictive performance of TCRmodel on a benchmark of recently determined TCR structures. (A) TCRs were modeled using structures from prior to 2016 as templates, and performance was compared with LYRA, another TCR modeling server (25), run with the same template restrictions. Backbone atom root mean square distances (RMSDs) between models and crystal structures were computed for full TCR models (Global), framework residues (FW), and individual CDR loops. Statistically significant lower RMSDs for TCRmodel (P < 0.05, two-tailed t-test) were observed for framework, CDR2α, CDR1β, and CDR2β. CDR3 loop refinement led to significantly improved RMSDs for CDR3α and CDR3β loops. (B) Performance comparison on the TCR benchmark, excluding templates from variable domains with >90% sequence identity to the modeled TCRs. Statistically significant lower RMSDs for TCRmodel (P < 0.05, two-tailed t-test) were observed for full models (Global), framework, CDR1α, CDR2α, CDR1β, and CDR2β. As in (A), CDR3 loop refinement in TCRmodel led to significantly improved RMSDs for CDR3α and CDR3β loops. Figure generated using the ggplot2 package in R (r-project.org).