Because no experimental structure of TGFBR2 exists, we queried to the PDB (Berman et al. 2000 (link)) using T-Coffee (Di Tommaso et al. 2011 (link)) and generated a homology-based model against the best available template using MODELLER (Sali and Blundell 1993 (link)) version 9.15. The template of adenine-bound ACVR2B, 2QLU (Han et al. 2007 (link)) exhibited 46% sequence identity for the modeled region. We generated 20 models and selected the model with the lowest DOPE (discrete optimized protein energy) score (Shen and Sali 2006 (link)) for further analyses. Model quality was assessed using PROCHECK (Laskowski et al. 1996 (link)), QMEAN (Benkert et al. 2009 (link)) and VADAR version 1.8 (Willard et al. 2003 (link)). Pathogenic and likely pathogenic missense variants were extracted from ClinVar (Landrum et al. 2014 (link)) and mapped to our TGFBR2 model. In silico mutagenesis was performed using Discovery Studio (Dassault Systèmes BIOVIA, Discovery Studio Modeling Environment, Release 2017, San Diego: Dassault Systèmes, 2017). Changes in folding energy upon mutation (ΔΔGfold) were computed using FoldX version 4 (Schymkowitz et al. 2005 (link); Van Durme et al. 2011 (link)).