TABLE 1.

Compilation of the available antibody structure prediction methods that leverage recent advances in machine learning. For each method, we describe the general goal (e.g., CDR prediction or whole variable region prediction), the accuracy of the most difficult region, the CDR-H3, its code/server availability, and the source paper. Please note that the CDR-H3 root mean square deviations (RMSDs) are not directly comparable as they could have been obtained from a different test set and are sometimes calculated in a different fashion, e.g., based on Cα or main chain heavy atom positions. As a baseline and reference point, we also include the AlphaFold2 predictions since many methods report values with respect to that method.

Method	Problem addressed	Model characteristic	CDR-H3 prediction accuracy	Corresponding AlphaFold2 accuracy	Availability	Source
DeepH3	CDR prediction	Residual neural network	2.2 Å backbone atoms are used	N/a	https://github.com/Graylab/deepH3-distances-orientations	Ruffolo et al. (2020)
Quaternion and Euler angle combined method	CDR prediction	Graph neural network	SAbDab benchmark: 2.29 Å	N/a	N/a	Son et al. (2022)
ABlooper	CDR prediction	Graph neural network based	RosettaAntibody benchmark: 2.49 Å; SAbDab latest structures: 2.72 Å. Backbone atoms were used	RosettaAntibody benchmark: 2.87 Å	https://github.com/oxpig/ABlooper	Abanades et al. (2022a)
DeepSCAb	Antibody side chain prediction	Residual neural network	Not applicable (side chain prediction)	N/a	https://github.com/Graylab/DeepSCAb	Akpinaroglu et al. (2022)
NanoNet	Heavy chain prediction	Residual network	RosettaAntibody benchmark: 2.38 Å; Nanobodies: 3.16 Å. Backbone atoms were used	Nanobodies: 2.88 Å	https://github.com/dina-lab3D/NanoNet	Cohen et al. (2022)
AbodyBuilder2	Variable region prediction	Based on AlphaFold2 structural module	AbodyBuilder2 benchmark: 2.81 Å. Backbone atoms were used	AbodyBuilder2 benchmark: 2.90 Å	https://github.com/oxpig/ImmuneBuilder	Abanades et al. (2022b)
EquiFold	Variable region prediction	SE(3)-equivariant neural network	IgFold benchmark: 2.86 Å (only N, Cα, C, and O RMSD)	IgFold benchmark: 3.02 Å	https://github.com/Genentech/equifold	Lee et al. (2022)
tfold-Ab	Variable region prediction	Based on AlphaFold2, using language models in the place of Evoformer	IgFold benchmark: 2.74 Å; SAbDab-22H1-Ab benchmark: 3.03 Å. Backbone atoms were used	IgFold benchmark: 3.02 Å; SAbDab-22H1-Ab benchmark: 3.18 Å	https://drug.ai.tencent.com/en	Wu et al. (2022)
xTrimoABfold	Variable region prediction	Based on AlphaFold2, using language models in place of Evoformer	1.25 Å (Cα only)	1.47 Å	N/a	Wang et al. (2022)
IgFold	Variable region prediction	Graph transformer using language model AntiBERTy	IgFold benchmark: 2.99 Å (backbone heavy atoms)	IgFold benchmark: 3.02Å	https://github.com/Graylab/IgFold	Ruffolo et al. (2022a)
AbFold	Variable region prediction	Based on AlphaFold2	AbFold benchmark: 3.04 Å, (backbone heavy atoms)	AbFold benchmark: 3.14 Å (backbone heavy atoms)	N/a	Peng et al. (2023)
AbBERT-HMPN	Sequence and structure generation	Deep graph neural network employing language models with generative capabilities	2.38 Å backbone atoms were used	N/a	N/a	Gao et al. (2022)
RefineGNN	CDR prediction and design	Graph neural network with generative capabilities	2.50 Å (Cα only)	N/a	https://github.com/wengong-jin/RefineGNN	Jin et al. (2021)
AbDockGen	CDR-H3 prediction, design, and antigen docking	Graph neural network-based with generative capabilities	Not applicable (docking scores reported)	N/a	https://github.com/wengong-jin/abdockgen	Jin et al. (2022)
DiffAb	Antibody sequence and the structure design	Diffusion method	Test set of 19 complexes: 3.246 Å (Cα only)	N/a	https://github.com/luost26/diffab	Luo et al. (2022)
DeepAb	Variable region prediction	Residual neural network	RosettaAntibody benchmark: 2.33 Å; therapeutics: 2.52 Å. Backbone heavy atoms were used	N/a	https://github.com/RosettaCommons/DeepAb	Ruffolo et al. (2022b)