LM-based antibody structure prediction methods work equally well with generic and Ab PLMs

Summary

Predicting the structure of antibodies using embeddings from antibody language models leads to equal or worse performance compared to using embeddings from generic protein language models (1,2). The former found that training IgFold with with ESM2-35M embeddings gave comparable performance to using the AntiBERTy embeddings used by default, while the latter obtained better performance on ABodyBuilder3 ProtT5 embeddings compared to IgBERT and IgT5.

Figures

Method	Embedder	Embedder Status	Meta-arch	lDDT-Cα	OCD	H Fr	H1	H2	H3	L Fr	L1	L2	L3
IgFold(paper)					3.77	0.45	0.80	0.75	2.99	0.45	0.83	0.51	1.07
IgFold(reproduced)	Antiberty(25M)	Freeze	IGFold	0.93	3.74	0.57	0.92	0.80	3.09	0.67	1.12	0.55	1.15
IgFold-variant1	ESM-2(35M)	Freeze	IGFold	0.92	3.76	0.62	0.87	0.94	3.06	0.49	0.90	0.51	1.15
IgFold-variant2	ESM-2(650M)	Freeze	IGFold	0.93	3.77	0.48	0.91	0.94	3.20	0.48	0.94	0.49	1.13
IgFold-variant3	ESM-2(35M)	Trainable	IGFold	0.93	3.88	0.51	0.89	0.85	3.14	0.51	1.00	0.50	1.10

Ref (1)

1.

Lee J, Han K, Kim J, Yu H, Lee Y. Solvent: A Framework for Protein Folding. 2023; Available from: https://arxiv.org/abs/2307.04603

2.

Kenlay H, Dreyer FA, Cutting D, Nissley D, Deane CM. ABodyBuilder3: improved and scalable antibody structure predictions. Bioinformatics. 2024;40(10). Available from: https://doi.org/10.1093/bioinformatics/btae576