Template CDR retrieval using embeddings

Summary

CDR templates with high structural similarity can be fetched by comparing the embeddings to those of templates in the PDB, fetching the top ten, and taking the medoid by Euclidean distance (1). This strategy was found to be more accurate than either DeepAb and AlphaFold2. On CDRH3 it was more accurate than OmegaFold, while on other CDRs it was about the same (results only show the LSTM model after fine-tuning, not the other two).

Figures

Model	Metric	Chain	CDR 1	CDR 2	CDR 3	Whole Fv
AbMAP-B	TM-Score	H	0.62 ± 0.007	0.67 ± 0.006	0.54 ± 0.007	0.86 ± 0.006
ProtBert	TM-Score	H	0.33 ± 0.003	0.31 ± 0.003	0.28 ± 0.003	0.69 ± 0.005
ESM-1b	TM-Score	H	0.49 ± 0.007	0.51 ± 0.006	0.44 ± 0.007	0.79 ± 0.004
DeepAb	TM-Score	H	0.51 ± 0.008	0.55 ± 0.008	0.24 ± 0.004	0.54 ± 0.005
OmegaFold	TM-Score	H	0.61 ± 0.007	0.67 ± 0.006	0.34 ± 0.005	0.79 ± 0.005
AlphaFold2	TM-Score	H	0.28 ± 0.003	0.30 ± 0.003	0.30 ± 0.004	0.65 ± 0.004
AbMAP-B	TM-Score	L	0.62 ± 0.007	0.76 ± 0.007	0.65 ± 0.007	0.89 ± 0.004
ProtBert	TM-Score	L	0.34 ± 0.003	0.60 ± 0.007	0.52 ± 0.010	0.80 ± 0.005
ESM-1b	TM-Score	L	0.41 ± 0.005	0.61 ± 0.007	0.39 ± 0.005	0.82 ± 0.004
DeepAb	TM-Score	L	0.40 ± 0.006	0.66 ± 0.009	0.38 ± 0.006	0.52 ± 0.005
OmegaFold	TM-Score	L	0.63 ± 0.006	0.69 ± 0.006	0.58 ± 0.006	0.83 ± 0.005
AlphaFold2	TM-Score	L	0.27 ± 0.003	0.36 ± 0.007	0.31 ± 0.003	0.58 ± 0.004
AbMAP-B	RMSD	H	0.43 ± 0.016	0.38 ± 0.013	0.43 ± 0.025	2.11 ± 0.082
ProtBert	RMSD	H	1.40 ± 0.030	1.21 ± 0.028	0.64 ± 0.031	3.07 ± 0.078
ESM-1b	RMSD	H	0.54 ± 0.017	0.76 ± 0.022	0.50 ± 0.023	2.69 ± 0.063
DeepAb	RMSD	H	0.56 ± 0.016	0.55 ± 0.015	0.81 ± 0.031	0.72 ± 0.028
OmegaFold	RMSD	H	0.35 ± 0.013	0.37 ± 0.013	0.75 ± 0.035	2.39 ± 0.067
AlphaFold2	RMSD	H	0.70 ± 0.032	0.37 ± 0.030	1.24 ± 0.063	4.40 ± 0.077
AbMAP-B	RMSD	L	0.41 ± 0.015	0.18 ± 0.006	0.44 ± 0.022	1.42 ± 0.044
ProtBert	RMSD	L	1.39 ± 0.041	0.15 ± 0.008	0.64 ± 0.044	0.76 ± 0.012
ESM-1b	RMSD	L	0.68 ± 0.022	0.22 ± 0.008	0.65 ± 0.028	2.38 ± 0.062
DeepAb	RMSD	L	0.72 ± 0.021	0.32 ± 0.011	0.84 ± 0.030	1.16 ± 0.084
OmegaFold	RMSD	L	0.40 ± 0.016	0.17 ± 0.007	0.46 ± 0.017	2.31 ± 0.076
AlphaFold2	RMSD	L	1.13 ± 0.046	0.24 ± 0.024	0.64 ± 0.033	4.79 ± 0.084

From (1)

See also

• CDR representations segregate into distinct clusters
• Contrastive learning on whole structures leads to learning of distinct substructures

1.

Singh R, Im C, Qiu Y, Mackness B, Gupta A, Joren T, et al. Learning the language of antibody hypervariability. Proceedings of the National Academy of Sciences. 2024;122(1). Available from: https://doi.org/10.1073/pnas.2418918121