Paired antibody LMs outperform equivalent unpaired LMs on sequence recovery for various regions

Summary

Antibody language models trained on paired heavy and light chains outperform equivalent models trained only on unpaired data (1). They also outperform generic PLMs.

Figures

Model	FWH1	FWH2	FWH3	FWH4	CDRH1	CDRH2	CDRH3	Total VH
AbLang (Olsen et al., 2022b)	0.9795	0.9667	0.9560	0.9808	0.9099	0.8845	0.5926	0.9105
AntiBERTy (Ruffolo et al., 2021)	0.9784	0.9653	0.9545	0.9775	0.9073	0.8821	0.5209	0.8998
ProtBert (Elnaggar et al., 2022)	0.8018	0.7607	0.7384	0.8463	0.6560	0.4556	0.2772	0.6821
IgBert-unpaired	0.9791	0.9655	0.9552	0.9798	0.9043	0.8841	0.5924	0.9099
IgBert	0.9810	0.9690	0.9576	0.9809	0.9130	0.8865	0.6012	0.9129
ProtT5 (Elnaggar et al., 2022)	0.9037	0.8539	0.8880	0.9142	0.7530	0.6292	0.3390	0.7932
IgT5-unpaired	0.9790	0.9671	0.9560	0.9825	0.9092	0.8839	0.6035	0.9121
IgT5	0.9820	0.9687	0.9574	0.9828	0.9150	0.8936	0.6196	0.9163

Model	FWL1	FWL2	FWL3	FWL4	CDRL1	CDRL2	CDRL3	Total VL
AbLang (Olsen et al., 2022b)	0.9663	0.9683	0.9707	0.9621	0.8911	0.9008	0.8385	0.9493
AntiBERTy (Ruffolo et al., 2021)	0.9786	0.9687	0.9748	0.9661	0.9066	0.8951	0.8444	0.9553
ProtBert (Elnaggar et al., 2022)	0.6597	0.7862	0.7827	0.6337	0.4690	0.4382	0.2901	0.6654
IgBert-unpaired	0.9804	0.9704	0.9739	0.9656	0.9081	0.8985	0.8461	0.9560
IgBert	0.9885	0.9738	0.9807	0.9740	0.9232	0.9149	0.8634	0.9647
ProtT5 (Elnaggar et al., 2022)	0.8456	0.9010	0.8799	0.8499	0.6961	0.6038	0.5172	0.8200
IgT5-unpaired	0.9809	0.9675	0.9752	0.9171	0.9076	0.9093	0.8423	0.9515
IgT5	0.9878	0.9735	0.9815	0.9784	0.9222	0.9163	0.8693	0.9656

Ref (1)

1.

Kenlay H, Dreyer FA, Kovaltsuk A, Miketa D, Pires D, Deane CM. Large scale paired antibody language models. PLOS Computational Biology. 2024;20(12):e1012646. Available from: https://doi.org/10.1371/journal.pcbi.1012646