Residue conservation and solvent exposure data perform comparably to PLMs at some property prediction tasks

Summary

Residue conservation and solvent exposure data perform comparably to PLMs at some property prediction tasks (1). The former was assessed using position-specific scoring matrices. However, as with language models, they perform worse than inverse folding on stability prediction.

Figures

Method	Model type	All	Stability	Binding	Expression	Activity	Fitness
−RSA	STR*	0.356	0.480	0.281	0.323	0.330	0.286
LOR	ALI*	0.427	0.447	0.375	0.390	0.452	0.414
RSALOR	STR & ALI	0.473	0.551	0.455	0.428	0.472	0.419
ProSST-2048 [13]	STR & pLM	0.522	0.638	0.527	0.527	0.486	0.441
PoET [14]	ALI & pLM	0.470	0.458	0.440	0.459	0.495	0.474
SaProt (650M) [19]	STR & pLM	0.462	0.565	0.441	0.482	0.459	0.375
VespaG [20]	pLM	0.461	0.479	0.415	0.450	0.489	0.440
TranceptEVE (L) [21]	ALI & pLM	0.450	0.424	0.405	0.447	0.489	0.458
GEMME [22]	ALI	0.447	0.452	0.367	0.430	0.477	0.444
EVE (ensemble) [23]	ALI	0.431	0.410	0.382	0.398	0.466	0.446
ESM2 (650M) [24]	pLM	0.428	0.496	0.382	0.409	0.431	0.381

Table from (1)

See also

• A subset of residues is sufficient to predict protein dG

1.

Tsishyn M, Hermans P, Rooman M, Pucci F. Residue conservation and solvent accessibility are (almost) all you need for predicting mutational effects in proteins. Bioinformatics. 2025;41(6). Available from: https://doi.org/10.1093/bioinformatics/btaf322