Table 2.

Spearman correlation between SAV effect prediction and DMS experiments^a

Method	Mean absolute ${\mathit{r}}_{\mathit{S}}$ (Eq. 11)	Median absolute ${\mathit{r}}_{\mathit{S}}$ (Eq. 11)
MSA-based
DeepSequence	0.50 ± 0.03	0.52 ± 0.03
GEMME	0.53 ± 0.02	0.56 ± 0.02
pLM-based
ESM-1v	0.49 ± 0.02	0.53 ± 0.02
VESPA	0.51 ± 0.02	0.53 ± 0.02
VESPAl	0.47 ± 0.02	0.47 ± 0.02

^aData sets: DMS39 [39 DMS experiments gathered for the development of DeepSequence (Riesselman et al. 2018)] with 135,665 SAV scores. Methods: DeepSequence: AI trained on MSA for each of the DMS experiments (Riesselman et al. 2018); GEMME: using evolutionary information calculated from MSAs with few parameters optimized on DMS (Laine et al. 2019); ESM-1v: embedding-based prediction methods (Meier et al. 2021); VESPA: method developed here using logistic regression to combine predicted conservation (ProtT5cons), BLOSUM62 (Henikoff and Henikoff 1992) substitution scores, and log-odds from ProtT5 (Elnaggar et al. 2021); VESPAl: “light” version of VESPA using only predicted conservation and BLOSUM62 as input. ± values mark the standard error