Table 3.

Comparisons for fitness predictors. Results were adopted from TopFit [20]. Performance was reported by average Spearman correlation over 34 DMS datasets and 20 repeats. Supervised model use ensemble regression from 18 regression models [20]

Zero-shot predictors
Model name	training set size
ESM-1b PLL [23, 33]	0.435
eUniRep PLL [127]	0.411
EVE [40]	0.497
Tranception [15]	0.478
DeepSequence [22]	0.504
Supervised models
Embedding name	training set size
Persistent homology [20]	0.263	0.432	0.496	0.534
Persistent Laplacian [20]	0.280	0.457	0.525	0.564
ESM-1b [23]	0.219	0.421	0.494	0.537
eUniRep [43]	0.259	0.432	0.485	0.515
Georgiev [127]	0.169	0.326	0.402	0.446
UniRep [21]	0.183	0.347	0.420	0.462
Onehot	0.132	0.317	0.400	0.450
Bepler [42]	0.139	0.287	0.353	0.396
TAPE LSTM [41]	0.259	0.436	0.492	0.522
TAPE ResNet [41]	0.080	0.216	0.305	0.358
TAPE transformer [41]	0.146	0.304	0.371	0.418