Table 1 Performance comparisons across different encoding types and model architectures for classification performance on the Atezolizumab dataset
From: The RESP AI model accelerates the identification of tight-binding antibodies
Encoding type | Model type | Num hidden layer weights | MCC on 5× CV for all class classification | AUC-ROC on 5× CV for RH03 vs rest classification | MCC on the test set for all-class classification |
|---|---|---|---|---|---|
One-hot (default) | Bayesian NN w/ ordinal regression (BNN-OR) | 84,090 | 0.717 ± 0.009 | 0.967 ± 0.001 | 0.721 |
Autoencoder | BNN-OR | 12,810 | 0.69 ± 0.006 | 0.966 ± 0.002 | 0.703 |
UniRep | BNN-OR | 114,930 | 0.62 ± 0.01 | 0.947 ± 0.003 | 0.638 |
ProtVec | BNN-OR | 3930 | 0.641 ± 0.003 | 0.852 ± 0.003 | 0.650 |
ESM-1b | BNN-OR | 39,330 | 0 (model did not converge) | – | – |
AbLang | BNN-OR | 23,970 | 0.636 ± 0.01 | 0.96 ± 0.002 | 0.664 |
AntiBertY | BNN-OR | 16,290 | 0.647 ± 0.007 | 0.961 ± 0.002 | 0.650 |
One-hot (default) | Fully connected net (FCNN) | 84,090 | 0.73 ± 0.01 | 0.973 ± 0.001 | 0.734 |
Autoencoder | FCNN | 12,810 | 0.731 ± 0.003 | 0.970 ± 0.001 | 0.734 |
UniRep | FCNN | 114,930 | 0.70 ± 0.01 | 0.963 ± 0.002 | 0.699 |
ProtVec | FCNN | 3930 | 0.690 ± 0.003 | 0.962 ± 0.002 | 0.683 |
ESM-1b | FCNN | 39,330 | 0 (model did not converge) | – | – |
AbLang | FCNN | 23,970 | 0.715 ± 0.003 | 0.969 ± 0.0008 | 0.719 |
AntiBertY | FCNN | 16,290 | 0.709 ± 0.008 | 0.968 ± 0.001 | 0.707 |
One-hot (default) | Random forest | NA | 0.673 ± 0.003 | 0.956 ± 0.003 | 0.672 |
Autoencoder | Random forest | NA | 0.7 ± 0.009 | 0.960 ± 0.003 | 0.708 |
