Table 1.
Overview of the tested models.
| Models | Architecture | Embedding | Year | Trainable parameters |
|---|---|---|---|---|
| TITAN Weber et al. (11) | Bimodal attention networks, pretrained with BindingDB. | Encoded peptides with SMILES, TCRs with BLOSUM62 and padded to the same length. | 2021 | 15,506,099 |
| DLpTCR Xu et al. (14) | Ensemble network out of: FCN, CNN and ResNet | depending on subNN: PCA on 500 amino acid indices, one-hot encoded or 20 different physicochemical properties (PCP) | 2021 | 10,454,869 |
| ERGO Springer et al. (13) | Autoencoder or LSTM Multilayer perceptron (MLP) | One-hot encoded and embedded with either LSTM or Autoencoder | 2020 | 580,299 (Autoencoder) or 6,557,421 (LSTM) |
| NetTCR2.0 Montemurro et al. (12) | CNN | Both sequences were encoded using the BLOSUM50 matrix | 2021 | 21,345 |
| ImRex Moris et al. (15) | CNN, L2 regularization penalty of 0.01. Dual-input CNN architecture | PCP interaction map between CDR3 and peptide sequence with 20x11x4 dimensions. | 2020 | 248,257 |