Table 1 DDI prediction performance of various machine learning models with different drug features as input. The p value compared with using GCAN features is added in brackets
Method | Feature | Macro-F1 | Macro-recall | Macro-precision |
|---|---|---|---|---|
DNN | Original | 90.1% ± 1.9% (0.001) | 90.7% ± 1.8% (0.0051) | 91.3% ± 2.3% (0.009) |
|  | Autoencoder | 91.3% ± 0.7% (0.0655) | 90.8% ± 0.9% (0.0223) | 93.2% ± 1.1% (0.6219) |
|  | GCAN | 93.3% ± 1.4% (–) | 93.9% ± 1.7% (–) | 93.7% ± 1.4% (–) |
Random forest | Original | 40.7% ± 1.8% (4E − 05) | 35.7% ± 1.5% (4.3E − 05) | 58.6% ± 1.4% (0.0008) |
|  | Autoencoder | 45.2% ± 2% (0.0004) | 39.9% ± 1.9% (0.0004) | 62.9% ± 2.3% (0.001) |
|  | GCAN | 57.6% ± 3% (–) | 51.6.9% ± 2.9% (–) | 75.7% ± 4.2% (–) |
MLKNN | Original | 40.5% ± 1.2% (1.2E − 05) | 34.7% ± 1.1% (1E − 05) | 54.9% ± 2.4% (2.9E − 05) |
|  | Autoencoder | 51.5% ± 1.5% (5.5E − 05) | 46.5% ± 1.9% (0.0001) | 63.5% ± 2% (6.6E − 06) |
|  | GCAN | 74.3% ± 2.1% (–) | 70.3% ± 1.9% (–) | 83.4% ± 2.2% (–) |
BRkNNaClassifier | Original | 29.9% ± 1.7% (1E − 05) | 23.4% ± 1.5% (9E − 06) | 52.2% ± 2.8% (4.2E − 05) |
|  | Autoencoder | 39.1% ± 1.3% (4.4E − 05) | 32.3% ± 1.3% (2.7E − 05) | 59.2% ± 2.1% (0.0003) |
|  | GCAN | 67.5% ± 2.4% (–) | 61.1% ± 2.4% (–) | 83.4% ± 3.3% (–) |