Table 5.
Prediction results of hERG blockers/nonblockers classification models developed by capsule networks with different architectures.
| Capsule network architecture | SE | SP | MCC | SD | Q (%) |
|---|---|---|---|---|---|
| Original CapsNet | 80.4% | 86.7% | 0.673 | 0.0141 | 84.1% |
| FC+FC | 82.6% | 86.7% | 0.694 | 0.0195 | 85.0% |
| Conv+FC | 82.2% | 86.4% | 0.687 | 0.0166 | 84.6% |
| Conv+FC+FC (Conv-CapsNet) | 88.6% | 89.1% | 0.774 | 0.0109 | 88.9% |
| Conv+Conv+FC+FC | 84.5% | 85.3% | 0.693 | 0.0142 | 84.9% |
| Conv+Conv+Conv+FC+FC | 81.9% | 86.9% | 0.685 | 0.0173 | 84.9% |
| One RBM | 83.1% | 86.5% | 0.694 | 0.0182 | 84.9% |
| Two RBMs (RBM-CapsNet) | 84.3% | 89.0% | 0.734 | 0.0160 | 87.0% |
| Three RBMs | 84.5% | 85.5% | 0.696 | 0.0160 | 85.0% |
| Four RBMs | 81.2% | 86.0% | 0.673 | 0.0108 | 83.9% |
| Five RBMs | 84.1% | 86.4% | 0.701 | 0.0156 | 85.4% |
*Conv, convolutional operation; FC, fully connected operation; RBM, restricted Boltzmann machine; Conv-CapsNet, convolution-capsule network; RBM-CapsNet, restricted Boltzmann machine-capsule network (The training set used was the Doddareddy's training set, and five-fold cross-validation was used to monitor the training performance. SE (%), sensitivity; SP (%), specificity; MCC, Matthew's correlation coefficient; SD, standard deviation; Q (%), overall accuracy). Conv-CapsNet and Conv-CapsNet showed the best performance.