Table 5 CNN hyperparameter space explored with hg19 and hg38 data through Bayesian optimization. A. Architecture and learning hyperparameters of the fixed-CNN; B. Architecture and hyperparameter space of the Bayesian-CNN models trained on the hg19 dataset; C. Architecture and hyperparameter space of the Bayesian-CNN models models trained on the hg38 dataset
Layers | Type | Units | Kernel | Activation | Notes |
|---|---|---|---|---|---|
A: fixed-CNN | |||||
3 | Convolutional | 64 | 5 | ReLU | – |
1 | Max pooling 1D | – | – | – | size 2 |
3 | Convolutional | 128 | 3 | ReLU | – |
1 | Max pooling 1D | – | – | – | size 2 |
3 | Convolutional | 128 | 3 | ReLU | – |
1 | Average pooling 1D | – | – | – | – |
1 | Dropout | – | – | – | Probability 0.5 |
2 | Dense | 10 | – | ReLU | – |
1 | Dropout | – | – | – | Probability 0.5 |
1 | Dense | 1 | – | Sigmoid | – |
Learning parameters | |||||
Learning rate | 0.002 | ||||
Batch size | 256 | ||||
Optimizer | Nadam | ||||
Epochs | 100 | ||||
Layers | Type | Units | Kernel | Activation | Notes |
|---|---|---|---|---|---|
Hyperparameter space | Hyperparameter space | ||||
B: Bayesian-CNN (hg19 dataset) | |||||
3 | Convolutional + batch norm | {32, 64, 128} | 5 | ReLU | – |
1 | Max pooling 1D | – | – | – | Size 2 |
1 | Convolutional + batch norm | {32, 64, 128} | {5, 10} | ReLU | – |
1 | Max pooling 1D | – | – | – | Size 2 |
1 | Flatten | – | – | – | – |
1 | Dense | {10, 32, 64} | – | ReLU | – |
1 | Dropout | – | - - | – | Probability 0.1 |
1 | Dense | {10, 32, 64} | – | ReLU | – |
1 | Dropout | – | – | – | Probability 0.1 |
1 | Dense | 1 | – | Sigmoid | – |
Learning parameters | |||||
Learning rate | 0.002 | ||||
Batch size | 256 | ||||
Optimizer | Nadam | ||||
Epochs | 100 | ||||
Layers | Hyperparameter space | Activation | |||
|---|---|---|---|---|---|
C: Bayesian-CNN (hg38 dataset) | |||||
No. of convolutional groups | \([0 \ldots 2]\) | ||||
No. of hidden convolutional layers, composing the group | \(\{0, \ldots , 3\}\) | ReLU | |||
No. of filters in the convolutional layer | \([0 \ldots 128]\) | ||||
2D kernel size in the convolutional layer | \([2 \ldots 8] \times [1,2]\) | ||||
Max pooling 2D | \([1 \ldots 8] \times [1,2]\) | ||||
Dropout | \([0 \ldots 0.5]\) | ||||
No. of dense groups | \([0 \ldots 2]\) | ||||
No. of hidden dense layers, composing the group | \(\{0, \ldots , 3\}\) | ReLU | |||
No. of units in dense layer | \([0 \ldots 64]\) | ||||
Dropout | \([0 \ldots 0.5]\) | ||||
Output | 1 | Sigmoid | |||
Learning parameters | |||||
Learning rate | 0.002 | ||||
l1 regularizer | 0.0001 | ||||
l2 regularizer | 0.0001 | ||||
Batch size | 256 | ||||
Optimizer | Nadam | ||||
Epochs | 100 | ||||