Multi-tasking (drug response, cell line classification, etc.) Uno Implemented in PyTorch.
- Error analysis for drug response (what kind of drugs/cells are more prone to larger errors);
- If we limit our predictions to only those of high-confidence how much does it improve our prediction skill?
- Outlier removal (and possible effect on performance);
- Inferencing with quantization (integer or half-precision);
- Hyper-parameter searching;
Python 3.6.4
PyTorch 0.4.1
SciPy 1.1.0
pandas 0.23.4
Scikit-Learn 0.19.1
urllib3 1.23
joblib 0.12.2
The default network structure is shown below:
An example of the program output for training on NCI60 and valdiation on all other data sources is shown below:
python3.6 ./launcher.py
Training Arguments:
{
"trn_src": "NCI60",
"val_srcs": [
"NCI60",
"CTRP",
"GDSC",
"CCLE",
"gCSI"
],
"grth_scaling": "none",
"dscptr_scaling": "std",
"rnaseq_scaling": "std",
"dscptr_nan_threshold": 0.0,
"qed_scaling": "none",
"rnaseq_feature_usage": "source_scale",
"drug_feature_usage": "both",
"validation_ratio": 0.2,
"disjoint_drugs": false,
"disjoint_cells": true,
"gene_layer_dim": 1024,
"gene_latent_dim": 512,
"gene_num_layers": 2,
"drug_layer_dim": 4096,
"drug_latent_dim": 2048,
"drug_num_layers": 2,
"autoencoder_init": true,
"resp_layer_dim": 2048,
"resp_num_layers_per_block": 2,
"resp_num_blocks": 4,
"resp_num_layers": 2,
"resp_dropout": 0.0,
"resp_activation": "none",
"cl_clf_layer_dim": 256,
"cl_clf_num_layers": 2,
"drug_target_layer_dim": 512,
"drug_target_num_layers": 2,
"drug_qed_layer_dim": 512,
"drug_qed_num_layers": 2,
"drug_qed_activation": "sigmoid",
"resp_loss_func": "mse",
"resp_opt": "SGD",
"resp_lr": 1e-05,
"cl_clf_opt": "SGD",
"cl_clf_lr": 0.01,
"drug_target_opt": "SGD",
"drug_target_lr": 0.01,
"drug_qed_loss_func": "mse",
"drug_qed_opt": "SGD",
"drug_qed_lr": 0.01,
"resp_val_start_epoch": 0,
"early_stop_patience": 20,
"lr_decay_factor": 0.98,
"trn_batch_size": 32,
"val_batch_size": 256,
"max_num_batches": 1000,
"max_num_epochs": 1000,
"multi_gpu": false,
"no_cuda": false,
"rand_state": 0
}
RespNet(
(_RespNet__gene_encoder): Sequential(
(dense_0): Linear(in_features=942, out_features=1024, bias=True)
(relu_0): ReLU()
(dense_1): Linear(in_features=1024, out_features=1024, bias=True)
(relu_1): ReLU()
(dense_2): Linear(in_features=1024, out_features=512, bias=True)
)
(_RespNet__drug_encoder): Sequential(
(dense_0): Linear(in_features=4688, out_features=4096, bias=True)
(relu_0): ReLU()
(dense_1): Linear(in_features=4096, out_features=4096, bias=True)
(relu_1): ReLU()
(dense_2): Linear(in_features=4096, out_features=2048, bias=True)
)
(_RespNet__resp_net): Sequential(
(dense_0): Linear(in_features=2561, out_features=2048, bias=True)
(activation_0): ReLU()
(residual_block_0): ResBlock(
(block): Sequential(
(res_dense_0): Linear(in_features=2048, out_features=2048, bias=True)
(res_relu_0): ReLU()
(res_dense_1): Linear(in_features=2048, out_features=2048, bias=True)
)
(activation): ReLU()
)
(residual_block_1): ResBlock(
(block): Sequential(
(res_dense_0): Linear(in_features=2048, out_features=2048, bias=True)
(res_relu_0): ReLU()
(res_dense_1): Linear(in_features=2048, out_features=2048, bias=True)
)
(activation): ReLU()
)
(residual_block_2): ResBlock(
(block): Sequential(
(res_dense_0): Linear(in_features=2048, out_features=2048, bias=True)
(res_relu_0): ReLU()
(res_dense_1): Linear(in_features=2048, out_features=2048, bias=True)
)
(activation): ReLU()
)
(residual_block_3): ResBlock(
(block): Sequential(
(res_dense_0): Linear(in_features=2048, out_features=2048, bias=True)
(res_relu_0): ReLU()
(res_dense_1): Linear(in_features=2048, out_features=2048, bias=True)
)
(activation): ReLU()
)
(dense_1): Linear(in_features=2048, out_features=2048, bias=True)
(res_relu_1): ReLU()
(dense_2): Linear(in_features=2048, out_features=2048, bias=True)
(res_relu_2): ReLU()
(dense_out): Linear(in_features=2048, out_features=1, bias=True)
)
)
================================================================================
Training Epoch 1:
Drug Weighted QED Regression Loss: 0.055694
Drug Response Regression Loss: 1871.18
Validation Results:
Cell Line Classification:
Category Accuracy: 98.98%;
Site Accuracy: 80.95%;
Type Accuracy: 82.76%
Drug Target Family Classification Accuracy: 1.85%
Drug Weighted QED Regression
MSE: 0.028476 MAE: 0.137004 R2: +0.17
Drug Response Regression:
NCI60 MSE: 1482.07 MAE: 27.89 R2: +0.53
CTRP MSE: 2554.45 MAE: 38.62 R2: +0.27
GDSC MSE: 2955.78 MAE: 42.73 R2: +0.11
CCLE MSE: 2799.06 MAE: 42.44 R2: +0.31
gCSI MSE: 2601.50 MAE: 38.44 R2: +0.35
Epoch Running Time: 110.0 Seconds.
================================================================================
Training Epoch 2:
...
...
Program Running Time: 8349.6 Seconds.
================================================================================
Overall Validation Results:
Best Results from Different Models (Epochs):
Cell Line Categories Best Accuracy: 99.474% (Epoch = 5)
Cell Line Sites Best Accuracy: 97.401% (Epoch = 60)
Cell Line Types Best Accuracy: 97.368% (Epoch = 40)
Drug Target Family Best Accuracy: 66.667% (Epoch = 23)
Drug Weighted QED Best R2 Score: +0.7422 (Epoch = 59, MSE = 0.008837, MAE = 0.069400)
NCI60 Best R2 Score: +0.8107 (Epoch = 56, MSE = 601.18, MAE = 16.57)
CTRP Best R2 Score: +0.3945 (Epoch = 37, MSE = 2127.28, MAE = 31.44)
GDSC Best R2 Score: +0.2448 (Epoch = 22, MSE = 2506.03, MAE = 35.55)
CCLE Best R2 Score: +0.4729 (Epoch = 4, MSE = 2153.30, MAE = 33.63)
gCSI Best R2 Score: +0.4512 (Epoch = 31, MSE = 2203.04, MAE = 32.63)
Best Results from the Same Model (Epoch = 22):
Cell Line Categories Accuracy: 99.408%
Cell Line Sites Accuracy: 97.138%
Cell Line Types Accuracy: 97.039%
Drug Target Family Accuracy: 57.407%
Drug Weighted QED R2 Score: +0.6033 (MSE = 0.013601, MAE = 0.093341)
NCI60 R2 Score: +0.7885 (MSE = 672.00, MAE = 17.89)
CTRP R2 Score: +0.3841 (MSE = 2163.66, MAE = 32.28)
GDSC R2 Score: +0.2448 (MSE = 2506.03, MAE = 35.55)
CCLE R2 Score: +0.4653 (MSE = 2184.62, MAE = 34.12)
gCSI R2 Score: +0.4271 (MSE = 2299.59, MAE = 32.93)
For default hyper parameters, the transfer learning matrix results are shown below:
Note that the green cells represents R2 score of higher than 0.1, red cells are R2 scores lower than -0.1 and yellows are for all the values in between.