Toxics, Vol. 10, Pages 706: Predicting Dose-Range Chemical Toxicity using Novel Hybrid Deep Machine-Learning Method

Figure 1. Schematic diagram of the Hybrid Neural Network (HNN) consisting of Convolutional Neural Network (CNN) and Feed-Forward Neural Network (FFNN). L, length of the SMILES string; M; N, number of filters (possibly different at each layer).

Figure 1. Schematic diagram of the Hybrid Neural Network (HNN) consisting of Convolutional Neural Network (CNN) and Feed-Forward Neural Network (FFNN). L, length of the SMILES string; M; N, number of filters (possibly different at each layer).

Toxics 10 00706 g001

Figure 2. The overall flowchart of the HNN-Tox toxicity prediction processes with detailed data preparation for each data source is indicated (see Methods section for details).

Figure 2. The overall flowchart of the HNN-Tox toxicity prediction processes with detailed data preparation for each data source is indicated (see Methods section for details).

Toxics 10 00706 g002

Figure 3. (A) Accuracy, (B) AUC, (C) sensitivity, and (D) specificity for the ChemIDplus oral data as predicted by HNN, RF, Bagging, AdaBoost, and the Ensemble methods with additional descriptors from ADMETlab and Canvas.

Figure 3. (A) Accuracy, (B) AUC, (C) sensitivity, and (D) specificity for the ChemIDplus oral data as predicted by HNN, RF, Bagging, AdaBoost, and the Ensemble methods with additional descriptors from ADMETlab and Canvas.

Toxics 10 00706 g003

Figure 4. (A) Accuracy, (B) AUC, (C) sensitivity, and (D) specificity for the ChemIDplus IP/IV/Sub/Oral data as given by HNN, RF, Bagging, AdaBoost, and the Ensemble methods.

Figure 4. (A) Accuracy, (B) AUC, (C) sensitivity, and (D) specificity for the ChemIDplus IP/IV/Sub/Oral data as given by HNN, RF, Bagging, AdaBoost, and the Ensemble methods.

Toxics 10 00706 g004

Figure 5. Predicted accuracy of T3DB data obtained via oral route of exposure with cutoffs at (A) 250 mg/kg, (B) 500 mg/kg, (C) 750 mg/kg, and (D) 1000 mg/kg.

Figure 5. Predicted accuracy of T3DB data obtained via oral route of exposure with cutoffs at (A) 250 mg/kg, (B) 500 mg/kg, (C) 750 mg/kg, and (D) 1000 mg/kg.

Toxics 10 00706 g005

Figure 6. The AUC for toxins LD50 data obtained via oral route of exposure with various cutoffs at (A) 250 mg/kg, (B) 500 mg/kg, (C) 750 mg/kg, and (D) 1000 mg/kg by HNN, RF, Bagging, AdaBoost, and the Ensemble methods.

Figure 6. The AUC for toxins LD50 data obtained via oral route of exposure with various cutoffs at (A) 250 mg/kg, (B) 500 mg/kg, (C) 750 mg/kg, and (D) 1000 mg/kg by HNN, RF, Bagging, AdaBoost, and the Ensemble methods.

Toxics 10 00706 g006

Figure 7. Accuracy for the combined data (toxin data from T3DB + animal toxicity data from EPA) with cutoffs at (A) 250 mg/kg, (B) 500 mg/kg, (C) 750 mg/kg, and (D) 1000 mg/kg by HNN, RF, Bagging, AdaBoost, and the Ensemble methods.

Figure 7. Accuracy for the combined data (toxin data from T3DB + animal toxicity data from EPA) with cutoffs at (A) 250 mg/kg, (B) 500 mg/kg, (C) 750 mg/kg, and (D) 1000 mg/kg by HNN, RF, Bagging, AdaBoost, and the Ensemble methods.

Toxics 10 00706 g007

Figure 8. AUC for the combined data (toxin data from T3DB + animal toxicity data from EPA) with cutoffs at (A) 250 mg/kg, (B) 500 mg/kg, (C) 750 mg/kg, and (D) 1000 mg/kg by HNN, RF, Bagging, AdaBoost, and the Ensemble methods.

Figure 8. AUC for the combined data (toxin data from T3DB + animal toxicity data from EPA) with cutoffs at (A) 250 mg/kg, (B) 500 mg/kg, (C) 750 mg/kg, and (D) 1000 mg/kg by HNN, RF, Bagging, AdaBoost, and the Ensemble methods.

Toxics 10 00706 g008

Figure 9. Accuracy and AUC for the multiclass classification (A) toxins Oral data and (B) toxins oral data + Oversampling by HNN, RF, Bagging, AdaBoost, and the Ensemble method.

Figure 9. Accuracy and AUC for the multiclass classification (A) toxins Oral data and (B) toxins oral data + Oversampling by HNN, RF, Bagging, AdaBoost, and the Ensemble method.

Toxics 10 00706 g009

Table 1. Accuracy % and AUC for the ChemIDplus oral data with 51 descriptors and 318 descriptors.

Table 1. Accuracy % and AUC for the ChemIDplus oral data with 51 descriptors and 318 descriptors.

No of DescriptorsHNN-ToxRFBagAdaAccuracy %31884.9684.8485.6282.53 5184.1182.0782.0576.24AUC3180.8970.9010.9020.862 510.8870.8650.8610.765

Table 2. Accuracy and AUC of Tox21 challenge data with 801 descriptors and SMILES for 12 different assays.

Table 2. Accuracy and AUC of Tox21 challenge data with 801 descriptors and SMILES for 12 different assays.

AhRARAREAR-LBDAromataseATAD5ERER-LBDHSEMMPP53PPAR-GammaAccuracyHNN-Tox89.197.883.298.292.493.790.896.796.689.693.194.7RF90.497.983.598.392.893.991.997.196.590.593.394.7Bag89.897.983.198.492.894.691.096.496.891.193.394.7Ensemble90.397.983.698.492.893.991.297.196.690.793.494.7AUCHNN0.8860.7830.7750.7560.7690.7820.7510.7410.7740.9170.8280.733RF0.8950.7180.7680.7090.7720.7590.7690.7570.7670.9190.7840.699Bag0.8970.7410.7590.7180.7610.7490.7510.7660.7760.9080.7650.711Ensemble0.9050.7670.7860.7200.7880.7810.7670.7810.7920.9280.8010.730DeepTox0.9280.8070.8400.8790.8340.7930.8100.8140.8650.9420.8620.861

留言 (0)

沒有登入
gif