Predicting Chemical Carcinogens Using a Hybrid Neural Network Deep Learning Method.

Area of Science:

• Environmental Science
• Toxicology
• Computational Chemistry

Background:

• Human exposure to environmental chemicals necessitates accurate carcinogenicity assessment.
• Existing methods for predicting chemical carcinogenicity require improvement for diverse chemical classes.

Purpose of the Study:

• To develop a novel hybrid neural network (HNN) method, HNN-Cancer, for predicting the carcinogenic potential of real-life chemicals.
• To evaluate the performance of HNN-Cancer against other machine learning models for binary classification, multiclass classification, and regression tasks.

Main Methods:

• Developed HNN-Cancer using a modified 3D array representation of 1D SMILES simulated by convolutional neural networks (CNN).
• Implemented binary and multiclass classification models using HNN-Cancer, random forest (RF), bootstrap aggregating (Bagging), and adaptive boosting (AdaBoost).
• Developed regression models using HNN-Cancer, RF, support vector regressor (SVR), gradient boosting (GB), and other machine learning algorithms.

Main Results:

• HNN-Cancer achieved 74% accuracy and ~0.81 AUC for binary classification of 7994 chemicals, outperforming other methods with 79.5% sensitivity and 67.3% specificity.
• For multiclass classification (1618 chemicals), HNN-Cancer, RF, Bagging, and AdaBoost achieved 70% accuracy and 0.7 AUC.
• HNN-Cancer and RF demonstrated a correlation coefficient (R) of ~0.62 in regression models, surpassing other tested machine learning methods.

Conclusions:

• The HNN-Cancer method demonstrates superior performance in predicting chemical carcinogenicity across various datasets and chemical classes.
• HNN-Cancer's predictive capabilities are comparable to existing literature models, even with more diverse chemical inputs.
• HNN-Cancer offers a valuable tool for identifying potentially carcinogenic chemicals, contributing to enhanced environmental and public health safety.