Deep-learning: investigating deep neural networks hyper-parameters and comparison of performance to shallow methods...

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Deep learning (DL) has gained popularity in artificial neural networks (ANNs).
• DL techniques show promise in quantitative structure-activity relationship (QSAR) and predictive toxicology.
• This study investigates DL performance in drug discovery and toxicology research.

Purpose of the Study:

• To explore hyper-parameter configurations for deep neural networks (DNNs).
• To compare DNN performance against established cheminformatics methods.
• To assess the robustness of machine learning models to noise.

Main Methods:

• Utilized the Caffe deep-learning framework and NVidia GPUs.
• Explored numerous DNN hyper-parameter configurations.
• Assessed model performance using statistical tests (Wilcoxon) and Matthews Correlation Coefficient (MCC).

Main Results:

• Optimized feed-forward DNNs demonstrated superior classification performance across various activity classes.
• Key hyper-parameters influencing DNN performance include activation function, dropout, number of layers, and neurons.
• Tuned DNNs statistically outperformed Naïve Bayes, kNN, RF, and SVM (p < 0.01).
• DNNs achieved higher average MCC values compared to other methods.
• Non-linear methods handled low noise (≤20%) well; Naïve Bayes excelled with higher noise levels (>30%).

Conclusions:

• Optimized DNNs are powerful tools for drug discovery and toxicology.
• Hyper-parameter tuning is critical for achieving high DNN performance.
• Model selection depends on noise levels, with Naïve Bayes being robust to significant noise.