Optimization of a Deep-Learning Method Based on the Classification of Images Generated by Parameterized Deep Snap a...

Area of Science:

• Computational toxicology
• Endocrinology
• Machine learning

Background:

• Chemicals can disrupt endocrine system homeostasis by affecting hormone receptors.
• Traditional chronic toxicity testing is costly and time-consuming.
• Computational methods, particularly deep learning (DL), offer promising alternatives for toxicity prediction.

Purpose of the Study:

• To evaluate the impact of DeepSnap DL parameters on prediction model performance.
• To identify optimal parameter thresholds for enhanced chemical toxicity assessment.
• To build a reliable DeepSnap-DL model for predicting CAR agonist activity.

Main Methods:

• Utilized the DeepSnap DL technique for Quantitative Structure-Activity Relationship (QSAR) analysis on 3D chemical structures.
• Investigated the influence of DeepSnap parameters (e.g., molecule split, zoom factor, atom size, bond radius/distance/tolerance) on validation loss.
• Trained and validated models using the Tox21 quantitative high-throughput screening (qHTS) database.

Main Results:

• Parameter tuning revealed quadratic relationships between DeepSnap parameters and validation loss, indicating optimal thresholds.
• A DeepSnap-DL model optimized with best-performing parameters achieved an Area Under the Curve (AUC) of 0.791 for CAR agonist prediction.
• The study demonstrates the effectiveness of parameter optimization for improving DL-based toxicity prediction accuracy.

Conclusions:

• Optimized DeepSnap parameters significantly enhance the reliability of DL models for chemical toxicity prediction.
• The proposed DeepSnap-DL approach provides a cost-effective and efficient tool for assessing risks of various chemicals.
• This method holds potential for broader applications in endocrine disruptor screening and risk assessment.