Data-Driven Toxicity Prediction: Advances in Machine Learning, Deep Learning, and Predictive Tools - A Systematic...

Area of Science:

• Computational toxicology
• Drug discovery and development
• Environmental safety

Background:

• High attrition rates in drug discovery are often due to toxicity.
• Machine Learning (ML) and Deep Learning (DL) offer advanced solutions for toxicity prediction.
• Emerging post-deep learning strategies are also being explored.

Purpose of the Study:

• To review ML, DL, and post-deep learning strategies for toxicity prediction in drug discovery and environmental safety.
• To summarize advancements in computational toxicology.
• To highlight the role of predictive models in reducing drug development failures.

Main Methods:

• Systematic literature search following PRISMA guidelines (2015-2025).
• Inclusion of studies using ML/DL for toxicity prediction with quantitative performance measures.
• Analysis of web-based prediction tools and platforms.

Main Results:

• Deep learning models (CNNs, GNNs) show high efficacy with large datasets.
• Techniques like transfer learning address data scarcity.
• Web-based applications offer multi-endpoint toxicity predictions with accuracy and interpretability.

Conclusions:

• Traditional ML methods (SVMs, Random Forests) are robust for smaller datasets.
• Deep learning architectures significantly improve predictive accuracy on large, complex datasets.
• Data-driven methods accelerate toxicity prediction but face challenges in data quality and interpretability.