TISBE: A Public Web Platform for the Consensus-Based Explainable Prediction of Developmental Toxicity.

Area of Science:

• Computational toxicology
• cheminformatics
• predictive toxicology

Background:

• Developmental toxicity (Dev Tox) is crucial for prenatal and neonatal health but poorly understood at the molecular level.
• Limited high-quality data and a lack of robust non-testing methods hinder Dev Tox research.
• Dev Tox is a significant area of animal testing in regulatory toxicology, necessitating alternative methods.

Purpose of the Study:

• To introduce TISBE (TIRESIA Improved on Structure-Based Explainability), a novel web platform for in silico developmental toxicity analysis.
• To provide an explainable artificial intelligence (XAI) framework for understanding chemical toxicity.
• To offer a reliable tool for predicting developmental toxicity and identifying responsible molecular fragments.

Main Methods:

• Development of TISBE, incorporating a larger dataset (1008 chemicals), a fragment-based XAI framework, a consensus classifier (5 ML models), and a novel applicability domain (AD) method.
• Training and validation using a 5-fold cross-validation approach.
• Application of TISBE to external datasets of bioactive compounds (ChEMBL) and pediatric drugs (TEDDY).

Main Results:

• TISBE achieved a median Matthews correlation coefficient of 0.410.
• High median sensitivity (0.984) and specificity (0.274) were observed during cross-validation.
• The platform successfully identified molecular fragments associated with toxicity or safety in tested compounds.

Conclusions:

• TISBE represents a significant advancement in in silico developmental toxicity assessment, offering explainability and improved reliability.
• The platform's ability to pinpoint toxic molecular fragments aids in understanding chemical risks.
• TISBE is freely available, supporting regulatory toxicology and drug development for enhanced child safety.