Tabular Foundation Models for Environmental Quantitative Structure-Activity Relationship: Robust Prediction with...

Area of Science:

• Environmental chemistry
• Computational toxicology
• Machine learning for science

Background:

• Environmental quantitative structure-activity relationship (QSAR) modeling faces challenges like small sample sizes, missing data, and imbalanced endpoints.
• Developing robust QSAR models is crucial for predicting chemical properties and environmental impact efficiently.

Purpose of the Study:

• To propose a modular framework addressing limitations in environmental QSAR modeling.
• To enhance data efficiency, handle missing values, and manage class imbalance in QSAR predictions.
• To provide data-driven guidance for sample size selection and ensure model interpretability.

Main Methods:

• A modular framework based on the TabPFN family, integrating zero-/few-shot prediction, robust probabilistic inference for missing data, and generative augmentation.
• Validation across hydroxyl radical rate prediction, singlet oxygen kinetics, and toxicity classification tasks.
• Utilizing data utility curves and a dynamic efficient-window identification strategy for sample size guidance.
• Employing SHAP and Partial Dependence Plots (PDP) for model interpretability and mechanistic insight.

Main Results:

• The framework demonstrated strong data efficiency across various environmental QSAR tasks.
• Stable predictive performance and key feature identification were maintained even with moderate missing data.
• The generative module improved conventional classifiers under extreme data scarcity and class imbalance, comparable to SMOTE.
• SHAP and PDP analyses revealed chemically meaningful trends aligned with existing mechanistic knowledge.

Conclusions:

• The proposed framework offers a flexible and reliable QSAR workflow for real-world data constraints.
• It enables cost-aware and interpretable modeling, particularly valuable when data is limited or imbalanced.
• The study provides practical tools for optimizing sample size selection and enhancing model robustness.