Tabular Foundation Models for Environmental Quantitative Structure–Activity Relationship: Robust Prediction with Imperfect Data

Environmental quantitative structure-activity relationship (QSAR) modeling is frequently constrained by small sample sizes, missing values, and end point imbalance. To address these limitations, we propose a modular framework based on the TabPFN family that integrates zero-/few-shot prediction, missing robust probabilistic inference, and generative augmentation. Validated across hydroxyl radical rate prediction, singlet oxygen kinetics, and toxicity classification tasks, the framework demonstrates strong data efficiency. Data utility curves combined with a dynamic efficient-window identification strategy provide data-driven guidance for sample size selection. The framework also maintains stable predictive performance and preserves key feature identification under moderate missingness. Under extreme data scarcity (<100 samples) and severe class imbalance, the generative module improves certain conventional classifiers, outperforming GMM-based augmentation and achieving performance comparable to SMOTE while providing limited benefit for TabPFN itself. SHAP and PDP analyses further reveal chemically meaningful trends consistent with current mechanistic knowledge. Overall, this study provides a flexible QSAR workflow for reliable, interpretable, and cost-aware modeling under real-world data constraints.