AI-Driven Nano-QSAR Framework for Predicting Carbon Nanotube Cytotoxicity: Overcoming High Dimensionality and Data...

Reliable prediction of nanomaterial toxicity from physicochemical properties remains a critical challenge in nanosafety assessment, particularly when experimental datasets are small and mechanistic validation is limited. This study proposes an artificial intelligence (AI)-driven nano-QSAR framework for predicting carbon nanotube (CNT)-induced cytotoxicity by systematically linking the physicochemical descriptor patterns with experimentally measured toxicity outcomes.

A curated dataset containing cytotoxicity measurements for 80 CNTs in THP-1 cells was analyzed. Each CNT was represented by 2142 physicochemical descriptors covering structural, surface, oxidation-related, and electronic properties. Descriptor redundancy and effective dimensionality were systematically examined using correlation analysis and principal component analysis (PCA). Models were implemented in Python 3.13 using scikit-learn, CatBoost, XGBoost, Optuna, and imbalanced-learn. Multiple classifiers were benchmarked within an imbalance-aware machine-learning framework, and performance was evaluated using stratified five-fold cross-validation and an independent held-out test set.

The descriptor space was highly compressible, with the first two principal components explaining 99.25% of the total variance. Descriptor-level analyses indicated that SCNO-, CCOX-, and COOX-related descriptor families, reflecting surface chemistry, oxidation-related properties, and electronic structure, were major contributors to cytotoxicity-associated variation. Among the evaluated algorithms, CatBoost consistently exhibited the best balance between overall predictive performance and minority-class detection, achieving an F1-score of 0.9086 and a balanced accuracy of 0.9615 on the independent test set.

This study demonstrates that robust prediction of CNT cytotoxicity can be achieved through the integration of low-dimensional descriptor-space analysis, hybrid feature engineering, and imbalance-aware machine learning. The results further suggest that the apparent complexity of CNT descriptor space can be effectively reduced to a limited number of physicochemical dimensions associated with biological responses. By establishing a structured modeling framework that prioritizes hazard identification over retrospective curve-fitting, this approach functions as a conservative, early-stage screening tool. Although broader external validation remains necessary, the consistent importance of surface chemistry, oxidation-related characteristics, and electronic structure supports their central role in regulating nano-bio interactions and cytotoxic responses. These findings establish a practical foundation for AI-assisted nanosafety assessment and mechanism-informed safer-by-design development of carbon nanomaterials.