Machine Learning Prediction of Analyte-Induced Fluorescence Perturbations in DNA-Functionalized Carbon Nanotubes

Abstract

Single-walled carbon nanotubes (SWCNTs) functionalized with single-stranded DNAs can function as near-infrared nanosensors for molecular analytes. However, predicting which analytes elicit strong optical responses for specific nanosensors remains challenging. We developed machine learning (ML) models to predict analyte-induced fluorescence changes in a DNA–SWCNT dopamine nanosensor. Using a data set of 63 small molecules sampling chemical space around dopamine, we encoded analytes with RDKit fingerprints, with or without HOMO and LUMO energies, and applied principal component analysis to identify structural motifs associated with optical response strength. We trained support vector regression and classification models using two strategies: ensembles of 200 models and cross-validation. Regression models achieved mean R2 values of 0.2–0.4, with cross-validation outperforming ensembles, while classifiers reached mean F1 scores of ∼0.8. Cross-validation performed best for predictions on a blind set of 21 molecules. These findings show that ML can capture structure–response patterns in modest data sets and guide in silico DNA–SWCNT nanosensor design.