Machine-Learning-Assisted Fluorescence Kinetic Fingerprinting of Seven Nucleotides with a Single Eu(III)-Functionalized HOF

Reliable discrimination of structurally similar nucleotides remains challenging in fluorescence sensing because their shared molecular frameworks often result in overlapping fluorescence responses. Here, a Eu(III)-functionalized hydrogen-bonded organic framework (Eu@MA-BPDC) was used as a single cross-reactive sensing unit to generate multidimensional kinetic fluorescence fingerprints for discriminating seven nucleotides: ATP, ADP, AMP, GTP, CTP, UTP, and GMP. Fluorescence responses at 400, 600, and 621 nm were recorded sequentially from each sample using a 10 s local acquisition window at each wavelength, and the resulting fixed-order multiwavelength kinetic features were analyzed using a radial basis function kernel support vector machine (RBF-SVM). The model achieved 100% accuracy, balanced accuracy, and macro-F1 score in nested stratified five-fold cross-validation and correctly classified 35 independently prepared validation samples. Furthermore, ATP quantification was achieved through ratiometric fluorescence analysis based on the I621/I400 ratio, showing a linear response over 100-500 μM (R2 = 0.9987), with mean recoveries of 97.48-103.38% in the presence of the tested inorganic salts and small molecules. By combining the multiwavelength dynamic response of a single Eu@MA-BPDC unit with machine learning, this method provides a fluorescence sensing strategy for classifying structurally similar nucleotides while reducing reliance on multimaterial sensing arrays.