We report a conjugated polyelectrolyte fluorescence-based biosensor P-C-3 and a general methodology to evaluate spectral shape recognition to identify biomolecules using artificial intelligence. By using well-defined analytes, we demonstrate that the fluorescence spectral shape of P-C-3 is sensitive to minor structural changes and exhibits distinct signature patterns for different analytes. A method was also developed to select useful features to reduce computational complexity and prevent overfitting of the data. It was found that the normalized intensity of 3 to 5 selected wavelengths was sufficient for the fluorescence biosensor to classify 13 distinct nucleotides and distinguish as little as single base substitutions at distinct positions in the primary sequence of oligonucleotides rapidly with nearly 100% classification accuracy. Photophysical studies led to a model to explain the mechanism of these fluorescence spectral shape changes, which provides theoretical support for applying this method in complicated biological systems. Using the feature selection algorithm to measure the relative intensity of a few selected wavelengths significantly reduces measurement time, demonstrating the potential for fluorescence spectrum shape analysis in high-throughput and high-content screening.
Keywords: biosensor; fluorescence; machine learning; nucleic acids; phosphate sensing.