Data Augmentation via Generative Adversarial Networks for Robust Prediction of NADH and pH from Limited Time-resolved Fluorescence Spectra

Circularly permuted yellow fluorescent protein (cpYFP)-based biosensors are powerful tools for biological sensing but are inherently pH-sensitive. This complicates quantitative analysis because the fluorescence signal convolves the effects of the target analyte with ambient pH fluctuations, hindering accurate disentanglement of their individual contributions. Achieving simultaneous monitoring of analyte concentration and pH is therefore essential to unlock the full potential of these biosensors. Here, using the cpYFP-based NADH biosensor SoNar as a model, we present a deep learning framework for the concurrent determination of NADH concentration and pH from its time-resolved fluorescence spectra. We address the critical issue of limited experimental data by developing a hybrid Generative Adversarial Network (GAN) augmented with a Long Short-Term Memory (LSTM) module for high-fidelity spectral data augmentation. This approach significantly enhances prediction accuracy. Our work establishes a robust strategy for dual-parameter quantification with cpYFP-based biosensors and provides a data-efficient solution that reduces reliance on extensive experimental trials.