Deep neural network-based detection of lead contamination via Förster resonance energy transfer in live cells

Lead (Pb), a heavy metal with extensive industrial applications, poses significant risks to human health and the environment. These detrimental effects have driven strict global and local regulations, such as World Health Organization and Taiwan's drinking water limits, to minimize exposure and protect public health. To ensure compliance with these regulatory standards, we engineered lead ion biosensors by fusing microbial metal-binding proteins with cyan (CFP) and yellow (YFP) fluorescent proteins. These biosensors undergo a lead-induced conformational change in live cells, bringing the donor and acceptor fluorophores into close proximity. This enables Förster Resonance Energy Transfer (FRET), resulting in a quantifiable fluorescence shift from the donor's blue emission to the acceptor's green emission. We have prototyped a portable device integrating this biosensor for field-deployable monitoring; however, precise quantification of fluorescent signals from smartphone-captured images remains a key challenge. To overcome this, we employed deep learning to analyze fluorescent signals in images and classify lead contamination levels relative to regulatory thresholds. Live-cell imaging was performed on HEK293T cells transfected with FRET-based biosensors. Lead with a range of concentrations were introduced to evaluate the performance of the biosensors. Fluorescence microscopy validated that 30- and 60-min incubation periods enable sensitive detection of lead concentrations, spanning levels both below and above regulatory limits. We meticulously acquired an extensive dataset of 1131 image pairs using our portable device and a smartphone, with images captured after 30- and 60-min reaction times. These images were then systematically partitioned into training and validation datasets to rigorously evaluate the AI model. A convolutional neural network with the EfficientNet architecture was developed and trained by these images. During the 80 training epochs, performance of the AI model improved consistently, shown by decreasing loss function values and increasing Youden's J statistic. Validation on independent image sets yielded area under the receiver operating characteristic curves (AUCs) of 72.0 % and 69.8 % for the 60- and 30-min reaction times, respectively. In conclusion, we proposed a novel combination of FRET biosensors with an EfficientNet-based model for field-deployable lead detection. The feasibility of this approach is supported by rigorous validation, demonstrating the platform's ability to accurately and robustly classify lead contamination levels relative to regulatory thresholds in a field-deployable context.