A deep learning-assisted whole-cell biosensor for catechol monitoring: Synergizing BphC recognition with a multi-feature fusion vision transformer.

Catechol (Cat) quantification typically requires benchtop instrumentation, restricting its application for on-site environmental analysis. To address this, we developed a portable image-based biosensing platform that integrates a genetically engineered whole-cell biocatalyst with a Multi-Feature Fusion Vision Transformer (MF-ViT). An Escherichia coli strain expressing 2,3-dihydroxybiphenyl-1,2-dioxygenase (BphC) was constructed to specifically convert Cat into a measurable yellow product. Molecular docking and molecular dynamics simulations provided a structural rationale for the sensor's analytical selectivity, demonstrating that specific bidentate coordination requirements prevent structural analogs from triggering false-positive signals. For signal decoding under ambient lighting, the MF-ViT model combined global visual representations with predefined colorimetric statistical features derived from the solution and background regions. This approach outperformed conventional simple RGB and convolutional neural networks (CNNs) analysis, achieving a high predictive accuracy (R2 = 0.944). The integrated biosensor exhibited a linear range of 12.5 to 400 nM and a limit of detection of 1.66 nM. Spiked recovery tests in seawater and tap water demonstrated that the smartphone-derived predictions were statistically consistent with standard spectrophotometric measurements. This approach provides a practical, algorithm-assisted strategy for the rapid screening of phenolic pollutants.