AI-Automated Portable FRET Platform for Ultra-Sensitive and Rapid On-Site Bacterial Lead Sensing

Authors:
  • YLYu-Sheng Liu
    Yu-Sheng Liu
  • PXPei-Yun Xie
    Pei-Yun Xie
  • HHHung-Chi Hsu
    Hung-Chi Hsu
  • WLWei-Qun Lai
    Wei-Qun Lai
  • YCYu-Fen Chang
    Yu-Fen Chang
  • PCPei-Ying Chiang
    Pei-Ying Chiang
  • YCYu-Ling Chiang
    Yu-Ling Chiang
  • MWMeng-Jer Wu
    Meng-Jer Wu

Abstract
Rapid on-site monitoring of lead (Pb) contamination is often limited by the fragility of mammalian-cell biosensors in an environmental setting and the manual image analysis that is labor intensive. This work demonstrates a stable, on-site portable sensing platform based on E. coli that expresses the high-sensitivity Met-lead 1.44 M1 FRET sensor and is encapsulated in a protective gelatin hydrogel matrix. To address data processing constraints, we present a novel AI-assisted pipeline combining feature extraction and Cellpose-SAM DL segmentation. This procedure for automated pixel-shift correction and high-accuracy ROI determination (with average precision, AP of 0.945 at threshold intersection over unit, IoU of 0.5) is integrated in this work allowing us to speed up analysis time drastically by a factor of over 30 compared to laborious manual processing. Our bacterial sensor has a high linearity (R2 > 0.97) and it performs consistently well in both aqueous and hydrogel forms. More importantly, the system exhibited a LOD of 2.0 ppb within 15 min, which is significantly below the WHO drinking water limit (10 ppb), and also all the functions remained intact after 4 h of warm environmental exposure. This AI-powered method can be used for both smartphone and microscopy imaging, allowing for real-time detection in a range of applications. Finally, a predictive “traffic-light” risk model was implemented for an easy interpretation of water quality. By closing the gap between laboratory-based sensing and field operation, this work provides a high throughput, real time, mobile solution for next generation environmental safety monitoring.

This publication has 35 references indexed in Scilit: