(M1530-06-35) Paper-Based Fluorophores and Machine Learning: A Versatile Tool for Accurate Identification of Pathogenic Bacteria

Purpose: Infectious diseases are the second leading cause of mortality in the world. The rise in mortality is attributed due to delayed or misdiagnosis of pathogens causing infections. Currently, microbiological, biochemical, and genetic analysis are deployed for identification of bacteria. However, these techniques are time-consuming, expensive, require large sample volumes as well as trained personnel. Therefore, the healthcare system demands for rapid, inexpensive, and accurate alternative strategies to improve response time and treatment regimens. Ratiometric fluorescent sensor array (3-hydroxyflavone derivatives) encapsulated in hyaluronic acid nanoparticles as a solution-based sensing tool were developed as an alternative to existing diagnostic tools. Despite the increase in sensitivity, this solution-based sensing tool showed similar shortcomings. To address this, our study focuses on paper microzone plates as a solid support for improved cost, sensitivity, identification and analysis time for detection of pathogenic bacteria.

Methods: Environment sensitive fluorescent sensor array (3-hydroxyflavone derivatives) were synthesized using modified Algar-Flynn-Oyamada reaction. Subsequently, paper microzone plates mimicking 96-well plates, were fabricated by photolithography using Grade 1 Whatman filter paper. Sensor dyes immobilized in five paper plates stored in the dark at room temperature were subjected to intra-plate and inter-plate stability and reproducibility of the dye’s signal. Fluorescence signal measured at λ_em = 485, 515, 525, 550, 575, 600 nm and λ_exc = 400 nm at various time points (t = 0, 1, 2, 16, 24 weeks) were subjected to intensometric and ratiometric analysis. Next, the bacteria-sensor dye interaction dynamics, interaction time, and number of bacteria were determined using whole cell or lysates of Gram-positive and Gram-negative bacteria. With the optimized parameters, the ability of the paper-based approach to identify 16 bacterial species and their Gram status was determined using pattern analysis and machine learning algorithms including neural networks and support vector machines. The classification accuracy was determined by using 75% of the dataset as ‘training dataset’ and remaining 25% as ‘testing dataset’. Further, the robustness of this tool was tested with biological samples including fecal culture from healthy and dysbiotic microbiota mice for diagnoses of inflammatory bowel diseases.

Results: 96-well paper plates were successfully fabricated by photolithography with hydrophilic wells surrounded by hydrophobic barriers (Fig. 1). This sensing tool required 150-fold less dye volume and sample amount. The intensometric and ratiometric analysis of the sensor dyes signals suggested that they were stable within the plates for up to 6 months post fabrication. Further, the signal intensity between five plates were consistent amongst all four dyes. Linear discriminant analysis of whole-cells or lysates showed that whole-cell based sensing showed greater discrimination between groups. Fluorescence microscopy comparing interaction of S. aurues and E. coli with sensor dyes, membrane stain – FM4-64, and nucleus stain – DAPI showed that the sensor dyes had increased interaction with the bacterial cell envelope than the cytoplasm. Further, the interaction between the bacterial samples and sensor dyes showed that the signal intensities and ratios of the multi-channel responses required about 40 minutes for the signal to stabilize and reached equilibrium after 60 minutes of incubation. Paper-based plates tested with various amounts of bacteria showed that about 1.6 × 10⁸ CFUs/mL was optimum for analyses. Samples with lower concentration of bacteria showed an intensity similar to that of the reference wells, possibly due to the lack of washing step. Next, increase in sample size to 16 species showed that approximately 12 species showed 94.5% and 95.9% accurate classification of bacterial species and Gram status, respectively using neural networks (Fig 2). The paper-based platform tested with fecal culture showed that it could differentiate between healthy and diseased mice using hierarchical clustering, an unsupervised machine learning algorithm (Fig 3).

Conclusion: Paper-based sensing approach using fluorescent sensor array provides a state-of-the-art versatile diagnostic tool for identification of bacteria. It is portable, rapid, inexpensive and does not require highly skilled personnel, thus translatable in low-resource settings. This approach will further be tested with fecal and blood cultures to differentiate between the two inflammatory bowel diseases – Crohn’s disease and ulcerative colitis. Thus, paper-based fluorescent sensor array integrated with machine learning opens avenues to translate into clinically relevant diagnostic tool for identification of infectious diseases.

References: 1. Svechkarev, D.; Sadykov, M. R.; Bayles, K. W.; Mohs, A. M., Ratiometric Fluorescent Sensor Array as a Versatile Tool for Bacterial Pathogen Identification and Analysis. ACS Sens 2018, 3 (3), 700-708.
2. Svechkarev, D.; Sadykov, M. R.; Houser, L. J.; Bayles, K. W.; Mohs, A. M., Fluorescent Sensor Arrays Can Predict and Quantify the Composition of Multicomponent Bacterial Samples. Front Chem 2020, 7, 916.
3. Klymchenko, A. S.; Pivovarenko, V. G.; Demchenko, A. P., Elimination of the Hydrogen Bonding Effect on the Solvatochromism of 3-Hydroxyflavones. J Phys Chem 2003, 107 (21), 4211-4216.
4. Carrilho, E.; Phillips, S. T.; Vella, S. J.; Martinez, A. W.; Whitesides, G. M., Paper microzone plates. Anal Chem 2009, 81 (15), 5990-8.
5. Laliwala, A.; Svechkarev, D.; Sadykov, M. R.; Endres, J.; Bayles, K. W.; Mohs, A. M., Simpler Procedure, and Improved Performance for Pathogenic Bacteria Analysis with a Paper-Based Ratiometric Fluorescent Sensor Array. Anal Chem 2022, 94 (5), 2615-2624.

Acknowledgements: 1) National Institutes of Health: R01 EB027662; 2) Bukey Memorial Funds


Fabrication of paper microzone plates with sensor array dyes (A) bright field. (B) Under UV radiation.


Machine learning algorithms for classification of bacterial species and Gram status


Dendrogram differentiation of healthy and diseased mice using hierarchical clustering analysis.