Improved Antibiotic Detection in Raw Milk Using Machine Learning Tools over the Absorption Spectra of a Problem-Specific Nanobiosensor

Abstract

In this article we present the development of a biosensor system that integrates nanotechnology, optomechanics and a spectral detection algorithm for sensitive quantification of antibiotic residues in raw milk of cow. Firstly, nanobiosensors were designed and synthesized by chemically bonding gold nanoparticles (AuNPs) with aptamer bioreceptors highly selective for four widely used antibiotics in the field of veterinary medicine, namely, Kanamycin, Ampicillin, Oxytetracycline and Sulfadimethoxine. When molecules of the antibiotics are present in the milk sample, the interaction with the aptamers induces random AuNP aggregation. This phenomenon modifies the initial absorption spectrum of the milk sample without antibiotics, producing spectral features that indicate both the presence of antibiotics and, to some extent, its concentration. Secondly, we designed and constructed an electro-opto-mechanic device that performs automatic high-resolution spectral data acquisition in a wavelength range of 400 to 800 nm. Thirdly, the acquired spectra were processed by a machine-learning algorithm that is embedded into the acquisition hardware to determine the presence and concentration ranges of the antibiotics. Our approach outperformed state-of-the-art standardized techniques (based on the 520/620 nm ratio) for antibiotic detection, both in speed and in sensitivity.

Keywords: absorption spectra; antibiotics; biosensors; gold nanoparticles; machine learning; nanotechnology; raw milk.

Figure 1

Aggregation of gold nanoparticles (AuNPs) and the corresponding absorption spectrum. (a) Microscopic images of non-aggregated AuNPs. (b) Microscopic images of aggregated AuNPs. These images where obtained using transmission electron microscopy for visual and morphological characterization (JEOL JEM 1200EX II Microscope, Electron Microscopy Laboratory, Universidad de Concepción). (c) Absorption spectrum of aggregated and non aggregated AuNPs. When the aggregation occurs, it produces a shift in the intensity peak of the absorption spectrum with the corresponding colorimetric change.

Figure 2

Development stages of the proposed device. (a) First laboratory prototype; (b) industrial design to address the drawbacks of the laboratory prototype; (c) microreactors and optical colimators; (d) prototype container and (e) key internal components: (I) an Ocean Optics STS VIS-NIR spectrometer, (II) P600 optical fibers that transmit from ultraviolet to the visible range spectrum, (III) an Ocean Optics HL-2000-HP-FHSA light source, (IV) 96-well plates for placing the samples to be measured and (V) a two-dimensional high-precision displacement unit to move the microplate and place the sample to be measured under the radiometer.

Figure 3

Block diagram of the optical data acquisition system. Firstly, a reference well is measured to estimate the aggregated spectrum of the container, the water and the dissolved NBS. The reference data are stored by the developed software. Secondly, each sample is measured in the same way and the NBS absorbance is obtained by subtracting the reference from the sample absorbance.

Figure 4

Adjustment of the measured absorption spectrum $r_j\left(\lambda \right)$ by means of Legendre functions of the first type. As the number of basis functions is increased, the better the fit of the approximation ${\widehat{r}}_j\left(\lambda \right)$. The approximation is shown with (a) ten basis functions and (b) twenty basis functions.

Figure 5

Root-mean-square error (RMSE) of the fit between the measured absorption spectrum $r_j\left(\lambda \right)$ and the approximation made with the Legendre functions ${\widehat{r}}_j\left(\lambda \right)$ for Kanamycin. We require $N=20$ basis functions to ensure an RMSE below 1%.

Figure 6

The spectrum measured from a sample, $r_j$, with unknown antibiotic presence and concentration level enters the algorithm. The support-vector machine (SVM) scheme determines whether an antibiotic is present or not. If the antibiotic is present in the sample, the tree is cascaded until one of the SVM is certain, meaning it has detected the concentration level.

Figure 7

Measured absorption spectra for the four antibiotics used in this work. The available repetitions for each antibiotic and each of the concentrations are shown. Although the curves show separability, it is not possible to estimate concentrations by visual inspection.