Drug-resistant Staphylococcus Aureus Bacteria Detection with the Combination of Surface-enhanced Raman Spectroscopy and Deep Learning Techniques

Abstract


47
Antimicrobial resistance is a growing problem globally, and 700,000 people die because of 48 resistant infections annually. By 2050, it will threaten 10 million lives a year 1 . Inappropriate 49 prescribing increases unnecessary antibiotic consumption, which triggers antimicrobial 50 resistance with a short period 2 . Antibiotic resistance can be prevented by prescribing the proper 51 antibiotics and raising public awareness. As another solution, new antibiotics can be discovered 52 to compensate for antibiotic resistance. However, the number of discovered and approved 53 antibiotics has declined between 1980 and 2014 3 . Hence, rapid, and correct diagnosis of 54 bacterial infections is required to prescribe the right antibiotic, and this is so crucial to curb 55 antibiotic resistance. 56 Antimicrobial susceptibility test (AST), categorized as phenotypic and genotypic, is 57 utilized to determine bacteria's antibiotic resistance. Phenotypic AST is reliable; however, it 58 contains a time-consuming culturing step. On the other hand, genotypic AST provides fast 59 results since it eliminates the need for culturing. Although it is highly sensitive, the existence 60 of resistance genes does not mean expressed resistance. Further, genotypic AST requires trained 61 personnel with advance knowledge 4 . Therefore, alternative diagnostic tools are needed for fast 62 and reliable detection of antibiotic resistance. 63 Surface-enhanced Raman spectroscopy (SERS) is a promising biomedical diagnostic 64 tool and span broad applications in the biomedical field 5-9 . Within the last two decades, it has 65 been successfully applied to discriminate bacteria as well 10-12 . Therefore, the SERS technique 66 also has a significant potential to detect bacteria's antibiotic resistance 13,14 . Although SERS 67 provides unique molecular information, SERS spectra of antibiotic-resistant and susceptible 68 bacteria show subtle spectral differences. Therefore, the SERS technique requires advanced 69 data processing algorithms to capture these minor differences. A vast majority of publications 70 have reported that machine learning techniques can be employed to discriminate antibiotic-71 resistant and susceptible bacteria by using data obtained from SERS [15][16][17][18] . 72 There are three main steps, including preprocessing, feature extraction, and 73 classification, to determine bacteria from the SERS data by using machine learning techniques. 74 Therefore, obtaining a classification model is very tedious and time-consuming due to the rigid 75 interdependency of the steps. Although some traditional machine learning techniques give 76 reasonable accuracy results to detect the type of bacteria, they have several disadvantages, 77 including overfitting, underfitting, requiring many user-supplied parameters, needing advanced 78 nonlinear optimization techniques, etc. Fortunately, these challenges can be overcome using 79 deep learning models whose achievement originates from large data volumes and sophisticated 80 computational abilities. Deep learning models can learn significant raw data patterns without 81 using advanced preprocessing and feature extraction techniques 19  including penicillin, cephalosporin, and carbapenem 30 . Undoubtedly, rapid and accurate 108 detection of antibiotic resistance profiles of S. aureus bacteria will both reduce morbidity and 109 mortality and slow down the development of antibiotic resistance. 110 We hypothesized that the cell wall structure of MRSA and MSSA might show some 111 differences due to the resistance mechanism. SERS able to reflect these differences at the 112 collected spectra. The discrimination of subtle spectral differences originates from the cell wall       The laser spot size was calculated as 1.3 µm (1.22 x λ / NA). The spectra were collected with a 159 5 µm step size to prevent overlapping. Two datasets were acquired on different days and a total 160 of 1500-1550 spectra were collected from each isolate. Hence, the total dataset consisted of 161 33,975 spectra. 1200 lines/mm -1 grating was used providing a spectral range from 550 to 1700 This process is as follows: where ci ϵ R Mx1 is the code, f is the encoding function, bi1 ϵ R Mx1 is the bias vector, and Wi1 ϵ 178 R MxN is the weight matrix of the encoder. Encoder part of an autoencoder is trained using 179 unsupervised fashion to dig significant feature information.

180
The decoder part reconstructs the input vector as x̂. Thus, an autoencoder tries to 181 generate its input at the output layer by minimizing the error as much as possible between input 182 x and output x̂. Decoding of ci is expressed as follows: where g is the encoding function, bi2 ϵ R Nx1 is the bias vector, and Wi2 ϵ R NxM is the weight 185 matrix of the decoder.

186
The objective function minimizing the error between the input and output is expressed: Two regularization term are added to equation (3) as seen in equation (4). λ is a 189 regularization term and used to prevent overfitting. β is the weight of the sparsity penalty term   Table 1.

314
A classifier performance can be measured different metrics and accuracy is one of them.  Table 2. The confusion matrix of SAE-based DNN is demonstrated in Fig. 5b as MSSA as seen in Fig. 5b.

335
The above results show that SAE-based DNN has better classification performance than 336 traditional classifiers. However, these findings should be supported with statistical analysis.  Competing interests 519 The authors declare no competing interests.

520
Data availability 521 The datasets generated during and/or analyzed during the current study are available from the 522 corresponding author on reasonable request.

524
Supplementary information is available for this paper.      Figure 1 General work ow of deep learning-based spectral data analysis for the discrimination of antibioticresistant bacteria.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. supplementaryinformationa.docx