Keywords: Raman spectroscopy, deep learning, process analytical technology, microbial contamination, convolution neural network, biologics, CHO cells
Real-time release of pharmaceuticals (small molecules and biologics) requires the ability to use in-process data to evaluate and ensure the quality of the final product [1]. Within biologics, determining sterility and measuring microbial contamination is especially important [2]. Traditional United States Pharmacopeia microbial testing methods depend primarily on culturing of microorganisms to determine bioburden and sterility [1, 3]. Since culturing and culture-dependent methods are slow (1-21 days), they cannot be used for real-time release testing. Nucleic acid-based technologies (polymerase chain reaction, next generation sequencing) have reduced the time for analysis to the order of hours but they still require sample preparation and thus, remain invasive methods of detection. Spectroscopic methods, such as Raman spectroscopy, on the other hand are non-invasive, rapid (minutes), and versatile (can detect a variety of microorganisms).
Although incidents of microbial contamination are rare, they can be extremely costly. For example, bioreactors can be operated at scales of about 15,000 L scale with media costs of $65/L and thus, a single contamination could lead to a loss of around $975,000 [1]. Thus, detecting contamination in a timely manner and monitoring critical control points is essential for real-time release. Recently, a proof-of-concept rapid microbiological screening system was able to detect Escherichia coli spiked into Chinese Hamster Ovary (CHO) cell line culture within three hours by using filtration (to separate CHO cells), microfluidics (to generate nanoliter-sized droplets), and an indicator dye (to measure the doubling time of bacteria) [4]. Since the method requires filtration and growth of bacteria, it is still limited to at-line or off-line use.
Raman spectroscopy measures the inelastic scattering of light due to molecular vibrations. It is possible to distinguish phenotypes of microorganisms based on their molecular composition [5]. Since the differences in the Raman spectra of different microbes can be subtle, the use of deep learning algorithms is essential to separate signal from noise. A recent demonstration of this approach on human pathogens achieved an accuracy of about 82% for distinguishing isolates of microbes [5].
In the current work, we apply Raman spectroscopy and deep learning to pharmaceutical contaminants and demonstrate detection and discrimination of 12 different microorganisms (encompassing Gram-positive bacteria, Gram-negative bacteria, and fungi listed in Table 1). We have used a TeflonTM-coated polished stainless steel substrate (Figure 1) to obtain high signal-to-noise ratios. We also demonstrate discrimination of bacterial contamination in a mixture with CHO cells. We achieve accuracies in the range of 95-100% for determining microbial identity (Figure 2).
Neural network-based microbial contamination classification: We used the convolution neural network (CNN) as a deep learning strategy to classify the microbial contaminants (and CHO cells) relevant to the pharmaceutical industry. The CNN consists of multiple hidden convolutional layers. In each layer, a certain number of filters convolve over the input map and abstract it into the feature map, which is passed to the next layer. Each layer extracts a pattern (which is determined during the optimization process) in the input data and passes the resulting feature maps to the next layer to search for higher-level patterns. The final output is passed into a fully-connected layer that converts the extracted feature maps into the probability distribution over various classes [18]. In our study, the input layer is Raman spectroscopy data obtained from different samples, and the output is the probability distribution over the 16 classes of samples (12 microbes, 1 CHO cell, 3 mixtures of CHO cells and microbes to represent Gram-positive bacteria, Gram-negative bacteria, and fungi). To evaluate the multi-class classification model, we use a confusion matrix shown in Figure 2. In this matrix, the vertical axis denotes the actual classes of samples, and the horizontal axis represents the predicted classes. In this study, we classified the samples into 16 categories. Using the confusion matrix, we can evaluate the performance of the model on every single class and learn about the type of microbe where the model has the weakest capability in recognition. In our study, the model has the lowest accuracy forStaphylococcus epidermis that is misclassified asPropionibacterium acnes in 5% of the cases. On the other hand, the model has very high accuracy in detecting the difference between microbes and microbes mixed with CHO cells. According to the confusion matrix, the average accuracy of the model is 98.19 ± 0.55% (the standard deviation is calculated over the 5 splits of training and validation sets in the LOOCV approach).
Attention map for classification: To explain the internal functionalities of proposed CNN, we use the recently developed Grad-Cam++ method [17]. This method uses a linear combination of positive partial derivatives of class scores with respect to last convolutional layers features as weights to provide the attention map of particular class labels. The resulting attention map helps us understand the regions that are important for CNN to predict the class of input data. In this case, we can identify the range of wavenumbers in spectral data of species that are significant in categorizing them, as shown in Figure 3. According to the attention maps for various species, we notice that any patterns after the largest peak in spectral data (2850-3050 cm−1) do not have any significance for the model, and CNN focuses on a range of wavenumbers before the largest peak, which is around (400 - 3050 cm−1) in our study. The results imply that the wavenumbers in the aforementioned wavenumbers are important in identifying the species.
Important features of Raman spectra for distinguishing microbial contaminants: The Raman spectra were collected in the wide range of 100 - 6000 cm-1 to avoid missing any minute variations within the different microbes. We collected 10 technical replicates by measuring the same dried sample from different points on the substrate (with 200 scans per point) and three biological replicates by repeating the experiment on three different days for each species of interest. The average (bold lines) of 6000 spectra/sample class of all the microbes/cells are depicted in Figure 3 where shaded regions indicate standard deviations.
The Raman spectra of all the microbes and CHO cells have prominent peaks of nucleic acids (1575, 1481, 812, 783 cm−1), proteins (1002 cm−1), and lipids (1658, 1448 cm−1) [19, 20]. A strong Raman shift found in all the microbes/CHO cells is around 2850-3050 cm−1. This region is found to be a non-specific organic >CH2 and –CH3 stretching modes [21]. Though a subtle difference can be observed between the spectra visually, high-throughput analysis requires an automated tool for discrimination [22]. Thus, CNN helped to classify the microbes and the CHO cells and to highlight which parts of the spectra had the most impact on discrimination between classes.
Although Raman spectroscopy typically suffers from low signal-to-noise ratios, here, the use of a polished stainless-steel substrate (Figure 4) has enabled concentration of the bacteria and reduction of background noise. The same substrate has been used in the past for detecting proteins at levels as low as 1 fmol [7].
The use of Raman spectroscopy has the following four advantages over other rapid microbial testing methods in the pharmaceutical industry: i) it can distinguish between several different types of microbes (spanning over Gram-positive bacteria, Gram-negative bacteria, and fungi), ii) it can provide a signal even in the presence of CHO cells and thus, does not require physical separation or filtration of the cell types before detection, iii) when a small number of scans are used, it is non-destructive and thus, the samples could be used for culturing or sequencing if needed for tracing the contaminant, and iv) collecting spectra requires less than a minute and thus, the technique can be used at-line in the production plant.
The use of convolution neural network and attention mapping enables the following three advances: i) high-accuracy classification despite only subtle differences between different classes, ii) when a training set has been incorporated, classification is rapid (in seconds), and iii) highlighting which parts of the spectra are relevant to classification helps understand the reasoning behind the classification (instead of using a completely black box approach).
The key limitations of the current study are: i) we used a high concentration of cells (105 cells/mL) to show proof-of-concept, ii) we dried the cells down before detection, and iii) we fixed the cells using glutaraldehyde before detection (due to concerns of biosafety).
In future studies, we aim to improve the sensitivity of Raman spectroscopy by using microfluidics and acoustic concentration. We also aim to detect the cells directly in a liquid sample to simplify the process. Our work serves as stepping stones for developing sensors for PAT and enabling real-time release of biologics.