Identification of a clonal population of Aspergillus flavus by MALDI-TOF mass spectrometry using deep learning
In 2015-2016, the mycology laboratory of the Marseille University Hospital was commissioned to search for the presence of mold in several batches of surgical masks with a musty smell. These batches came from various suppliers.
Aspergillus flavus culture and identity
The masks were introduced into a mixing bag (Gosselin, France) and washed in 15 mL of Tween® 80 at 0.1% in saline solution for 10 min in a stomacher (Homogenius HG400, Mayo International, Italy). The liquid was then transferred to a 15 ml Falcon tube and centrifuged for 10 min at 3500 rpm. The supernatant was discarded and the pellet was resuspended in 500 µL wash solution. The next 500 µL were immediately cultured on a Sabouraud-Chloramphenicol Petri dish. Aspergillus flavus colonies were identified from 23 batches of masks using the recently released MSI-2 application (which allows online identification of MALDI-TOF fungal mass spectra)13 and subjected to microsatellite typing with 32 randomly selected isolates from the daily workflow of four French university hospitals (Hôpital de la Pitié Salpêtrière in Paris, CHU Bordeaux, CHU Toulouse and CHU Montpellier) to obtain a wide diversity of microsatellite profiles. The list and origin of the 55 isolates can be found in Supplementary Table 1.
All A. flavus isolates were typed using the protocol and microsatellites described by Hadrich et al. in 201014. This publication showed that a combination of five of the twelve microsatellites tested constituted the most parsimonious panel, achieving a Simpson’s diversity index (D) greater than 0.95. For our study, we used these five microsatellites: AFLA1, AFLA3, AFLA7, AFPM3 and AFPM7. However, the AFPM3 marker (a marker with a complex motif ((AT)6AAGGGCG(GA)) was dropped because no positive PCR was obtained for 21 of the 23 isolates of Aspergillus flavus recovered from the masks. Of the four remaining microsatellites, the AFLA3, AFLA7 and AFPM7 markers could be amplified for all samples and were used to construct an UPGMA tree including the 55 isolates to be tested. The UPGMA tree was constructed using the website http://genomes.urv.cat/UPGMA/ and considering our data as categorical values. For 10 isolates from the panel of patient isolates, the AFLA1 marker could not be read; therefore, we used the value of this AFLA1 marker only to confirm clonality and strengthen the clone definition of the isolates. Simpson’s diversity index using three (AFLA3–AFLA7–AFPM7) or four (AFLA3–AFLA7–AFPM7–AFLA1) markers used here is almost equal to 0.88 with our data.
MALDI-TOF MS Protein Extraction
Each isolate of A. flavus were first thawed as they had been stored in preservative liquid with ceramic canisters at -80°C for up to 3 years. One or two bids were used for culture on Sabouraud chloramphenicol gentamicin (SCG) agar and incubated for one week at 30°C to obtain sporulating isolates. After this incubation period, each isolate was subcultured onto a new SCG plate and incubated for three (“Day-3”) and five (“Day-5”) days at 30°C. This two-step subculture allowed similar growth for all isolates, as some of them took longer to revive. At both culture ages, protein extraction using consecutive incubations of formic acid and acetonitrile was performed as previously described.15.
Workflow for MS Acquisition
The 55 isolates were cultured twice, with a delay of one month, under the same culture conditions, on SCG culture media (Oxoid, Dardilly, France). The first batch of culture was used for the CNN training process, while the second batch of culture was used for testing only. The workflow is schematized in Fig. 1.
After determining the microsatellite profile of each of the 55 isolates and extracting the proteins from all the isolates, the number of clonal and non-clonal deposits was balanced for the training part of the study (Session S1). For this purpose, the protein extracts of the clones were deposited in two spots on the MALDI-TOF MS target, while the other extracts were deposited in only one spot. The 55 isolates were deposited on the same target. To multiply the number of spectra available for training, four targets were prepared in parallel, and the spectra were acquired again the following day, using a new calibration of the Microflex.
The second session of spectral acquisitions (S2) aimed to test the neural network developed from the training session (S1). Here, we deposited the protein extract on a single spot per isolate on the target, as it would be done under routine conditions. However, deposits were made on three targets which were evaluated on three different Microflex mass spectrometers (one at the mycology laboratory of the Saint-Antoine hospital in Paris (SAT-MS), one at the mycology laboratory (MYCO- MS: the same as for the learning phase) and one in the bacteriology laboratory of the Pitié Salpêtrière hospital (BACT-MS)). Our goal was to test the robustness of the results across devices.
Selection of spectra for training versus testing
To increase statistical reliability, we performed cross-validation of CNN with iterative experiments (n = 30) in which 44 isolates (80% of isolates) were randomly selected to train a neural network with the corresponding mass spectra obtained during the first culture session (S1). For each of the 30 iterations, the mass spectra corresponding to the remaining 11 isolates (20% of the isolates) were discarded to avoid sharing isolates in the training and test sections. Instead, for these 11 isolates, the spectra were selected from the second session (S2) of cultures obtained one month after the first culture session and were included in the test set. After each training/testing iteration, the training data was reset and a new random selection of 44 isolates for training using the S1 spectra and 11 isolates for testing using the S2 spectra was performed. The selection of isolates for training/testing was done using a stratified random sampling method. Each training set has approximately 980 spectra and each test set has 22 spectra.
For each experiment, sets of mass spectra used for training and for testing were obtained from cultures, and mass spectra acquisitions were made one month apart. The two different culture ages were grouped together for the training phase, while they were analyzed separately or grouped together to test the neural network. Additionally, while we used a single mass spectrometer to acquire the spectra used for training, we acquired the spectra of the isolates assigned to the tests on three devices. The aim was to test whether a neural network trained on one device maintained the ability to identify additional isolates cultured at a different time and with spectra acquired on another device.
Convolutional Neural Network (CNN): System Architecture and Training Process
The raw data extracted from the fid files of each of the spectra was treated in the same way before being included in the CNN. In total, approximately 22,000 intensity values (corresponding to 22,000 m/Z) composing a spectrum were subjected to a three-step preprocessing. First, the baseline was subtracted using the asymmetric least squares smoothing method16. Second, a Fourier transform was applied to smooth the spectra to avoid small intensity variations. Third, the peaks were selected by calculating the derivative of the intensity and then detecting the sign changes in the derivative of the spectra17. The processing of the spectra is schematized in FIG. 2.
A CNN is a special type of deep learning model for data processing that has a grid model and is designed to automatically and adaptively learn the spatial hierarchies of features from low to high level models. In our case, we work with a 1D CNN, because the spectra are represented in a one-dimensional way. The idea is to see if the CNNs manage to read the spectra as images and capture local patterns that are peaks of interest and thus manage to classify them. As a first intention to test a neural network on fungal mass spectra, we built a simple CNN using TensorFlow v2.0.0.
The CNN model contains 1 input layer, 1 convolutional layer, 1 max clustering layer, 1 flatten layer, 1 fully connected layer and 1 softmax18 layer to form the output prediction. The number of filters and the kernel size were set to 8 and 16, respectively, for the convolutional layer. The activation function of the convolutional layer and the fully connected layer were rectified linear units (ReLU). The softmax function was applied in the last layer to produce the prediction probability on the two output classes: Clones (CL) and Non-Clones (NC).
Categorical cross-entropy was selected as the loss function and Adam’s algorithm was selected as the optimizer. The learning rate was set to 0.001 and the number of epochs was set to 100. A schematic of this CNN can be found in Fig. 3.
For each spectrum submitted to the CNN, the result is expressed as a prediction (probability that the spectrum matches each of the output classes), and the probability that a spectrum matches a clone plus the probability that it matches a no -clone is always one. Accuracy (percentage of correct identifications), precision (proportion of true positive categorizations among positive categorizations = positive predictive value) and recall (ability to identify clones = sensitivity) for the 30 iterations are calculated and compiled to assess the CNN.