This was a retrospective longitudinal study of 76 clinical isolates of S. pneumoniae from pediatric patients diagnosed with IPD admitted between 2016 and 2023 to Hospital Pequeno Príncipe, a 372-bed academic pediatric hospital in Curitiba, Paraná, southern Brazil. All isolates were stored in sheep blood with 20% glycerol at −80°C (Figure 1A).

Bacterial colonies were cultivated on 5% sheep blood Columbia agar base (CBAB 5%) (NEWPROV, Pinhais - PR, Brazil) at 37°C in an atmosphere of 5% CO₂ for 24 h. Identification was carried out using matrix-assisted laser desorption ionization time-of-flight mass spectrometry (MALDI-TOF MS) (Bruker Daltonics GmbH, Bremen, Germany) on a Microflex LT Bruker MALDI-TOF MS with flexControl v3.4 software (Bruker Daltonics GmbH, Bremen, Germany) (Figure 1B). Confirmation was achieved by standard methods (WHO, 2011) species were identified by the susceptibility on the optochin test and on sodium deoxycholate (bile salt) solution.

Serotyping was carried out at the Instituto Adolfo Lutz (IAL) of the National Reference Laboratory for Bacterial Meningitis and Invasive Pneumococcal Infections of the Brazilian Ministry of Health. S. pneumoniae was serotyped using pneumotest-latex agglutination and the Quellung reaction using antisera, both from Staten’s Serum Institute (Copenhagen, Denmark) and according to the manufacturer’s instructions. Non-typeable (NT) isolates identified by Quellung were also verified by the deduction of serotypes/serogroups using sequential multiplex PCR for 41 serospecificities, following the gene targets and protocols described by the Centers for Disease Control and Prevention, Atlanta, GA, United States (Carvalho et al., 2010).

Polysaccharide-based phenotypic typing was conducted using the IR Biotyper^® system (Bruker Daltonics, Germany), located at Laboratório Central do Estado do Paraná (LACEN/PR). The selected bacterial isolates were first cultured on CBAB 5% (NEWPROV, Pinhais - PR, Brazil) at 37°C in 5% CO₂ for 24 h. The obtained bacterial colonies were subsequently cultured on MHF agar (Mueller Hinton agar +5% horse blood +20 mg/L β-NAD, NEWPROV, Pinhais - PR, Brazil) and incubated at 37°C in 5% CO₂. Suspensions of the bacterial isolates were prepared in IR Biotyper^® suspension vials containing metal beads (IR Biotyper^® kit, Bruker Daltonics GmbH, Bremen, Germany) according to the manufacturer’s instructions. Portions of the culture colonies were removed using a 1 μL disposable inoculation loop, and a homogeneous suspension was prepared in 50 μL of 70% ethanol. After homogenization, 50 μL of deionized water was added. The samples were then agitated at 1,350 rpm for 5 min (Thermomixer, Eppendorf, Hamburg, Germany). Subsequently, 15 μL of each sample were dispensed in quintuplicate onto a reusable 96-well silicon IR Biotyper^® plate (Bruker Daltonics GmbH, Bremen, Germany), with each spot serving as a technical replicate. For database generation, each isolate was sub-cultivated for 2 or 3 days (biological replicates) to accommodate technical variability. Additionally, 10 μL of the two standards, IRTS1 and IRTS2 (IR Biotyper^® kit, Bruker Daltonics GmbH, Bremen, Germany), were included in duplicate on each plate per run (Bruker, 2021; Figure 1C). One spot was left empty to measure the background after each spot. Spectra of poor quality, such as those with absorbance values outside the 0.4–2.0 window, were excluded from analysis, aiming to include at least 12–15 spectra of each isolate.

Spectra were acquired (transmission mode between wave numbers 4,000–500 cm⁻¹) and processed by OPUS software V.8.2.28 (Bruker Optics, Germany) on an IR Biotyper^® with the corresponding IR Biotyper^® software V4.0 (Bruker Daltonics) for data analysis. Spectral splicing was performed by default settings to the polysaccharide region of 1,300–800 cm⁻¹. For spectral distance visualization, dendrograms were created using the Euclidean distance as the exploration method with the UPGMA linkage type (average). Scatter plots in both 2D and 3D were visualized using dimension reduction techniques, namely Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA). To correct for technical variance, LDA was employed as a preprocessing step, capturing 95% of the variance with a maximum of 30 principal components, with isolate ID serving as the target group for the visualization of the 2D and 3D scatter plots. The isolates were labeled according to the corresponding Quellung reaction outputs to compare the results of the two techniques.

Among the 76 S. pneumoniae isolates, 18 distinct serotypes were identified using Quellung serology (3, 6C, 7F, 7C, 9 N, 10A, 16F, 14, 15A, 15B, 15C, 18C, 19A, 22F, 23A, 23B, 33F, and 35B). These same isolates were tested on the IR Biotyper^® equipment to create a database for developing a classifier capable of distinguishing the different serotypes. All the spectra generated and stored in the newly created database were incorporated into the training data set. Additionally, two other machine-learning algorithms, a support vector machine with a radial basis function (RBF) kernel and another with a linear kernel, were evaluated. However, these algorithms failed to achieve a high classification accuracy (>95%) during development and were subsequently discontinued. Subsequently, the IR Biotyper^® software’s machine learning algorithms were trained to process the spectra and build the classifier using an ANN, repeating this process over multiple cycles (300 cycles), and producing a confusion matrix to evaluate the accuracy of the resulting classification model. It was determined that it was not feasible to differentiate between serotypes 15ABC, 10A/16F, and 7CF; these were amalgamated into a single group for future classification models. Ultimately, the classifiers’ accuracy was tested through an internal validation trial utilizing all the generated spectra, which were split into two sets: a training dataset and a testing/validation dataset containing 384 and 459 spectra, respectively. This process also produced a confusion matrix as an output file, which depicted the accuracy of the predicted outcomes in three colors (green, red, and yellow), confirming the reliability of the serotypes classified with our developed classifier.

Of the 76 isolates, 48 (63.2%) were isolated from blood, 15 (19.7%) from pleural fluid, 11 (14.5%) from cerebrospinal fluid, one (1.3%) from ascitic fluid and one (1.3%) from thoracentesis fluid samples. Eighteen distinct serotypes were identified, of which serotypes 19A, 6C, and 3 were the most prevalent (Figure 2). The three most prevalent serotypes [19A (gray), 6C (light pink) and 3 (green)] show the largest “clouds” in the 3D plot (Figure 2). In this plot, the first, second and third principal components are displayed, which already show some distinct clusters. For the further analysis, all 30 principal components were explored to see distinct clusters and therefore the possibility of distinguishing/differentiating all the serotypes measured. Supplementary Figure S1 shows a deviation plot of the spectra generated by the readings of the three prevalent serotypes in our study, displaying their spectral differences. The IR Biotyper^® spectra for serotypes 19A, 6C, and 3 were analyzed using a deviation plot, which presents the median spectra of each serotype group along with their respective standard deviations displayed as a shade. In this deviation plot, the spectral differences between serotypes 19A, 6C, and 3 are shown, indicating that these can be differentiated using FT-IR. With the use of machine learning algorithm, which is incorporated in the IR Biotyper^® software, a classification model could be made. The purpose of a classifier is to evaluate spectra without the need of manual exploratory data analysis. This means that when a classification model is applied in the software (automatically triggered for pneumococci), it immediately predicts the outcome of any unknown measured spectrum such as the serotype. The created pneumococci serotype classifier can then be used in the future for automated serotype classification, during measurement of pneumococci isolates. Different serotypes can be observed in the dendrogram (Supplementary Figure S2), where the serotypes are represented in different colors on the right vertical axis. The cutoff value (8.216) of the dendrogram was an automatically calculated cutoff value (based on Simpson’s diversity index × mean coherence), which was selected to visualize the different serogroups and observe the clustering of the samples individually. The program has a limitation on the number of spectra (up to 500 spectra) to generate the dendrogram. To generate Supplementary Figure S2, the “average” function was used, which means that the spectra (both technical and biological representatives) are presented as “average spectra.” In this model, serotypes 9 N, 10A, 16F, 7C, and 7F are not distinguishable in the dendrogram, but this case does not create a problem in our classification model. For a dataset consisting of a total of 843 spectra obtained in the study, the use of a dendrogram to visualize clustering may not be recommended due to the complexity of visualizing hierarchical relationships. Dendrograms are effective for smaller datasets. An alternative to overcome this limitation is to use dimension reduction techniques, such as LDA. By applying LDA, the informative features of the spectra are retained, simplifying the data and revealing patterns that would otherwise cannot be seen in the dendrogram.

From the spectra obtained from 76 isolates representing 14 distinct serotypes (Table 1), a classifier was developed using all available machine learning algorithms in the IR Biotyper^® software. Among these algorithms, the Artificial Neural Network (ANN) machine learning demonstrated the highest possible accuracy. Upon creating the classifier, the confusion matrix immediately displayed the model’s classification accuracy. However, with the initial classifier differentiating all 14 serotypes, the accuracy threshold of >95% was not achieved due to misclassifications between serotypes 10A, 16F, 7C, 7F, 15A, 15B, and 15C, which was also visually evident in the 3D scatter plots from the LDA. For example, the confusion matrix showed that serotype 10A (actual class, vertical axis) was predicted as serotype 16F (predicted class, horizontal axis), indicating that these two serotypes could not be distinguished within this dataset. As a result, serotypes 10A and 16F were merged into a single label, 10A/16F, in the classifier. Similarly, serotypes 7C and 7F, as well as 15B and 15C, were combined. As a result, to create our final classifier, from the 76 isolates tested, 843 spectra were generated, classifying 11 distinct serotypes and three serogroups (Table 1), achieving an accuracy of 100% after 300 training cycles of the ANN (Figure 3A). The classifier was then ready for use, but before implementation, an internal validation was conducted to assess its robustness. Consequently, the entire dataset was partitioned into two groups: Group 1, the training set with “reduced” data, encompassed 384 spectra (Figure 3B); Group 2, used to test and validate the classifier, included 459 spectra (Figure 3C). Isolates in the training dataset displaying variance in their respective serotypes were selected. Following the 300-cycle training of the neural network, model using these two groups, an accuracy of 98% was achieved (Figure 4).