A New Strategy for Canine Visceral Leishmaniasis Diagnosis Based on Molecular Spectroscopy and Machine Learning


 Visceral leishmaniasis is a neglected disease caused by protozoan parasites of the genus Leishmania. The successful control of the disease depends on its accurate and early diagnosis, which is usually made by combining clinical symptoms with laboratory tests such as serological, parasitological, and molecular tests. However, early diagnosis based on serological tests may exhibit low accuracy due to lack of specificity caused by cross-reactivities with other pathogens, and sensitivity issues related, among other reasons, to disease stage, leading to misdiagnosis. In this work was investigated the use of mid-infrared spectroscopy and multivariate analysis to perform a fast, accurate and easy canine visceral leishmaniasis diagnosis. Canine blood sera of non-infected, Leishmania infantum, and Trypanosoma evansi infected groups were studied. The data demonstrate that principal component analysis with machine learning algorithms achieved an overall accuracy above 85% in the diagnosis.


Introduction
Leishmaniasis is a group of neglected infectious diseases, affecting over 100 countries worldwide (1), caused by a protozoan parasite from over 20 Leishmania species. Three leading clinical forms of the disease affect humans: visceral leishmaniasis (VL), cutaneous leishmaniasis, and mucocutaneous leishmaniasis. In Brazil, VL is caused by Leishmania infantum, and its fatal if untreated in over 95% of cases. The successful treatment depends on its accurate, and early diagnosis (2).
There is a high occurrence of canine visceral leishmaniasis (CVL) in Brazil, and dogs become the main reservoir of the protozoan. The human infection is mainly caused by the bites of infected female phlebotomine sand ies, responsible for parasite transmission from dogs to humans (3). In this scenario, the euthanasia of infected dogs has been adoped as one of the main strategies for disease control.
Therefore, more accurate VL testing in dogs became essential to prevent and control the disease and avoid the controversial culling of dogs with false-positive diagnosis delivered by the available diagnostic tests.
Tests recommended by the World Health Organization exhibit some drawbacks. Based on the correct identi cation of the parasite under a microscope, the parasitological test shows high accuracy. However, it depends on the capture of the parasite by aspiration of skin lesions, bone marrow, and lymph nodes, which is invasive and presents low diagnostic sensitivity (4). Enzyme-Linked Immune Sorbent Assay (ELISA), indirect uorescent antibody (IFA), and immunochromatographic assay are the most common serological tests. They show high sensitivity but low speci city, often exhibiting cross-reactivity with other pathogens, such as the Trypanosoma genus (5)(6)(7)(8). Molecular diagnosis by polymerase chain reaction (PCR), which replicates speci c DNA fragments, is highly accurate, but it is a time-consuming and expensive test, hindering its use on a large scale (4,9,10).
Most alternative strategies to leishmaniasis diagnosis have been based on the development of biosensors. Impedance-based biosensors depend on the immobilization of a DNA fragment on an electrode; among them, NiO nanostructured lm was used for L. donovani diagnosis (11), and Au nanoleaves were employed in a highly sensitive sensor for L. major (12). In turn, L. infantum diagnosis was carried out using either Au electrodes (13) or Au nanoparticles dispersed in a polyaniline matrix (14).
Gold nanoparticle-based colorimetric biosensors from kDNA were developed to diagnose and differentiate L. major, L. tropica, and L. infantum (15). Electrical capacitance measurements of interdigitated electrodes with immobilized proteoliposomes detected speci c anti-L. amazonensis antibodies and distinguished them from Trypanosoma cruzi (16). Biosensors can provide accurate results to different species of Leishmania. Nevertheless, the need for DNA manipulation, electrical probes, chemicals, and, in some cases, controlled storage conditions di cult in its implementation as standard and widely deployed diagnostic tests (17). New alternatives to overcome these limitations have emerged from multivariate analysis of molecular spectra of bio uids, skin tissue, and cell samples to disease diagnosis.
Fourier transform infrared (FTIR) and Raman spectroscopy have shown great ability to enable the detection of infections diseases. The have successfully discriminated Candida albicans isolates, allowing candidiasis diagnosis (18). Discirminated strains of Staphylococcus aureus from other pathogens (19), and also identi catied bacteria related to dairy cow diseases (Escherichia coli, Trueperella pyogenes and Streptococcus spp (20,21)) was possible by using FTIR. Serological screening of echinococcosis, a zoonotic parasitic disease, was carried out using Raman spectroscopy (22). Cerebrospinal uid investigated by micro-Raman spectroscopy allowed the meningitis diagnosis (23). Malaria diagnosis by FTIR spectroscopy was developed from infected blood spectra (24). Other virus related diseases could be diagnosticated from blood serum and spectroscopic analysis, such as dengue by Raman spectroscopy (25), and hepatitis B and C by FTIR (26). Papilloma virus diagnosis was performed from Raman spectra of cervical secretion (27), and FTIR spectra of cervical exfoliated cells (28).
Studies based on molecular spectroscopy diagnosis related to Trypanosomatidae family are limited to an investigation of L. chagasi (sin.=L. infantum), T. cruzi and T. rangeli strain discrimination (29), and Phlebotomine sand ies specimen's classi cation based on FTIR spectroscopy (30). This is the rst report about canine blood serum evaluated by FTIR spectroscopy and machine learning algorithms for the Leishmania infection diagnosis. Three groups of canine blood serum were investigated, non-infected, L. infantum, and T. evansi infected animals. Different spectral range, and group combination were evaluated to provide the higher overall accuracy to the correct diagnosis. These results show the ability of moelcular spectorscopy to provide a sensitive and speci c analysis that is rapid anc cost effective making it suitable for wide scale tranlsation to enable effective canine disease control.

Blood serum collection
The samples were obtained from the "Laboratório de Biologia Molecular, FAMEZ (Faculdade Veterinária e Zootecnia)" biobank from "Universidade Federal do Mato Grosso do Sul (UFMS)", Brazil. The samples were collected with approval of the UFMS Ethical Committee of Research in Animals (CEUA 765/2016).
All methods were carried out in accordance with relevant guidelines and regulations.
Canine blood serum samples were collected from 48 dogs, and distributed into three groups: (i) 20 animals naturally infected by L. infantum; (ii) 8 animals naturally infected by T. evansi; and (iii) 20 noninfected animals (control group). The L. infantum infected animals exhibited clinical symptoms, such as onychogryphosis, lymphadenopathy, and progressive weight loss, etc. ELISA and immunochromatographic assays were employed as validation tests. The control group showed no clinical symptoms and negative results in the validation tests. The T. evansi infected group was con rmed by PCR and DNA sequencing, as described elsewhere (31). Serological tests excluded Leishmania infection in this group.

Sample preparation
The canine blood serum samples were removed from storage at -20 o C, and allowed to thaw at room temperature (27 o C) for 15 min. After homogenization, three deposition steps were performed by depositing 30 μL of serum on silicon oxide (SiO 2 ) substrate, followed by drying at 40 o C for 30 min. Each sample was analysed in triplicate, and a total of 144 samples were measured (120 from the L. infantum and non-infected animals, and 24 from the T. evansi infected animals).

FTIR spectra acquisition and data analysis
The mid-infrared spectra were acquired in a Fourier transform infrared spectrophotometer (Perkin Elmer, Spectrum 100) with an attenuated total re ectance accessory. The spectra were collected from 4000 to 600 cm -1 with a resolution of 4 cm -1 and 5 scans. The FTIR spectrum of each sample was obtained by averaging the spectra acquired in three different spots, from the center to the substrate's border, to avoid any inhomogeneity in the sample composition due to the drying process (32).
The average FTIR spectra of canine blood sera were investigated using MATLAB R2015b software after standard normal variate (SNV) pretreatment. Initially, the data were dimensionally reduced by principal component analysis (PCA) to evaluate the potential separation of the samples. The PCA analysis reduces the original amount of data into a new set of uncorrelated variables (principal components, PCs) by a linear orthogonal transformation of this data, revealing the data variance. The component scores (PCs plot) reveal the data separation enabling a predictive model for sample classi cation. Additionally, it is possible to identify the main spectral data variance by analyzing the weight associated with the linear transformation used to obtain the component score (loadings) (33).
Machine learning (ML) algorithms based on k-nearest neighbors (KNN), support vector machine (SVM), and discriminant analysis (DA) methods were performed by using the PCA data with the aid of MATLAB R2018b software. The number of PCs used in ML training was optimized to avoid over tting. Brie y, the DA algorithm uses the training data set to determine the contour between the classes distributions, classifying the test sample according to this criterion. The SVM organizes the data in a spatial distribution and groups it in categories through the optimization of a hyperplane between classes. SVM did not not assume a normal distribution, differing from DA in this point. The hyperplane can be linear or non linear, being adjusted to reach the best performance. The KNN is also based on the spatial distribution, the classi cation is based on the label of the k nearest neighbors. The number of closest neighbors for classi cation can be changed: k = 1 (Fine KNN) and k = 10 (Medium KNN). In this case, Euclidean distance is considered. In order to weight the distance between samples, the weigthed KNN classi er can be applied usind some speci c function (34).
These methods were compared, and the accuracy results on the validation set were summarized accordingly. The validation process employed was the leave one out cross-validation (LOO-CV). In this case, a sample was withdrawn from the training set and used to test the predictive modeling power. The process is repeated, alternating the withdrawn sample to test all sample sets. The sample classi cation was performed adopting the proper ML algorithms from the validation process so that each sample was tested and assigned to its corresponding group. The samples were correctly classi ed when the predicted class, from ML classi cation, corresponded to the true class identi ed from the standard method.

Results And Discussion
The average FTIR spectrum for non-infected, L. infantum, and T. evansi infected canine blood serum are shown in Fig. 1, exhibiting remarkable similarity for all investigated groups. The vibrational bands can be assigned to proteins, carbohydrates, lipids, and fatty acid molecules. The vibrational assignments and molecular groups in the canine blood serum can be are exhibited in Table 1.
The broad band centered around 3276 cm − 1 is assigned to the overlap of molecular vibrational modes of protein and water molecules. The variation of water content in biological samples is common, and signi cative variation between the samples can contribute to a wrong discrimination between the groups. The drying step, and the same sample preparation conditions for all saples are important to not cause interefence in our data discrimination due to water content.
The absorption in the 1700 − 1450 cm − 1 spectral range is attributed to proteins molecules ( Table 1). The main antibodies produced in response to Leishmania antigens belong to this group of molecules.
Although, very similar FTIR spectra were obtained, including for the non-infected group, probably due to the overlap, for example, of amide I and II absorption bands in proteins. This spectral range is very infomration rich and useful for sample discrimination as previously shown by other researchers (26, 35,   36). In addition, other molecules that are assyaed by FTIR (e.g. carbohydrates, lipids, and fatty acids) can be equally important for sample classi cation; the immune response of the organism during infection may impact these molecules, and consequently, cause subtle spectral alterations. Table 1 Vibrational assignments and related molecular groups for canine blood serum FTIR spectra (37)(38)(39)(40)(41)(42)(43)(44)(45)(46) Figure 2 shows the score plot and loadings for the 1800 − 800 cm − 1 range, which excludes the O-H/N-H overlapped bands, and lipid/amide II contributions for sample classi cation. Data selection is a rational alternative to improve the clustering in PCA analysis (30), when highly correlated data are used. Detailed results related to other spectral ranges can be found in the Supplementary Material (Figures S1-S4).
The score plot exhibited in Fig. 2(a), shows a broad data distribution with no cluster separation for noninfected, L. infantum, and T. evansi infected canine blood serum. The loadings, Fig. 2(b), indicate that the rst three PCs are responsible for over 91% of the data variance, mainly in the spectral range assigned to proteins around 1600 cm − 1 , including vibrational modes of immunoglobulins. However, the group separation was not effective by PCA because the infection did not induce measurable spectral differences between the groups with basis on the observed vibrational bands .
The PCA analysis was applied to 1x1 group classi cation to nd a better strategy to sample identi cation. Figure 2(c) shows the score plot for non-infected versus L. infantum infected canine blood serum with no apparent clustering. The three rst PCs represent 92.34% of data variation. The loading data exhibited in Fig. 2(d) is similar to the one shown in Fig. 2(b), with prominent peaks around 1600 cm − 1 . Analogous behavior was observed for the score plot and loadings for non-infected versus T. evansi (Figs. 2(e) and 2(f)) and L. infantum versus T. evansi infected canine blood (Figs. 2(g) and 2(h)). Therefore, 1x1 group classi cation did not enable cluster formation and separation, probably due to the high data correlation.
Machine learning algorithms that rely on the PCA data were then employed to maximize sample classi cation accuracy, testing different spectral ranges and group comparisons. The results are described in the Supplementary Material ( Figures S5 and S6). infantum infected group, with overall accuracy of 85% by using the rst 4 PCs, responsible for 95.67% of data variance, in the quadractic SVM. However, the best accuracy achieved for non-infected versus L.
infantum group classi cation, was 87.5% in the 1700 until 1450 cm − 1 range, by using the rst 10 PCs with Quadratic Discriminant Analysis ( Figure S6).
In turn, the classi cation of L. infantum and T. evansi infected samples yielded an overall accuracy of 100% using either linear or cubic SVM with 10 PCs (Fig. 3). Identical accuracy was achieved in the classi cation of non-infected and T. evansi infected sera by applying linear SVM with 6 PCs. The weighted KNM algorithm successfully classi ed only the non-infected and T. evansi infected samples, while linear SVM classi ed both groups. Cubic and quadratic SVM with 10 PCs also provided a 100% overall accuracy to classify L. infantum and T. evansi infected sera. infantum (85%), and 16 non-infected samples (80%) (Fig. 4(a)). Applying quadratic DA with 10 PCs for the data set from 1700 to 1450 cm − 1 , the overall accuracy classi cation of 87.42% was determined for the non-infected and L. infantum classi cation. The number of samples correctly classi ed was the same for L. infantum, and only 2 samples (10%) were incorrectly classi ed in the non-infected group (Fig. 4(b)). The 100% overall accuracy was reached in the classi cation of both L. infantum and T. evansi infected sera employing linear or cubic SVM with 10 PCs (Fig. 4(c)), and non-infected and T. evansi utilizing linear SVM or weighted KNN with 6 PCs in the 1800 − 800 cm − 1 range (Fig. 4(d)).
ML algorithms allied to PCA were able to distinguish T. evansi infection successfully in the group comparisons and provide good classi cation accuracy of L. infantum infection when compared with standard methods (4,9,11). Additionally, methods based on these strategies could present advantages such as easy sample manipulation, low cost, and fast diagnosis. There is no need for sample preprocessing and any input, and the collection and analysis of the spectra are performed in a few minutes. For comparison, serological tests currently available for CVL diagnosis require prior processing of biological samples and inputs such as antibodies, enzymes, and buffers. ELISA or IFA tests take around 60 to 180 min to deliver the diagnosis.
The false-positive classi cation for non-infected versus L. infantum versus T. evansi infected samples ( Fig. 4(a)) and non-infected versus L. infantum infected samples ( Fig. 4(b)) is not a signi cant issue since the present proposal may be used as a trial test. Positive tested animals would be tested again by high accurate methods before the specialist takes any decision. On the other hand, the false-negative classi cation was kept constant for three or two group tests, requiring, for example, new methods of sample preparation or measurement to improve the overall accuracy.
Additionally, plasma protein kinetics is used as a biomarker for diagnosis and clinical follow-up of many infectious diseases (47, 48). Acute-phase proteins (APPs) such as haptoglobin (Hp), serum amyloid A (SAA), and C-reactive protein (CRP) have been investigated in dogs infected by L. infantum (49). Compared with the increase in speci c antibodies (IgG), plasma proteins may change earlier due to infections, thus favoring an earlier disease diagnosis (49). In the present study, of the 20 animals in the control group (not infected), 4 (in the LI x N x TE comparison) and 2 (LI x N) were classi ed as LI (infected by L. infantum), which could be possible cases of recent infection, with an increase in APPs, but with no detectable IgG production. Thus, the animals could be infected but without clinical symptoms and with negative results in the serological tests. It is important to highlight that the animals used in this study did not undergo any selection, providing samples with high heterogeneity from animals with different sexes, ages, stages of infection, nutritional condition, and so on. Thus, the approach reported in this work presents a remarkable ability to classify animals infected by either L. infantum or T. evansi, taking into account the high inhomogeneity of the data closely related to the clinical reality.
Conclusions FTIR measurements and machine learning algorithms together with PCA provided an alternative for canine visceral leishmaniasis diagnosis, discriminating non-infected and L. infantum, T. evansi infected animals with about 85% of classi cation accuracy. L. infantum from T. evansi infected animals with 100% accuracy. The high inhomogeneity of the samples that constituted each group proved that this approach is a robust test for canine visceral leishmaniasis and trypanosomiasis diagnosis.  Figure 1 Average FTIR spectra for non-infected (black line), L. infantum (red line), and T. evansi (blue line) infected canine blood serum. The main vibrational assignments are indicated, below each spectral band, and their related molecular groups (proteins, carbohydrates, lipids, and fatty acids) are identi ed by different colors infected, canine blood serum. (c, d) non-infected, and L. infantum infected, canine blood serum. (e, f) L. infantum, and T. evansi infected, canine blood serum. (g, h) non-infected, and T. evansi infected, canine blood serum.

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