Molecular analyses
Results of molecular analyses
The sequences obtained are available in GenBank (Supplementary Information 1). Sequence alignments were 399 bp for COI and 587 bp for 28S including gaps.
Phylogenetic analysis
Our molecular analysis (Fig. 1) with both markers generated seven supported clusters, six of which were in agreement with the morphological determination (i.e. C. alazanicus, C. brunnicans, C. circumscriptus, C. furcillatus, C. nubeculous and C. pictipennis). However, one cluster (i.e. two species) corresponded to undistinguished C. clastrieri and C. festivipennis.
Block diagram of the study.
In addition, the COI mtDNA tree shows that C. furcillatus is the sister of the “C. clastrieri/festivipennis” clade. Indeed, C. pictipennis is the sister species of C. brunnicans while C. circumscriptus is positioned between the two clades.
Moreover, the 28S rDNA tree shows that C. pictipennis is the sister of the “C. clastrieri/festivipennis” clade. The other species are positioned in several places without a clade.
Intra- and inter-specific comparison
The COI Genbank sequences show little intraspecific divergence in both C. clastrieri (0.1 ± 0.1%) and C. festivipennis (1.2 ± 0.4%). The interspecific difference between C. clastrieri and in C. festivipennis is 0.7 ± 0.2%.
Small intraspecific divergences with COI sequences were observed in our sample: C. alazanicus (1.2 ± 0.4%), C. brunnicans (0.7 ± 0.2%), C. circumscriptus (2.2 ± 0.5%), C. clastrieri (0.3 ± 0.1%), C. festivipennis (0.4 ± 0.1%), C. furcillatus (1.5 ± 0.4%), C. nubeculosus (0.2 ± 0.1%) and C. pictipennis (1.1 ± 0.3%).
Finally, C. festivipennis and C. clastrieri—grouped in the same main clade—showed small interspecific distances (0.4 ± 0.2%); these were not identified as separate species based on DNA barcodes. We therefore decided to create a new group (C. clastrieri/festivipennis clade) based on interspecific distance. The overall mean genetic distance (K2P) computed for the different species of Culicoides was found to be 16.6 ± 1.4%. Interspecific K2P values for different (Table 1) species and taxa ranged from 27.3% (between C. furcillatus and C. nubeculosus; between C. circumscriptus-and C. furcillatus) to 17.2 ± 2.1% (between C. circumscriptus and the C. clastrieri/festivipennis clade) for our samples. For the COI Genbank sequences, we observed approximatively the same proportion and the same species (Table 1). We remarked very little interspecific divergence between our sample of the C. clastrieri/festivipennis clade and the C. clastrieri/festivipennis Genbank clade (0.6 ± 0.4%).
Analysis from 28S rDNA sequences did not show any intraspecific divergence whatever the taxa (0.000) with the exception of C. nubeculosus (0.1 ± 0.1%) and C. festivipennis/C.clastrieri (0.1 ± 0%). The overall mean genetic distance (K2P) computed for the different species of Culicoides was found to be 2.1 ± 0.03%. Interspecific K2P values for different species (Table 1) and taxa ranged from 1.2% (between C. circumscriptus and C. furcillatus; C. furcillatus and C. brunnicans, the main C. clastrieri/festivipennis clade and C. furcillatus) to 5.3 ± 0.9% (between C. circumscriptus and C. nubeculosus).
Morphometric and morphological analyses
In all, 148 specimens identified as C. alazanicus (n = 10), C. brunnicans (n = 27), C. circumscriptus (n = 27), C. clastrieri (n = 21), C. festivipennis (n = 20), C. furcillatus (n = 14), C. nubeculosus (n = 19) and C. pictipennis (n = 20) were analysed with 11 wing landmarks/specimens (Fig. 2).
Trees obtained from nucleotide analysis of: (a) COI mtDNA; (b) 28S rDNA (with MP method) sequences of C. alazanicus, C. brunnicans, C. circumscriptus C. clastrieri, C. festivipennis, C. furcillatus, C. nubeculosus and C. pictipennis and bootstrap values are shown in nodes (1000 replicates).
Principal component analyses
Principal component analysis (PCA) was used to observe possible grouping trends.
Firstly, we performed a first normed PCA using the “Wing landmarks” model. The first three axes accounted for 76%, 15% and 8% of the total variance, which suggests a weak structuration of the data. This was confirmed by a scatterplot of PCA axes 1 and 2 that was unable to separate the species (Fig. 3).
Principal component analysis (PCA): percentage of variance explained for each PCA dimension and results.
Secondly, we performed a first normed PCA on the “Wing morphological characters” model. The various specimens of each species are represented by a single point suggesting a close correlation of wing morphological characters. This model, without variance, is not validated and does not permit species separation.
We studied the “Full wing (landmarks and morphological, characters)” model through a normed PCA on raw data. C. clastrieri could be clearly separated from C. festivipennis. The first five axes accounted for 40%, 25%, 12%, 10% and 5% of the total variance. The scatterplot separated unambiguously and without overlap C. clastrieri-C. festivipennis on the one hand and the six species on the other hand (Fig. 3).
Finally, we performed a first normed PCA on the “Full model” (Morphological characters—wing, head, abdomen, legs—and wing landmarks). The first nine axes accounted for 26%, 23%, 22%, 10%, 8%., 4%, 3%, 2% and 1% of the total variance, which reveals good structuration of the data. This was confirmed by a scatterplot of PCA axes 1 and 2 that presents the same topology as the wing morphological model (Fig. 3).
This supports discrimination according to the species’ wing pattern. Similarly, and some body pattern characters could be used to identify Culicoides from the clastrieri/festivipennis clade better and quicker. With that objective in mind, we performed analyses on three datasets: (1) “Wing landmarks” (11 landmarks); (2) “Full wing” (38 items) and (3) the “Full model” that includes 71 items.
Discriminant analyses
PLS-DA and sPLS-DA models were used in order to discriminate the extremes (i.e. the most sensitive and most robust groups) using the three datasets (species, models and components) as described. The accuracy and the balanced error rate (BER) for the two models were compared and are summarised in Supplementary Information 2 and Fig. 4.
Balanced error rate (BER) choosing the number of dimensions. Performance and ncomp selection.
The tuning step of the number of components to select showed that 16 components were necessary to lower the BER (Fig. 4A,B) for the “Wing landmarks” data. The AUC values with 16 components are as follows: C. alazanicus (0.97, p < 0.001), C. brunnicans (0.98, p < 0.001), C. circumscriptus (1.00, p < 0.001), C. clastrieri (0.97, p < 0.001), C. festivipennis (0.89, p < 0.001), C. furcillatus (0.97, p < 0.001), C. nubeculosus (1.00, p < 0.001) and C. pictipennis (1.00, p < 0.001). After 16 components, the AUC values are approximately comparable (Fig. 4).
From the performance plot (Fig. 4), we observe that the overall error rate and the BER are similar for the “Full wing” model (Fig. 4C,D) and the full model (Fig. 4E,F), and decrease when components increase from one to eight. The error rates stabilise after nine components for PLS-DA and sPLS-DA models for “Full wing” (Fig. 4C,D). The AUC values with nine components (Supplementary Information 2) are as follows: C. alazanicus (1.00, p < 0.001), C. brunnicans (1.00, p < 0.001), C. circumscriptus (1.00, p < 0.001), C. clastrieri (1.00, p < 0.001), C. festivipennis (1.00, p < 0.001), C. furcillatus (1.00, p < 0.001), C. nubeculosus (1.00, p < 0.001) and C. pictipennis (1.00, p < 0.001).
In contrast, the error rates stabilise after eight components for PLS-DA and sPLS-DA for the full model (Fig. 4E,F). The AUC values with eight components (Supplementary Information 2) are as follows: C. alazanicus (1.00, p < 0.001), C. brunnicans (1.00, p < 0.001), C. circumscriptus (1.00, p < 0.001), C. clastrieri (1.00, p < 0.001), C. festivipennis (1.00, p < 0.001), C. furcillatus (1.00, p < 0.001), C. nubeculosus (1.00, p < 0.001) and C. pictipennis (1.00, p < 0.001).
A perfect result would be an AUC of 1.0 obtained using eight components with PLS-DA (Fig. 5) and seven with sPLS-DA using the “Full model” (Fig. 6). For the “Full wing” model, nine components with PLS-DA (Fig. 5) and seven with sPLS-DA (Fig. 6) are needed to obtain an AUC of 1.0.
The ROC curve of the “Full wing” and “Full model” obtained with partial least squares discriminant analysis (PLS-DA) according to the components. A perfect result would be an area under the curve (AUC) of 1.0.
The ROC curve of the “Full wing” and “Full model” obtained with sparse partial least squares discriminant analysis (sPLS-DA) according to the components. A perfect result would be an area under the curve (AUC) of 1.0.
The most discriminating characters for the “Full wing” and “Full model” with PLS-DA and sPLS-DA are summarised in Supplementary Information 3 and 4 respectively.
With the PLS-DA classifier and the “Full model”, we observe that a large number of characters are necessary to separate species. According to the components, 39 items were identified for C. clastrieri (× 11, × 2, × 1, WingM18, × 4, y2, AnM54, y11, Palp.M51, y9, × 9, × 2, Wing.M11, y2, y5, y11, y6, Ab.M30, Ab.M29, y1, y4, y5, Wing.M8, Ant.M54, y3, Palp.M50, Wing.M2, Ab.M28, × 1, y10, Ant.M52, Wing.M7, Wing.M8, , Wing.M13, × 7, Phr.M48, y8, × 11, Ab.M30) and eight for C. festivipennis (Wing.M12, Ant.M52, WingM12, WingM4, y11, Ant.M58, Ant.M54, WingM11). For example, for C. pictipennis, seven items (y2, × 5, y10, y3, × 8, Wing.M4, × 6) were observed. For the other species, see details in Supplementary Information 3. Fewer descriptors are needed for species discrimination with the sPLS-DA model than with the PLS-DA model. Fourteen descriptors were needed for C. clastrieri (y3, × 8, Wing.M2, y4, Wing .M9, y5, y1, y8, Palp.M51, × 4, × 11, Wing.M20, Ant.M56, Wing.M12) and one for C. festivipennis, (Wing.M12). Only two items are needed to identify C. pictipennis (Ant.M57, Wing.M25) and three for C. furcillatus (Ab.M33, Ant.M58, Ab.M29). For the other species, see details in Supplementary Information 4.
With the PLS-DA classifier and the “Full wing”, we observe that many items are necessary to discriminate species. According to the components, 31 items were identified for C. clastrieri (× 1, × 2, × 3, × 4, × 6, × 9, × 10, × 11, y1, y2, y3, y4, y5, y6, y7, y9, y11, WinM.2, WingM.3, WingM.6, WingM.7, Wing.M8, Wing.M9, Wing.M10, Wing.M11, Wing.M13, Wing.M14, WingM.15, WingM.16, Wing.M18, Wing.M19) and eight for for C. festivipennis (× 1, × 7, y10, Wing.M1, WingM.10, Wing.M11, Wing.M12, Wing.13). For the other species, see details in Supplemental Data S3. In contrast, with the sPLA-DA classifier, only a few descriptors are needed for species classification. Five descriptors were needed for C. clastrieri (y1, y2, y3, y4, y5) and one for C. festivipennis, (Wing.M12). For the other species, see details in Supplementary Information 4.
Source: Ecology - nature.com
