A flow chart of the conventional modern geometric morphometric analysis. The standard pipeline comprises several steps (left). First, a set of landmarks are chosen based on distinct anatomical features believed to be equivalent in some sense (e.g., biologically homologous) among all the specimens (papionins), and their cartesian coordinates are scanned and recorded. Next, the landmark data are superimposed using GPA through translation, rotation, and scaling. Next, after performing PCA, the scatterplots of the first two or three PCs are used to assess patterns of shape variation based on the distribution of the observations. The outcome of the pipeline (right) is depicted for the benchmark data. For brevity, only three taxa are shown on the top right.

PC scatterplots of the shape space of the lower molar of NR2 and other Hominins (left). The analysis combines the enamel–dentine junction (EDJ) and the cemento-enamel junction (CEJ). The Kernel Density Estimation (KDE) plots the distribution of the distances of each Homo species from NR fossils in the space of the corresponding PCs (right). The legends of the right sub-figures present the adjusted (Benjamini/Hochberg) p-values of the T-test for the means of the distances of each Hominin population and Neanderthals from NR fossils. Legend guide: African Early Homo sapiens (AEHs), African Middle Pleistocene Homo (MPH), Early Homo sapiens (EHs), European Middle Pleistocene Homo (EMPH), Homo erectus (He), Recent Homo sapiens (RHs). For consistency, the colours of the species are similar to those of Hershkovitz et al. (22) PC plots.

Comparing two hypotheses of papionin phylogenetic relationships. (a) Hypothesised phylogenetic tree of the extant Papionini from molecular (mtDNA and Y chromosome) data. (b) the most parsimonious tree, using Macaca rather than Pan as the outgroup. The figure was obtained from Gilbert et al. (105) (their Figure 2) under the exclusive PNAS License to Publish [Copyright (2007) National Academy of Sciences].

Plots depicting the papionin benchmark data. A) PC 1-2, B) PC 1-3, C) PC 2-3, D) Confusion matrix of a 2NN classifier with 5-fold CV, E) Dendrogram of Procrustes distances of the samples using agglomerative clustering, F) Decision boundary of the 2NN classifier after t-SNE embedding with 5-fold CV.

The accuracy and balanced accuracy of the eight classifiers.

The accuracy was evaluated in a five-fold stratified CV on the benchmark dataset and the six alteration cases. The numbers in parentheses are the balanced accuracies.

Plots depicting the papionin data after the removal of the Lophocebus albigena (green) taxon. A) PC 1-2, B) PC 1-3, C) PC 2-3, D) Confusion matrix of a 2NN classifier with 5-fold CV, E) Dendrogram of Procrustes distances of the samples using agglomerative clustering, F) Decision boundary of the 2NN classifier after t-SNE embedding with 5-fold CV. The insets in the subfigures are the corresponding benchmark PC scatterplots (Figure 4A-C).

Plots depicting the papionin data after removing the Macaca mulatta (orange) taxon. A) PC 1-2, B) PC 1-3, C) PC 2-3, D) Confusion matrix of a 2NN classifier with 5-fold CV, E) Dendrogram of Procrustes distances of the samples using agglomerative clustering, F) Decision boundary of the 2NN classifier after t-SNE embedding with 5-fold CV. The insets in the subfigures are the corresponding benchmark PC scatterplots (Figure 4A-C).

Plots depicting the papionin data after the first case of sample removal. 11 Lophocebus albigena (green) samples (index in benchmark dataset: 52, 56, 60, 63, 64, 66, 67, 73, 75, 76 and 77). A) PC 1-2, B) PC 1-3, C) PC 2-3, D) Confusion matrix of a 2NN classifier with 5-fold CV, E) Dendrogram of Procrustes distances of the samples using agglomerative clustering, F) Decision boundary with 5-fold CV. The insets in the subfigures are the corresponding benchmark PC scatterplots (Figure 4A-C). The red arrow in subfigures A, B, and C marks the Lophocebus albigena (pink) sample whose position in PC scatterplots is of interest.

Plots depicting the papionin data after the second case of sample removal. 19 samples from both Lophocebus albigena (green) (index in benchmark dataset: 52, 56, 60, 63, 64, 66, 67, 73, 75, 76 and 77) and Cercocebus torquatus (blue) (index in benchmark dataset: 80, 81, 82, 83, 84, 88, 91 and 94) A) PC 1-2, B) PC 1-3, C) PC 2-3, D) Confusion matrix of a 2NN classifier with 5-fold CV, E) Dendrogram of Procrustes distances of the samples using agglomerative clustering, F) Decision boundary with 5-fold CV. The insets in the subfigures are the corresponding benchmark PC scatterplots (Figure 4A-C). The red arrow in subfigures A, B, and C marks the Lophocebus albigena (pink) sample whose position in PC scatterplots is of interest.

The LOF outlier scores.

Cercocebus torquatus (blue) was treated as an outlier in three cases; A) each sample separately, B) 6 samples together, and C) the whole group together.

Plots depicting the papionin data after the first case of landmark removal. Landmarks 23, 24, 25, 26 and 27 were removed. Scatterplots of the papionin data after the first case of landmark removal. A) PC 1-2, B) PC 1-3, C) PC 2-3, D) Confusion matrix of a 2NN classifier with 5-fold CV, E) Dendrogram of Procrustes distances of the samples using agglomerative clustering, F) Decision boundary with 5-fold CV. The insets in the subfigures are the corresponding benchmark PC scatterplots (Figure 4A-C).

Plots depicting the papionin data after the second case of landmark removal. Landmarks 11, 12, 16, 17, 18, 30 (symmetrical pair of 12) and 31 (pair of 13) were removed. PC 1-2, B) PC 1-3, C) PC 2-3, D) Confusion matrix of a 2NN classifier with 5-fold CV, E) Dendrogram of Procrustes distances of the samples using agglomerative clustering, F) Decision boundary with 5-fold CV. The insets in the subfigures are the corresponding benchmark PC scatterplots (Figure 4A-C).

Plots depicting the papionin data in the first extreme case. A) PC 1-2, B) PC 1-3, C) PC 2-3, D) Confusion matrix of a 2NN classifier with 5-fold CV, E) Dendrogram of Procrustes distances of the samples using agglomerative clustering, F) Decision boundary with 5-fold CV. The insets in the subfigures are the corresponding benchmark PC scatterplots (Figure 4A-C). the corresponding benchmark PC scatterplots (Figure 4A-C).

Plots depicting the papionin data in the second extreme case. A) PC 1-2, B) PC 1- 3, C) PC 2-3, D) Confusion matrix of a 2NN classifier with 5-fold CV, E) Dendrogram of Procrustes distances of the samples using agglomerative clustering, F) Decision boundary with 5-fold CV. The insets in the subfigures are the corresponding benchmark PC scatterplots (Figure 4A-C).

Detecting outliers using local outlier score and KDE plots. Each Cercocebus torquatus (blue) sample is treated as an outlier (red arrow, A1-O1 t-SNE plots) and the KDE of the scores are plotted for each case (A2-O2). The radius of the circle around each sample is proportional to the LOF score. The red dot in the KDE plots shows the LOF score of the outlier in each case.

Detecting outliers using local outlier score and KDE plots. Six Cercocebus torquatus (blue) samples (indexes: 84,85,86,87,90 and 92) are treated as outliers (A1 t-SNE plots) and the KDE of the scores are plotted (A2). The radius of the circle around each sample is proportional to the LOF score. The red dots in the KDE plots show the LOF score of the outliers.

Detecting outliers using local outlier score and KDE plots. All Cercocebus torquatus (blue) samples are treated as outliers (A1 t-SNE plot) and the KDE of the scores are plotted (A2). The radius of the circle around each sample is proportional to the LOF score. The red dots in the KDE plots show the LOF score of the outliers.

The 31 landmarks of Table S1. The wireframes connect the landmarks.