American Ginseng (PQ) adulterated with varying levels of Asian Ginseng (Panax ginseng, or PG) was used to illustrate an application of the POI (Harnly et al., 2013). For the study, 44 PQ samples were obtained from the Wisconsin Ginseng Board (harvested over 3 years from 20 different farms in Wisconsin) and 8 PG samples grown in China and acquired from American Herbal Pharmacopoeia and commercial sources. Samples were analyzed by flow injection mass spectrometry (FIMS) which yielded spectra with more than 1,000 ions.

The SSTM and SITM (Table 1) for the POI analysis were chosen as 98% and 90% PQ, respectively. Spectra for the test materials were synthesized mathematically using appropriate ratios of the 100% PQ and 100% PG spectra. In all, 344 (43 PQ × 8 PG) spectra were generated for each test material. The unsupervised PCA score plot (not shown) showed visual separation but the computed confidence limits did not allow detailed statistical analysis. Ironically, one-class modeling was used to provide detailed statistical analysis of the data although it was not the focus of the paper. The Q statistic served as the combined metric (Figure 4).

Figure 4A shows the Q statistic plot for 100% PQ, 100% PG, and 98% PQ with confidence limits based on one-class modeling for both 100% PQ and 98% PQ. The one-class model for 100% PQ has a sensitivity of 95% (42/44) and a specificity of 100%. The one-class model for 98% PQ, the SSTM, has a sensitivity of 98% (342/344) and a specificity of 100%. Figure 4B shows a similar plot with 90% PQ, the SITM. Based on the one-class model for 98% PQ, the 90% PG has a specificity of 99% (341/344).

A follow up study verified the previous numerical dilution model with physical dilution of 100% PQ with 100% PG (Harnly et al., 2013). The more labor intensive nature of the physical dilutions restricted the number of samples. Five samples of 100% PQ (arbitrarily selected from the samples in the previous study) were diluted with two samples of PG at ratios of 95:5, 9:1, 8:2, 6:4, 4:6, and 2:8. Figure 5A shows that unsupervised PCA provided a near linear progression from 100% PQ to 100% PG on the X-axis indicating that PQ concentration was the primary source of variance. Figure 5B shows the influence plot (Q statistic versus Hotelling T²) based on a one-class model of 100% PQ. While the Hotelling T2 residuals failed to discriminate between the different PQ purities, the Q statistic provided excellent separation of the five PQ concentrations. A plot of the square root of the Q statistic showed a linear relationship with the PQ concentration (plot not shown).

An additional follow up study examined the ability of FIMS to discriminate between PQ:PG ratios of 99:1, 98:2, 97:3, and 96:4 (unpublished data) to determine the level of adulteration of PQ that could be detected. Figure 6 shows the influence plots (Q statistic versus Hotelling T²) for the 4 purity levels of PQ. Once again, Q statistic was more useful than Hotelling T² for discriminating between populations. The sensitivity for 100% PQ was 95% and the specificity was 57%, 90%, 99%, and 100% for PG contamination of 1%, 2%, 3%, and 4%, respectively.

This study demonstrates the utility of one-class modeling for detecting adulteration. The Q values for physical and mathematical dilution showed excellent agreement. The digital model, based on the pure spectra (100%) of the authentic material and the adulterant, provides a much simpler means of evaluating the ability of the method to detect adulteration.

The POI studies used the term SIMCA (soft independent modeling of class analogy) instead of the one-class modeling. SIMCA, as explained earlier, consists of a series of one-class models. In all the studies reported in this paper, only a single class was used for modeling, a single PCA model. For the ginseng studies and all those that follow, the 95% confidence limit was chosen as the statistical limit. Other statistical limits can be selected by researchers. The Q and Hotelling T² statistics were based on only one principal component, thus mitigating the possibility of over fitting. In addition, pre-processing for all the PCA calculations consisted of unit vector normalization of the samples (i.e., the sum of the squares of the sample intensities are set to 1.0) and autoscaling (normalization of each variable by its standard deviation) with mean centering of each variable.

One-class modeling was used to distinguish A. racemosa L. (Ranunculaceae) from other Actaea species and commercially available roots and supplements (Harnly et al., 2016). FIMS and proton nuclear magnetic resonance spectrometry (NMR), two metabolic fingerprinting methods, and DNA sequencing were used to identify and authenticate the A. racemosa species. For this study, authentic A. racemosa botanical reference materials were acquired from four sources: American Herbal Pharmacopoeia (AHP), Strategic Sourcing (SS), North Carolina Arboretum (NCA), and the National Institutes of Standards and Technology (NIST). The NCA samples were triplicate samples collected from 22 sites across the eastern US.

Initial analyses of the Actaea species furnished by AHP using FIMS and NMR gave similar result as shown in Figure 7. PCA score plots (Figures 7A, B) for both methods showed a pattern differentiating A. racemosa from the other Actaea species. This differentiation was statistically verified by one-class modeling based on A. racemosa (Figures 7C, D). DNA sequencing using two independent gene regions (ITS and psbA-trnH) confirmed the metabolic fingerprinting results. Although not the point of this review, DNA sequencing provided identification for root materials but was inconsistent for the supplements. It was assumed that processing of the supplements destroyed the DNA in some cases.

The simplicity of categorization in Figure 7 was complicated when A. racemosa reference samples from the four sources were compared. A PCA score plot (Figure 8A) and a influence plot (combined one-class model plots for AHP and SS reference materials) (Figure 8B) demonstrated that the reference materials were statistically different. The plots show three separate groups with the NIST samples in the same quadrant as those from SS. Sensitivity for the AHP and SS samples was 94% and 91%, respectively, and 91% for the un-modeled NCA samples in the upper right quadrant, exceeding the 95% confidence limits for AHP and SS, i.e., outliers compared to the other reference materials. Interestingly, the NCA samples collected from 22 sites across the US showed a similar lack of uniformity. These data suggested that sample handling and storage, local environmental conditions, and/or genetic drift influenced the plant metabolic profiles. It was also suggested that endophytic bacteria might have a significant influence on the metabolic profile.

A follow-up study took a closer look at the differences between the reference samples and the effect of pre-processing on PCA and one-class modeling patterns (Harnly and Upton, 2024). In Figure 9, samples from all four sources were combined to serve as reference samples and used for one-class modeling. The variance of the combined reference materials exceeded that of each modeled individually, i.e., the model for AHP in Figure 7C. Sensitivity was 97% for the combined A. racemosa samples and the specificity was 56% for other Actaea species, 55% for commercial Actaea roots, and 100% for commercial Actaea supplements. These data demonstrated the difficulty of differentiating between the Actaea species and that some of the commercial root samples are correctly identified as A. racemosa. None of the commercial Actaea supplements exhibited metabolic profiles similar to the reference samples, most likely due to the influence of sample preparation.

Figure 10 shows an alternate perspective for one-class modeling data. If the modeled data in Figure 9 are viewed from the side, they can be displayed as frequency plots. This is a more intuitive view, providing the mean and distribution of the populations. Figure 10 shows only the frequency plots for A. racemosa and the other Actaea species. Alternate pre-processing methods had little impact on the separation of the two populations. Use of only variables with a low variance provided improved overlap of the A. racemosa from the different sources (data not shown) but reduced the ability to discriminate between species (data not shown). Use of variables that enhanced discrimination between species (based on an f-test) also failed to enhance differentiation as the differences between the reference samples from the 4 different sources were also increased, i.e., the frequency plot for A. racemosa broadened (data not shown).

This study demonstrated that not all reference materials are created equal, use of multiple sources for reference materials increases the variance of the model, discriminating between species is difficult because the same metabolites are present but in a slightly different ratios, and finally, there is no simple pre-processing for improving discrimination between species.

Preparation of commercial supplements from botanical ingredients results in changes in the chemical composition of the supplement. This makes it difficult to authenticate supplements based on the raw ingredients. One approach to comparing the composition of different samples is cross-correlation of the two spectra (Harnly et al., 2017). Ions found in both the reference samples and supplements spectra will provide an enhanced numerical product while those missing from either will yield a product close to zero. Comparison of the cross-correlated spectra can be achieved using PCA and eliminates the need for a normalization factor. As usual, however, one-class modeling provides a clearer analysis with statistical evaluation.

E. purpurea (L.) Moench aerial samples were chosen as the test material and cross-correlation and one-class modeling were used to distinguish E. purpurea (L.) aerial samples from E. purpurea and Echinacea angustifolia roots and supplements. Authentic E. purpurea aerial and root samples and E. angustifolia root samples were obtained from Missouri Botanical Gardens, American Herbal Pharmacopoeia, and commercial sources. Twenty-three liquid and solid (tablets and capsules) supplements were purchased locally; 11 were single-ingredient supplements containing only E. purpurea and E. angustifolia aerial or root material and 13 were mixed or unknown Echinacea supplements labeled to contain mixtures of E. purpurea with E. angustifolia and/or Echinacea pallida (Nutt.) and consisting of either aerial or root material. Samples were analyzed by FIMS.

One-class modeling based on E. purpurea aerial samples showed that all other species and plant parts were significantly different from E. purpurea aerial (Figure 11). Thus, the compositional profiles of E. purpurea aerial could be differentiated from that of the E. purpurea roots and E. angustifolia roots and from any of the single ingredient or mixed supplements. The sensitivity for the model in Figure 11 was 99% and the specificity was 98%. These data reaffirm previous observations that most raw materials are not suitable for authenticating supplements. The preparation process for commercial samples (e.g., extraction, back extraction, drying, powdering, and possible addition of other materials) can result in significant differences in the chemical composition and their levels. It should be noted that his does not necessarily reflect on the supplements efficacy.

Figure 12 shows the average EPA spectrum cross-correlated with itself (auto-correlation), each of the EPA solid supplements (EPAS) (EPAS1-EPAS5), and EPA liquid supplements (EPAL). Only EPAS5 appears to be significantly different. One-class modeling of the cross-correlated spectra of the average EPA with each of the individual EPA samples (plot not shown) showed that the EPAS and EPAL spectra, with the exception of EPAS5 were the same as the cross-correlated EPA population. These results provide some interesting conclusions. First, since the correlation spectra for EPAS1-EPAS4, and EPAL were similar, the supplements contained similar extracted components. This means that the starting ingredients and the processes used by the manufacturers were similar. Second, since the auto-correlation spectrum of EPA was similar to the correlation spectra of EPA with EPAS1-EPAS4 and EPAL, the reference samples in this study were similar to the raw ingredients used by the manufacturers. Finally, since the extracted compounds were the same, the analytical extraction process used in this study was similar to the preparation method employed by the manufacturers. These data also indicate that supplement EPAS5 was produced using either a different starting ingredient or a different extraction process.

Figure 13 shows a one-class model for mixed supplements based on the cross-correlated spectra for EPA, EPAS1-EPAS4, and EPAL. All mixed supplements that were known to contain aerial plant material fell within the 95% confidence level, even if the species were not specified. All supplements with root material, even if the species were not specified, fell outside the 95% confidence limit. The exceptions falling outside the 95% confidence limit were supplements composed of both species and both plant parts. This suggests that perhaps the aerial portion was of insufficient concentration to provide a clear cross-correlated spectrum.

The data in Figures 12, 13 indicate that cross correlation of supplements with the raw ingredient accurately identifies ions found in both spectra. Ions common to both spectra provide an enhanced multiplicative product whereas ions found in only one spectrum provide a product close to zero. Thus, cross-correlation and one-class modeling can identify commonalities of the raw ingredients and supplements.

Maca data were presented earlier for the general description of one-class modeling. This study was a collaboration with the American Botanical Council; Gaia Herbs; Hong Kong Baptist University; Charles Sturt University & Therapeutic Research, TTD International Pty Ltd; and NSF International Authenticity Laboratory (Geng et al., 2020). There is some question as to whether the plant material is accurately identified as L. meyenii Walpers or Lepidium Peruvianum Chacon (Meissner et al., 2015). Samples consisted of 39 commercial maca supplements from 11 manufacturers, 31 unprocessed maca roots grown in Peru and China, and an historic non-tuber maca sample from Peru. Samples were analyzed using FIMS and DNA next-generation sequencing (NGS).

Initial, untargeted PCA placed all the maca samples in 3 classes: commercial maca samples, roots grown in Peru, and roots grown in China (Figure 3A). With one-class modeling the commercial Maca was readily differentiated from the raw root materials (Figures 3C, D). A similar approach, selecting either the Chinese or Peruvian samples for the model, showed that either could be differentiated from the other two classes (data not shown).

One-class modeling was also combined with analysis of variance (ANOVA) to differentiate roots based on country (China and Peru) and color (black, red, and yellow) (Figure 14). ANOVA was used to isolate the mean variance for each experimental factor (country, processing, and color) and cross factor to provide residuals free of the variance of the other factors (Harnly et al., 2014). One-class modeling of the individual factor residuals provided plots in Figures 14B–F. These data show, not surprisingly, that the metabolite composition correlates with country and color.

Metabolite profiling using ultra-high performance liquid chromatography-high resolution mass spectrometry (UHPLC-HRMS) in combination with PCA loadings was used to annotate the compounds responsible for differentiating between country and color. Genetically, all samples were confirmed to be similar and to be L. meyenii Walpers based on NGS at 3 gene regions (ITS2, psbA, and trnL) and comparison to recorded sequences of vouchered standards.

The results of this study show that one-class modeling can be used in combination with ANOVA to determine the statistical significance of differences arising from experimental factors associated with the samples. The metadata associated with genetics, country of origin, year of origin, climate, and handling can have a major impact on a plant’s metabolite profile and one-class modeling provides a versatile in deconvoluting the interaction of the factors.