We used two related methods of extracting caries patterns in the permanent dentition from surface-level caries data. PCA yielded many moderate-to-weak patterns, possibly indicating a high degree of noise or sporadic (non-patterned) occurrence of dental caries. Moreover, PCs 1-3 closely recaptured the DMFT, PFS, and SMS indices, an observation that suggests these a priori
caries phenotypes may reflect the predominant patterns of decay in the permanent dentition, although cumulatively they account for only 37% of the variability. Some PCs were heritable, whereas many were not, which suggests that genetic patterns of decay may be separable from non-genetic patterns. Unlike PCA, FA did not yield factors that clearly recaptured a priori
phenotypes, with the exception that FAC1 was correlated with PFS. Maxillary incisors contributed heavily to FAC2, which is consistent with previous studies that used multidimensional scaling [24
] and cluster analysis [23
] to explore caries patterns in the primary dentition and showed maxillary incisors formed the second cluster (after other smooth surfaces). Ten factors were insufficient to explain the variability of the data, cumulatively accounting for approximately 45%.
Like PCA, FA yielded some factors that were highly heritable indicating that certain caries patterns may be due to genetic etiologies while others may be due to non-genetic etiologies. Because the caries patterns presented in this manuscript are more precisely and agnostically defined than a priori
phenotypes, we conservatively conclude that specific patterns represented by FAC3 and FAC6 are heritable, rather than generalizing to broader surface categories such as SMS. Interestingly, the strongest genetic contribution identified was for FAC3, which was 65.3% heritable (compared to 41.8% for D1MFS index) which suggests that FAC3 may be a better phenotype for gene discovery than a priori
caries phenotypes. A similar conclusion can be made for PC7 (50.3% heritable). These results are generally consistent with a previous study comparing PCA and FA that showed FA may better capture underlying genetic signals from correlated phenotype measurements (although both methods perform quite similarly) [35
]. Non-heritable PCs and FACs, presumably due to effects of non-genetic risk factors, may be preferred phenotypes for future epidemiological studies of environmental risk factors for dental caries.
The severity of caries significantly increased with age (or age2) for most patterns (results not shown). Heritability estimates were calculated while simultaneously modeling age, age2, and sex, although very similar heritability estimates were obtained in unadjusted models for all patterns except PC1 which exhibited decreased heritability when covariates were omitted (results not shown). These results are sensible given that altogether, age, age2 and sex accounted for about 10% of variance in PC1, but very little variance for the other PCs and FACs.
One of the challenges of using agnostic methods such as PCA and FA to identify underlying patterns of dental decay (devoid of a priori surface classifications) is in interpreting the findings. While some patterns, such as PC1 (defined by near-uniform loadings across most tooth surfaces and therefore representing global extent of decay), and FACs 7, 8 and 10 (each defined by contributions of a single pre-molar), are readily interpretable, other PCs and FACs may be difficult to relate back to the original variables. Moreover, there is no clear method of distinguishing biologically relevant patterns attributable to distinct risk factors from sporadic patterns due to noise. Sensitivity analysis showed that patterns represented by PCs 1-9 and FACs 1-6 were stable, whereas PC10 and FACs 7-10 were moderately stable. The overall stability lends credence to the notion that PCs and FACs considered in this study are not due to chance alone.
This study benefits from the large sample of related individuals with detailed surface-level caries assessment, which facilitated caries pattern extraction and heritability estimation. An additional strength of the analysis was using two different but related methods of extracting caries patterns from the data, which, most importantly, did not use a priori pattern definitions.
Despite these strengths, several limitations of this study warrant discussion, including inherent limitations to assigning tooth surfaces as carious or not. First, caries assessment by visual inspection, though suitable for obtaining data on large numbers of individuals and of sufficient quality for research purposes, may under-represent the true level of disease. Moreover, teeth missing due to decay, for which all surfaces count as carious, and approximal lesions which are often treated by two-surface restorations (leading to filled occlusal surfaces despite absence of decay) may cause caries assessment errors. Likewise, the quality of caries assessment may not be uniform across surfaces of the permanent dentition, which may have caused additional "noise" in the caries measurement. Lastly, prophylactic restorations may inflate caries assessment. These limitations are unavoidable for cross-sectional (i.e., single time point) study designs of dental caries. However, appropriate modeling techniques, such as methods of pattern extraction including PCA and FA, may aid in overcoming theses limitations of the caries assessment.