Commentary on: Elkins KM , Garloff AT , Zeller CB . Additional predictions for forensic DNA phenotyping of externally visible characteristics using the ForenSeq and Imagen kits. J Forensic Sci. 2023;68(2):608–13. https://doi.org/10.1111/1556‐4029.15215
Manfred Kayser · Journal of Forensic Sciences · 2023
In their article published by the Journal [1], Elkins et al. state that 15 SNPs included in the commercial ForenSeq and Imagen kits have “prediction capabilities” for 15 externally visible characteristics (EVCs) not targeted by these kits, that these additional EVCs “can be predicted” with these SNPs, and that these SNPs “can be used” in cold case investigations. Based on the data they present, the authors overstate their findings. Moreover, they understate previously established scientific knowledge on the complexity of the genetic basis and genetic prediction of human appearance traits. Below, I provide brief explanations. In their study [1], the authors' work was restricted to the inspection of the SNPs targeted by these two forensic kits in published outcomes of previously performed genome-wide association studies (GWASs). As a result, they highlighted 15 SNPs for which statistically significant associations with 15 EVCs not targeted by these kits (off-target EVCs) were reported in previous GWASs. The authors did not present empirical evidence that these SNPs actually have predictive value for these off-target EVCs. The authors did not present empirical evidence that these off-target EVCs can actually be predicted with these SNPs. Such evidence is not available from genetic association testing (done in GWASs) and can only be obtained from dedicated statistical prediction analyses, which the authors did not perform in this study. Based on the work described in their article [1], it appears that the authors simply assumed from the significant association results they found previously reported for these SNPs, that these SNPs have “prediction capabilities” for these off-target EVCs and that these additional EVCs “can be predicted” with these SNPs, without any further testing. However, the association is not a prediction. That an SNP is significantly associated with a phenotypic trait does not automatically mean that it has predictive value for that trait, nor does having a predictive value of an SNP mean that a trait can actually be predicted with such SNP with any practically useful prediction accuracy. Whether a significantly associated SNP has predictive value needs to be established empirically via dedicated prediction analyses using genotype–phenotype data independent of those used for GWAS-based SNP discovery. Whether an SNP with demonstrated predictive value can actually predict such a trait depends on the strength of the predictive effect, which needs to be established empirically from building and validating statistical prediction models using genotype–phenotype data that are independent of those used for GWAS-based SNP discovery. In their study, the authors did not carry out prediction analyses and did not present such empirical evidence. Because the association is not a prediction, simply assuming “prediction capabilities” and “can be predicted” from significant associations, as the authors did, is not justified and reflects overstating of their findings. Furthermore, by stating that these 15 off-target EVCs “can be predicted” with these 15 SNPs [1], the authors understate the wealth of previously established scientific knowledge on the complexity of the genetic basis and genetic prediction of human appearance traits. Even if these 15 SNPs have predictive value for these 15 off-target EVCs, which first would need to be empirically established via dedicated prediction analyses, current knowledge expects it to be extremely unlikely that their predictive effect would be large enough so that these 15 off-target EVCs can actually be predicted with these 15 SNPs. At least not on the prediction accuracy level required in Forensic DNA Phenotyping. Notably, the authors motivate their study by Forensic DNA Phenotyping, present their findings in the context of Forensic DNA Phenotyping, and state that these SNPs “can be used” in cold case investigations (and they published their paper in a forensic journal) [1]. As a consequence of the complex genetics of human appearance, where (except for red hair color) many genes determine any given trait with small phenotypic effects per each gene, consequently leading to small predictive effects per each SNPs, appearance traits cannot be predicted with single SNPs, nor with a few SNPs, except for blue and brown eye color. At least not with the prediction accuracy required in Forensic DNA Phenotyping. This is the reason why, with the only exception of eye color, all existing prediction tools for human appearance traits that are applied in Forensic DNA Phenotyping include many SNPs, that is, dozens to hundreds, for which the predictive effects were empirically established and validated in prediction modeling analyses (for a recent summary of the field of Forensic DNA Phenotyping see [2]). Overstating findings and understating previously established scientific knowledge, as the authors did with their article [1], does not help in further improving Forensic DNA Phenotyping so that it can contribute more and more effectively towards solving serious crime cases that otherwise remain unsolved because the perpetrators are not yet known with their forensic STR profiles to the law enforcement agencies during the time of the case investigation.