The risk of over-interpreting in silico predictions as clinical evidence: a comment on Roy et al. (2024)

Abstract Roy et al. reported an extensive computational analysis of nonsynonymous single-nucleotide variants (nsSNPs) in the human CYLD gene and characterized several variants as potentially disease- or cancer-associated. Although such analyses can be valuable for generating hypotheses, computational predictions should be clearly distinguished from clinical or functional evidence. Among the 18 “high-risk nsSNPs” prioritized by the authors, only p.Glu747Gly (E747G) has established clinical evidence supporting pathogenicity, whereas the remaining variants lack sufficient evidence for pathogenic classification and should be considered variants of uncertain significance pending further evidence. We highlight three major concerns: the lack of independence among some prediction tools used for variant prioritization; the interpretation of somatic mutation clustering as evidence for germline disease association; and the presentation of variants with no established clinical or experimental evidence as disease-associated. Under the ACMG/AMP framework, computational evidence can provide supporting evidence but cannot substitute for clinical, population, segregation, or functional evidence. We suggest that the terminology and conclusions of the original article be revised to distinguish computationally predicted effects from demonstrated disease associations.