RE: Exploring the cross-cancer effect of circulating proteins and discovering potential intervention targets for 13 site-specific cancers

To the Editor:

We are writing in response to a recently published article by Sun et al. in the Journal (1). This study applied mendelian randomization, an epidemiologic approach that uses germline genetic mutations as instruments (“proxies”), to evaluate the effect of 3991 plasma proteins on risk for 13 cancers. The authors constructed instruments by identifying genetic variants strongly associated with each protein and report finding “convincing causal evidence” of the effects of 39 proteins on risk for 7 cancers.

To provide a valid test of the causal null hypothesis, mendelian randomization requires the strong and unverifiable assumption that the genetic instruments do not influence an outcome (eg, cancer risk) through pathways independent to the exposure (eg, a protein), termed horizontal pleiotropy (2). It is widely accepted that genetic variants that reside in or near a protein’s coding gene (cis-protein quantitative trait loci) are more likely to satisfy this assumption than are variants that are distal to a protein’s cognate gene (trans-protein quantitative trait loci) (3,4) because of the greater likelihood that cis-protein quantitative trait loci directly affect the transcription or translation of the protein itself, thus having direct biological relevance and greater specificity to the protein being instrumented. In contrast, trans-protein quantitative trait loci frequently associate with large numbers of proteins, along with traits upstream of proteins (eg, body mass index, dietary patterns) that can confound protein-outcome associations (termed correlated horizontal pleiotropy) and factors associated with sample handling (2,5). A reliance on trans-protein quantitative trait loci without evidence from cis-protein quantitative trait loci when constructing genetic instruments is therefore more likely to induce bias and runs counter to standard practice in protein mendelian randomization analysis.

In their study, Sun et al. neither restricted their instruments to cis-protein quantitative trait loci nor acknowledged the distinction between cis-protein quantitative trait loci and trans-protein quantitative trait loci. In fact, analyses for only 17 of 56 proteins reported as putatively causal in initial mendelian randomization analyses and mapped to a unique gene in Ensembl or with known alternate cognate gene name included cis-protein quantitative trait loci (defined as ±1 megabase from the cognate gene) in the instruments. Further, it is not possible from the results Sun et al. provided to assess whether the role of these 17 proteins in cancer risk was informed or refuted by evidence from these cis-protein quantitative trait loci. Instead, the majority of proteins tested were exclusively instrumented using potentially pleiotropic trans-protein quantitative trait loci—73% of proteins (2780 of 3827 proteins with gene coordinates available in Ensembl) did not include any cis-protein quantitative trait loci in instruments. One highly pleiotropic single-nucleotide variation (formerly single-nucleotide polymorphism) in particular, rs10922098, was used to instrument 1115 distinct proteins, approximately 30% of all proteins tested.

A consequence of this widespread use of trans-protein quantitative trait loci as instruments is the inability to confidently conclude whether putative “causal effects” of any of the proteins identified in this analysis are driven by effects of these proteins or that of other traits influenced by the same instrument, representing violation of a core mendelian randomization assumption. Importantly, although Sun et al. report finding “no significant pleiotropy” in their analyses, the authors’ use of 1) co-localization on trans-protein quantitative trait loci and 2) “pleiotropy-robust” models that have limited statistical power to detect horizontal pleiotropy when instruments consist of small numbers of variants does not provide credible evidence against findings being driven by horizontal pleiotropy (6). Indeed, the authors’ use of “pleiotropy-robust” models that implicitly remove or down-weight outlying variants (eg, median and mode-based estimators) are known to induce bias in settings where variants that appear to be outliers are in fact the most biologically reliable (eg, if instruments consist of many trans-protein quantitative trait loci but only 1 cis-protein quantitative trait locus) (7). The combined extensive use of trans-protein quantitative trait loci as instruments along with inadequate sensitivity analyses to evaluate the presence of horizontal pleiotropy thus calls into question the robust causal claims relating proteins to cancer risk that Sun et al. make.

Protein mendelian randomization remains a potentially powerful approach for enhancing our understanding of the molecular drivers of carcinogenesis and for prioritizing novel therapeutic targets for cancer prevention and treatment. To realize the promises of protein mendelian randomization, careful evaluation of methods employed and interpretation of findings must accompany any analysis using this approach.

Data availability

No data are reported in this correspondence.

Author contributions

James Yarmolinsky, PhD (Conceptualization; Writing—original draft; Writing—review & editing), Ioanna Tzoulaki, PhD (Writing—review & editing), Marc J. Gunter, PhD (Writing—review & editing), Ruth C. Travis, DPhil (Writing—review & editing), George Davey Smith, DSc (Writing—review & editing), Karl Smith-Byrne, DPhil (Conceptualization; Writing—original draft; Writing—review & editing).

Funding

J.Y. and I.T. are supported by the National Institute for Health and Care Research Imperial Biomedical Research Centre. R.C.T. and K.S.B. are supported by Cancer Research UK (grants No. C8221/A29017 and C16077/A29186) and UK Research and Innovation grant No. 10063259.

Conflicts of interest

The authors have no conflicts of interest to report.

Acknowledgements

The funders did not play a role in the writing of the correspondence or in the decision to submit the correspondence for publication.