Limitations of Explainability for Established Prognostic Biomarkers of Prostate Cancer

Abstract

High-throughput technologies do not only provide novel means for basic biological research but also for clinical applications in hospitals. For instance, the usage of gene expression profiles as prognostic biomarkers for predicting cancer progression has found widespread interest. Aside from predicting the progression of patients, it is generally believed that such prognostic biomarkers also provide valuable information about disease mechanisms and the underlying molecular processes that are causal for a disorder. However, the latter assumption has been challenged. In this paper, we study this problem for prostate cancer. Specifically, we investigate a large number of previously published prognostic signatures of prostate cancer based on gene expression profiles and show that none of these can provide unique information about the underlying disease etiology of prostate cancer. Hence, our analysis reveals that none of the studied signatures has a sensible biological meaning. Overall, this shows that all studied prognostic signatures are merely black-box models allowing sensible predictions of prostate cancer outcome but are not capable of providing causal explanations to enhance the understanding of prostate cancer.

Are There Limits in Explainability of Prognostic Biomarkers? Scrutinizing Biological Utility of Established Signatures

“…We would like to emphasize that usually this distinction is not made explicitly but implicitly [22]. Examples for such a dual usage are abundant in the literature and some specific instances thereof are provided by markers that even entered the clinical practice [23], e.g., MammaPrint [24], Oncotype DX [25] or Prosigna/Pam50 [26].…”

“…Extending to disorders with complex etiology rather than Mendelian heredity such as cancer, in [22,32] the biological meaning of biomarkers has been further challenged. Specifically, instead of only selecting genes randomly to form new signature sets, as in [18], in [22,32] all signature genes and all genes involved in the same biological processes as the signature genes were removed (Remark: Even genes involved in proliferation were removed). In Figure 3, we show a visualization of the conceptual difference of these studies.…”

“…Figure 3B shows the biological processes containing such signature genes highlighted in magenta. While the study by [18] performed a random sampling from all genes, which can include genes in the original signature (corresponding to the weak removal I), the studies in [22,32] removed not only all signature genes but in addition all genes belonging to the same biological processes as the signature genes (corresponding to the strong removal III). Hence, the available pool of selectable genes is much smaller and those genes do not have any GO-terms in common with the signature genes.…”

“…3 B shows the biological processes containing such signature genes 144 highlighted in magenta. While the study by [18] performed a random sampling from 145 all genes, which can include genes in the original signature (corresponding to the weak 146 removal I), the studies in [22,32] removed not only all signature genes but in addition all For a weak removal of biological meaning there are two ways, one allows genes from all BPs (I) and one that removes only the signature genes (II) (A). In contrast, a strong removal selects only genes from GO-terms that do not contain any signature gene (B).…”

“…Interestingly, despite the further reduction in the size of the pool of selectable genes, the studies in [22,32] showed that even among those remaining genes there exist random gene sets that perform similarly in the prognostic prediction task. Such gene sets are called surrogate gene sets.…”