Prognostic relevance of gene-expression signatures

TL;DR

This paper investigates the limitations of gene-expression signatures in cancer prognosis, demonstrating that current signatures cannot exceed 80% accuracy due to inherent biological constraints and missing information.

Contribution

It introduces the concept that prognostic signatures are fundamentally limited and provides evidence from a large set of signatures showing a maximum achievable prognostic power.

Findings

Maximum prognostic accuracy is around 80%.

Over 50% of available information remains unexploited.

Emergent effects are crucial for improving prognostic signatures.

Abstract

Cancer prognosis can be regarded as estimating the risk of future outcomes from multiple variables. In prognostic signatures, these variables represent expressions of genes that are summed up to calculate a risk score. However, it is a natural phenomenon in living systems that the whole is more than the sum of its parts. We hypothesize that the prognostic power of signatures is fundamentally limited without incorporating emergent effects. Convergent evidence from a set of unprecedented size (ca. 10,000 signatures) implicates a maximum prognostic power. We show that a signature can correctly discriminate patients' prognoses in no more than 80% of the time. Using a simple simulation, we show that more than 50% of the potentially available information is still missing at this value.