Performance Is Not All You Need! Comment on “Unsupervised Machine Learning for Differential Analysis in Proteomics”

In their recently published article (DOI: 10.1021/acs.analchem.5c03117), Xu et al. argue that the detection of differentially abundant proteins in proteomic experiments should go beyond traditional statistical methods and should leverage unsupervised machine learning for anomaly detection. Shedding light on this category of methods is insightful, and the reported performances are promising. However, we believe the benchmarking angle of this article is restrictive. First, the reported performance increments are associated with overstated theoretical differences. Second, an excessive focus on the performances could lead proteomic investigators to undermine their usual elicitation of the biological question. As both reasons pertain to the researchers' empowerment of machine learning tools and of computational workflows, we believe it is important to formulate complementary guidelines.