Despite advances in proteomic technologies, idiosyncratic data issues, for example, incomplete coverage and inconsistency, resulting in large data holes, persist. Moreover, because of naïve reliance on statistical testing and its accompanying p values, differential protein signatures identified from such proteomics data have little diagnostic power. Thus, deploying conventional analytics on proteomics data is insufficient for identifying novel drug targets or precise yet sensitive biomarkers. Complex-based analysis is a new analytical approach that has potential to resolve these issues but requires formalization. We categorize complex-based analysis into five method classes or paradigms and propose an even-handed yet comprehensive evaluation rubric based on both simulated and real data. The first four paradigms are well represented in the literature. The fifth and newest paradigm, the network-paired (NP) paradigm, represented by a method called Extremely Small SubNET (ESSNET), dominates in precision-recall and reproducibility, maintains strong performance in small sample sizes, and sensitively detects low-abundance complexes. In contrast, the commonly used over-representation analysis (ORA) and direct-group (DG) test paradigms maintain good overall precision but have severe reproducibility issues. The other two paradigms considered here are the hit-rate and rank-based network analysis paradigms; both of these have good precision-recall and reproducibility, but they do not consider low-abundance complexes. Therefore, given its strong performance, NP/ESSNET may prove to be a useful approach for improving the analytical resolution of proteomics data. Additionally, given its stability, it may also be a powerful new approach toward functional enrichment tests, much like its ORA and DG counterparts.
Keywords: bioinformatics; networks; proteomics; systems biology; translational biology.