treeClust improves protein co-regulation analysis due to robust selectivity for close linear relationships

Abstract

Gene co-expression analysis is a widespread method to identify the potential biological function of uncharacterised genes. Recent evidence suggests that proteome profiling may provide more accurate results than transcriptome profiling. However, it is unclear which statistical measure is best suited to detect proteins that are co-regulated. We have previously shown that expression similarities calculated using treeClust, an unsupervised machine-learning algorithm, outperformed correlation-based analysis of a large proteomics dataset. The reason for this improvement is unknown. Here we systematically explore the characteristics of treeClust similarities. Leveraging synthetic data, we find that tree-based similarities are exceptionally robust against outliers and detect only close-fitting, linear protein -protein associations. We then use proteomics data to demonstrate that both of these features contribute to the improved performance of treeClust relative to Pearson, Spearman and robust correlation. Our results suggest that, for large proteomics datasets, unsupervised machine-learning algorithms such as treeClust may significantly improve the detection of biologically relevant protein -protein associations relative to correlation metrics. every dataset, as the optimal choice of measure depends on various characteristics of a given dataset, such as the frequency of outliers and missing values.A recent, fundamental change to the expression profiling setup was made possible by improvements in the field of quantitative proteomics: the use of protein abundances rather than mRNAs as readout for gene activity. This increases the accuracy of gene function prediction, because protein abundances are better indicators of gene function than mRNA levels, at least in human (9-11) and mouse (12) . We have recently reported ProteomeHD, a dataset that quantifies the response of 10,323 human proteins to 294 biological perturbations using isotope-labelling mass spectrometry (13) ( https://www.biorxiv.org/content/10.1101/582247v1 , in revision). ProteomeHD is a heterogeneous dataset, incorporating a wide range of perturbation experiments from different laboratories, such as inhibitor treatments, differentiation time courses and cancer cell line comparisons. We compared different coexpression measures for their ability to detect proteins that are co-regulated in response to these perturbations. Surprisingly, we found that the unsupervised machine-learning algorithm treeClust (14, 15) provided a striking improvement over established correlation-based metrics. However, the reason for this improvement remained unclear, because treeClust is a novel algorithm that works in a fundamentally different way to previously used coexpression metrics.The treeClust algorithm uses recursive partitioning (16-18) to create decision trees. Such trees are normally used for supervised classification or regression tasks. In contrast, treeClust uses decision trees to calculate a dissimilarity measure in an unsupervised manner. To do so, treeClust first creat...

The human proteome co-regulation map reveals functional relationships between proteins

“…However, experiments with synthetic data also show that treeClust works best for large datasets with 50 samples or more, depending on additional parameters such as the frequency of missing values. Traditional correlation analysis may be better suited for smaller gene expression datasets 54 .…”