Multi-classifier proteomics to define complexes yields new chromosomal proteins

Abstract

In a recent issue of Cell, Ohta et al. (2010) report a method of quantitative proteomics coupled with bioinformatic analysis for the identification of associated components in complex mixtures. Using this approach, they assayed the protein composition of mitotic chromosomes, identifying 4029 associated proteins, 562 of which are previously uncharacterized.

The isolation of subcellular fractions and the identification of their constituents are instrumental to deciphering how cellular processes work. Unlike the genome, however, the proteome varies both temporally and across cell types, making it challenging to precisely define on a global scale. Currently, mass spectrometry provides the most comprehensive and versatile tool in large-scale proteomics. Combined with classic cell biological methods of isolation, mass spectrometry-based approaches have successfully identified hundreds of new protein components of specific organelles and subcellular structures ranging from mitochondria to the spliceosome to the mitotic spindle (Yates et al., 2005). Nevertheless, many key challenges to this approach remain. For example, the complexity of the spectra and proteomics data produced from tens of thousands of peptides can easily overwhelm the available bioinformatics tools. Beyond this, the Achilles’ heel (and perhaps the most severe limitation) of the approach is distinguishing real constituents of a complex from simple contaminants. The purity and the integrity of organelle preparations are frequently mutually exclusive properties -- the more complex the structure, the more likely that every effort to increase purity (and remove contaminants) will cause the loss of important but weakly associated true components.

In a recent Resource article in Cell, Ohta et al. (2010) propose a solution to these problems. The key to their approach is the use of Multi-Classifier Combinational Proteomics (MCCP) to identify functional relationships between proteins present in complex mixtures. There are three key parts to this method (Figure 1). First, the mass spectrometry is done quantitatively through the use of a heavy/light, Stable Isotope Labeling with Amino acids in Cell culture (SILAC) approach (Ong et al., 2002). Second, a set of “classifiers” (abundance, enrichment in the isolated complex, the degree of “exchangeability” of each protein within the complex, and any property(ies) unique to the complex being analyzed, such as the effects of specific mutations) is defined. Third, functional relationships between proteins are uncovered through a bioinformatics approach using what is called a random forest analysis (Breiman, 2001), a machine learning approach for detecting patterns in the behavior of groups of components so as to reveal their functional relationships and cluster them into functional relevant groups. The underlying logic and application of the random forest method will very likely be immediately apparent only to the bioinformatics expert, but for the rest of us it is important to point out three crucial strengths of this method. First, it enables analysis of data sets that contain missing values (e.g., when a protein is missing from one of several data sets, which is very common in mass spectrometry analysis). Second, it can be expanded indefinitely by adding more data sets. Third, data can be incorporated from any experimental approach in which the proteins of interest are sorted systematically (e.g. data from microarrays, yeast two hybrid screens, systematic protein pull-downs, protein phosphorylation and localization). In principle, this approach can be used to study any complex structure and may be especially useful for dynamic organelles and structures with transient components.

Figure 1. Using multi-classifier combinational proteomics (MCCP) to define mitotic chromosomal proteins.

Chromosome and non-chromosome samples were prepared using customized experimental procedures (gray) exploiting heavy (¹³C) or light (¹²C) amino acid labeling. Quantitative mass spectrometry of isolated peptides (green), multi-dimensional combination analysis of six different classifiers (yellow) and random forest analysis (orange) predicted chromosomal proteins (purple) that were experimentally tested (blue) using GFP fusions.

As a proof-of-principle of the power of their multi-classifier proteomic methods, Ohta et al (2010) focus on the analysis of the composition of mitotic chromosomes. The chromosome, a name coined in 1888 by the anatomist Heinrich Wilhelm Gottfried von Waldeyer-Hartz, has long been known to be a highly condensed, organized form of chromatin that appears as cells enter mitosis. However, its protein composition still remains undefined despite more than a century of intense study, largely because it is virtually impossible to purify large quantities of mitotic chromosomes free of cytoplasmic contaminants that adhere through nonspecific hydrophobic or electrostatic interactions. In contrast to previous smaller scale proteomic studies of mitotic chromosomes (Gassmann et al., 2005; Morrison et al., 2002; Porter et al., 2007; Takata et al., 2007; Uchiyama et al., 2005), Ohta et al. (2010) used large scale quantitative mass spectrometry to identify chromosomal proteins in the context of intact chromosomes. Using chromosomes purified from chicken cells, SILAC mass spectrometric techniques, and follow-up bioinformatics analyses, Ohta et al. (2010) identified 4029 mitotic chromosome-associated proteins. Using six classifiers that they defined and subsequent random forest analysis, they classified ~1530 of these proteins as “hitchhikers” – that is, non-chromosomal proteins remaining in the chromosomal preparation (which the authors argue are different from conventional contaminants in that they represent in vivo associations, rather than in vitro artifactual association after cell lysis).

The key to the success of the MCCP approach in distinguishing bona fide protein associations from the non-specific is the definition of six different classifiers, each providing an independent measure of the association of a protein with the complex being studied, in this case mitotic chromosomes (Figure 1). The first classifier – abundance – is perhaps the most obvious. The authors determined this value for each protein in mitotic chromosomes using a standard curve produced from proteins of known abundance and a scaled protein abundance index based on the number of peptides observed and the spectral count for each protein. (As expected, histones scored highly here, comprising almost half of total chromosomal protein mass.) A second classifier exploited the use of heavy/light isotopic labeling of amino acids to estimate the enrichment of individual proteins in an isolated chromosomal fraction (light) versus the amount in a cytoplasmic extract (labeled with heavy ¹³C amino acids). Classifier III measured the ability of a protein to stably bind to (light) chromosomes during an in vitro incubation in a (heavy) cytosolic supernatant. Classifier IV exploited differences in chromosome composition that follow from the loss of a component that is a key player in mitotic chromosome condensation and structural integrity. Using a conditional deletion of the condensin subunit SMC2, correctly condensed chromosomes from SMC2-containing (light labeled) cells were mixed with samples from SMC2-depleted (heavy labeled) cells and the relative abundance of each protein determined by the heavy/light ratios. To focus on centromere proteins and those associated with the spindle microtubule-binding kinetochore that assembles onto the centromere early in mitosis, Classifier V was modeled after Classifier IV except with a focus on the chromosomal abundance of each protein after deletion of the non-essential kinetochore component Ska3/Rama1. Finally, Classifier VI was defined by performing bioinformatic analysis on the domains in each protein of interest, as a means of identifying chromosomal proteins by the presence of domains more frequently found in known chromosome-bound proteins than cytosolic proteins.

Each classifier on its own lacks a clear boundary between bona fide chromosomal proteins and cytosolic proteins, but the strength in the MCCP approach comes from the combination of different classifiers and random forest analysis to allow for simultaneous integration of the information present in all of the proteomics classifiers. Indeed, the combined proteomic classifiers assigned 118 of 125 known centromere proteins correctly as chromosomal, a 94.4% success rate. Of the 4029 chromosomal proteins identified, 562 were previously uncharacterized. To validate the strength of the prediction that these are indeed chromosomal proteins, the authors analyzed the localization of 50 of these uncharacterized proteins tagged with GFP and showed that 88% localized as predicted. In addition, this analysis demonstrated that 13 of these 50 newly cloned proteins are associated with kinetochores in mitosis, thus expanding the list of known centromere proteins beyond CENP-Z up to CENP-36. Among these, the authors further characterized CENP-32, showing that its reduction by siRNA disrupted spindle morphology and chromosome alignment. Perhaps most provocatively, the analysis predicts that 97 more centromere proteins remain to be discovered in the pool of the 510 remaining uncharacterized chromosomal proteins uncovered in the current study. Identification of these proteins and characterization of their role in the structure and function of centromere will serve as a starting point for in-depth systematic analysis of mechanisms involved in mitotic chromosome segregation.