Assessing the functional structure of genomic data
1Department of Computer Science, Princeton University, 35 Olden Street, Princeton, NJ 08540 and 2Lewis Sigler Institute for Integrative Genomics, Carl Icahn Laboratory, Princeton University, Princeton, NJ 08544, USA
*To whom correspondence should be addressed.
 |
ABSTRACT |
|---|
 TOP ABSTRACT 1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
Motivation: The availability of genome-scale data has enabled an abundance of novel analysis techniques for investigating a variety of systems-level biological relationships. As thousands of such datasets become available, they provide an opportunity to study high-level associations between cellular pathways and processes. This also allows the exploration of shared functional enrichments between diverse biological datasets, and it serves to direct experimenters to areas of low data coverage or with high probability of new discoveries.
Results: We analyze the functional structure of Saccharomyces cerevisiae datasets from over 950 publications in the context of over 140 biological processes. This includes a coverage analysis of biological processes given current high-throughput data, a data-driven map of associations between processes, and a measure of similar functional activity between genome-scale datasets. This uncovers subtle gene expression similarities in three otherwise disparate microarray datasets due to a shared strain background. We also provide several means of predicting areas of yeast biology likely to benefit from additional high-throughput experimental screens.
Availability: Predictions are provided in supplementary tables; software and additional data are available from the authors by request.
Contact: ogt{at}princeton.edu
Supplementary information: Supplementary data are available at Bioinformatics online.
|
1 INTRODUCTION |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
The technological developments of the past several decades have driven a continuing expansion of our understanding of molecular biology and a similar expansion in the analysis techniques applied to this data. In particular, genome-scale assays for coexpression (Eisen et al., 1998; Spellman et al., 1998), genetic interactions (Giaever et al., 2002; Tong et al., 2004), physical interactions (Gavin et al., 2002; Ho et al., 2002), protein localization (Huh et al., 2003) and regulatory networks (Harbison et al., 2004; Zhu and Zhang, 1999) have all opened up new opportunities for computational data mining that have been richly explored. Data such as these have been used in a variety of machine learning and other computational contexts (Franke et al., 2006; Jansen et al., 2003; Karaoz et al., 2004; Lee et al., 2004; Troyanskaya et al., 2003).
As the amount of available genome-scale data has continued to increase, it has become possible to ask higher level questions about the systems-level functional associations between entire pathways and processes. These associations represent the complex interplay between linked biological processes: DNA replication and mitosis are distinct cellular processes, for example, but they are functionally associated in their biological goals (cell division), regulation and genetic participants. Understanding this network of associations between processes is a critical link between functional relationships at the single-gene level and phenotypes at the organismal level.
By deriving an understanding of large-scale functional structure based directly on genome-scale datasets, we also gain an understanding of the data itself. An examination of the pathways and processes perturbed by whole-genome experiments allows those experimental results to be described in terms of their functional activity. For example, microarrays performed under conditions of heat shock and oxidative stress might both show functional activity related to an environmental stress response; this similarity of functional activity reveals biological commonalities between otherwise disparate experiments. By combining these two lines of inquiry—functional associations between processes and functional similarities between datasets—we gain insight into unexpected relationships in existing data, and we can direct experimenters to biological areas that are currently unexplored. All of these analyses deal with the high-level functional structure of genome-scale data and biological processes, which allows us to answer increasingly complex questions using the ongoing flood of high-throughput data.
We present such an analysis of functional associations among 141 biological processes and over 180 datasets (spanning >950 publications, >2300 microarray conditions, and several thousand interaction, localization and sequence-based data) in Saccharomyces cerevisiae, where a functional association entails co-operation, coregulation or other interaction between pathways and processes to perform a cellular task. These associations are derived by examining functional relationships between many individual genes, which are in turn predicted in a process-specific, probabilistic manner from heterogeneous data integration. This provides a global view of the functional structure of biological processes in yeast, including the degree of data-driven associations between processes, the experimental cohesiveness of gene behavior within each process, and the coverage of individual biological processes by currently available data. Likewise, we obtain measures of functional activity within each dataset—that is, which biological processes are covered by a dataset, independently of experimental platform. This high-level functional analysis technique is not specific to yeast and is extensible to any organism with a sufficiently large body of experimental data.
This analysis of functional structure produces a number of findings useful for guiding future experimental efforts and further computational studies. Specifically, we provide maps of data-driven associations between biological processes and of similar functional activities among datasets. By examining associations between processes, we observe several biological processes that could benefit from additional high-throughput data coverage, including ion homeostasis and transport and mitochondrion organization. We also highlight biological processes likely to be performed by currently uncharacterized genes (e.g. autophagy). Similar functional activities among datasets demonstrate commonalities in several large microarray studies and consistency between protein localization, synthetic lethality and protein–protein interaction screens. These similarities also expose specific biological relationships, such as a subtle effect due to strain background we discovered in three otherwise diverse microarray datasets. All of these relationships are fundamentally driven by similarities in gene and protein response across hundreds of datasets, and this high-level analysis of such large-scale functional structure is valuable for guiding future experimentation and in understanding systems-level associations among biological processes.
|
2 METHODS |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
In summary, we analyzed the large-scale structure of functional relationship networks predicted based on Bayesian integration of genomic data. Functional associations between biological processes from the Gene Ontology (GO; Ashburner et al., 2000) were derived by further integration and analysis of these networks in a context-sensitive manner. Functional activity information for each dataset was calculated during the integration process, and this was used to further characterize functional similarities between datasets. The resulting process/process, process/dataset and dataset/dataset association networks were mined for subgraphs and interactions of high weight. All network visualization was performed using Graphviz from AT&T (Gansner and North, 2000).
2.1 Data collection and gold standard generation
2.1.1 Data collection
The data employed in this study is a union of that from Hibbs et al. (2007) and Myers and Troyanskaya (2007). Non-expression data includes pairwise physical and genetic interaction data from a variety of databases (Alfarano et al., 2005; Stark et al., 2006), protein localization (Huh et al., 2003), and sequence and TFBS similarities (Harbison et al., 2004; SGD, 2006). Pairwise interaction data were represented as binary presence/absence values; where applicable, interaction profile similarities were calculated between genes from binary data using an inner product. For details, see Myers and Troyanskaya (2007).
Expression data was collected from 
80 publications comprising 120 datasets and 2300 conditions as described in Hibbs et al. (2007) and initially processed as described in Huttenhower et al. (2006). Datasets containing fewer than four experiments were initially merged, creating a merged microarray set that was subsequently processed identically to the remainder of the datasets. Each of these was converted from expression values to gene pair similarity scores using Pearson correlation normalized using Fisher's z-transform (David, 1949) and subsequently z-scored:
|
| (1) |
|
| (2) |
, and the final similarity between two genes z(gi, gj) is the pair's Fisher score minus the mean Fisher score µf divided by the Fisher score SD 
f (both over all gene pairs).
After z-scoring, each expression dataset was quantized using the binnings (–
,–1.5), [–1.5, –0.5), [–0.5, 0.5), [0.5, 1.5), [1.5, 2.5), [2.5, 3.5), [3.5, ); these represent steps of 1 SD in z-score space. Mutual information was calculated between the resulting sets of discrete values, and any pairs of datasets sharing >15% of the possible information were merged by averaging z-scores. PISA (Kloster et al., 2005) modules (a biclustering algorithm) were also calculated for the expression data collection and transformed into pairwise scores for our analysis by counting the number of times each pair of genes coclustered after 500 iterations. These biclusters offered an orthogonal analysis of the microarray data capable of providing different information than the normalize correlation scores.
2.1.2 Gold standard generation
To perform supervised learning, we generate a gold standard of known functionally related and unrelated gene pairs. Biological processes of interest were selected from the GO (Ashburner et al., 2000) using a method based on Myers et al. (2006). The standard developed in Myers et al. (2006) is specific to S.cerevisiae; using a similar voting method and polling six biologists, a set of 433 GO terms were selected for this study to be experimentally informative independent of organism. Of these 141 have at least 10 gene annotations in S.cerevisiae, and these were selected as processes (gene sets) of interest (Supplementary Table 1).
|
An answer set was derived from these processes of interest as described in Huttenhower et al. (2006). Gene pairs coannotated to any of the 141 terms were considered to be related. A gene pair was unrelated in the gold standard if (1) the two genes were both annotated to some term in the set of 141, (2) the genes were not coannotated to any of these terms and (3) the terms to which the genes were annotated did not overlap with hypergeometric P-value <0.05. All other gene pairs were omitted from the standard (i.e. they were neither related nor unrelated for training and evaluation purposes).
For context-specific learning, this answer set was decomposed into subsets relevant to each process of interest. A gene pair was considered to be relevant to a biological process if either (1) both genes were annotated to the process or (2) one of the two genes was annotated to the process and the pair was unrelated in the standard (i.e. not coannotated to another process).
2.2 Bayesian analysis
2.2.1 Learning Bayesian classifiers
One naive Bayesian classifier (Neapolitan, 2004) was learned per biological process of interest; experiments with other network structures were shown to provide negligible performance improvements (Huttenhower and Troyanskaya, 2006). Briefly, a global classifier was learned in which the class to be predicted was gene pair functional relationships (as defined in the gold standard) and each dataset formed one node in the network. One hundred and forty-one function-specific networks were learned with identical structures, each using a subset of the global gold standard as described above. When fewer than 25 gene pairs were available for a particular dataset/relationship combination, the global probability distribution was used for that condition. This defines the predicted probability of functional relationship between genes as a weight:
|
| (3) |
All Bayes network manipulation was performed with a combination of custom C++ software and the SMILE library from the University of Pittsburgh Decision Systems Laboratory (Druzdzel, 1999).
2.2.2 Predicting functional relationships
Each naive Bayesian classifier directly implies a functional relationship network in which nodes represent genes and edge weights consist of the posterior probabilities of functional relationships between gene pairs. The 141 function-specific networks were combined to form a predicted global interaction network by transforming each network's edge weights to z-scores (subtracting the mean predicted probability and dividing by their SD) and averaging each gene pair's weight across all available networks.
2.3 Functional relationship and dataset enrichment predictions
2.3.1 Process/process relationships
As described above, for the purposes of this analysis, a biological process was defined as a set of related genes. The strength of a predicted functional relationship between two processes F and G was calculated as the average edge weight in the global interaction network within the edge set:
|
| (4) |
Similarly, the functional cohesiveness of a process was measured as the ratio of the average edge weight in the process to the average edge weight incident to the process:
|
| (5) |
2.3.2 Process/dataset relationships
The predicted enrichment of each dataset within each biological process was derived from the conditional probability tables learned for that dataset's node within the appropriate function-specific Bayesian classifier. Specifically, the predicted enrichment for process F in dataset D was calculated as the weighted sum of the difference in posterior probability of functional relationship induced in F's classifier by evidence from each possible value of D:
|
| (6) |
0.005. The exact value may differ due to rounding in this example. The estimated coverage of a process in currently available data was calculated as the average of rel(F,D) over all datasets in our study.
2.3.3 Dataset/dataset relationships
This calculation of predicted process/dataset enrichments results in a vector of 141 values in the range [0, 1] for each dataset. To determine the functional similarity between two datasets, each value is first transformed to a log ratio against the average across all datasets:
|
| (7) |
2.3.4 Gene/function relationships
For the purpose of predicting gene function based on ‘guilt by association’ with known genes in some process, the connectivity of a gene to a process was assessed as follows. Each gene/process pair was assigned a functional association score equal to the ratio of its average probability of functional relationship to the process over the process's cohesiveness:
|
| (8) |
2.3.5 Robustness
A robustness study was carried out by randomly shuffling data points within each dataset prior to Bayesian learning. The resulting networks had average dataset functional enrichment scores of 4.46x10–5±1.57x10–4, biological processes cohesiveness of 1.37±1.32, and association between processes of 7.14x10–3±0.0293, the last due to the greatly reduced differentiation between processes. In contrast, the averages for these values in Supplementary Tables 1–3 are 2.43x10–4±6.02x10–4, 15.1±35.9, and 1.94x10–3±0.141, respectively.
2.3.6 Dense subgraphs
An implementation of a modified greedy heuristic for discovering heavily weighted subgraphs (Charikar, 2000) was used to mine interaction networks for cohesive modules. Briefly, to discover each module within the network of interest, a node set was initialized with the most cohesive pair in the network. Nodes were added to this set greedily based on edge weight until no node could be added without reducing the average cohesiveness of the node set below the network baseline. The average edge weight of the set was then subtracted from each edge between nodes in the set, and the process was iterated to discover the next module. In pseudocode:
- N=argmax{gi,gj} cohes({gi, gj})
- Loop:
- g=argmaxgcohes(N
{g})
- If cohes(N
- N=N
- If |N|>2, output N
- Let
be the average edge weight among nodes in N
- For each gi, gj
N
- Repeat from 1
|
3 RESULTS |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
By analyzing functional associations among biological processes and functional similarities between high-throughput datasets in a purely data-driven manner, we summarize knowledge from thousands of whole-genome experiments in a biologically informative way. This includes descriptions of the cohesiveness, data coverage and associations of biological processes (Fig. 1), which can guide experimenters towards promising targets for future experimental work (Table 1). Datasets can also be compared based on functional activity, allowing the detection of large-scale functional similarity between the effects of experimental perturbations (Figs 2 and 3). These analyses provide an important global summary of interplay between pathways, and they identify processes, process associations and dataset similarities likely to benefit from experimental investigation.
|
|
|
3.1 Discovering data-driven functional associations between biological processes
Two or more biological processes can interact and work together to perform cellular functions in a manner analogous to a relationship between individual genes. A pair of genes might be functionally related if they operate in the same complex, pathway or transcriptional module. Our focus is at a higher level, where two processes might be functionally associated if they interact to achieve the same cellular goals; for example, nutrient sensing and the translation of new proteins at the ribosomes are distinct processes, but they interact to allow controlled cellular growth. These process–process associations are thus an extension of gene functional relationships: processes are functionally associated if they achieve related cellular goals, and we predict such an association if their constituent genes behave similarly in datasets determined to be good functional indicators. A small segment of our predicted process association network appears in Figure 1, made up of only the most confidently associated biological processes (see Supplementary Table 1 for complete results).
The edges in this process association network summarize information regarding the interactions between biological processes. A single biological process is internally cohesive in the currently available experimental results if its constituent genes also show strong individual functional relationships. If most gene pairs within a process are confidently functionally related, that process is reflected well by the available data: its annotations are in agreement with measured cellular behavior. If gene pairs within a process are related with low confidence, it often indicates an area of biology where further experimentation or annotation efforts may be most beneficial. The cohesiveness of biological processes in Figure 1 is represented by node color, where more cohesive processes appear in brighter yellow.
Finally, we also determined the degree to which each biological process is covered by available data. Our integration method provides a statistical measure of how active each biological function is within each dataset; we can thus sum over all datasets to estimate a biological process’ total representation within the data. This coverage measure is summarized by border width in Figure 1, with thicker borders indicating well-covered processes. Cohesive biological processes (yellow nodes) not covered well by available data thus represent promising candidates for future investigation: they show evidence of strong functional similarity, but they may not yet have been specifically targeted by high-throughput studies.
This interplay between functional associations, cohesiveness and data coverage is evident in several of the example processes in Figure 1. Ribosome biogenesis and rRNA metabolism, for example, are processes strongly evident in most microarray data (Myers et al., 2006), and this ubiquity is demonstrated by their extremely strong coverage and association. They are not as cohesive as many other processes, however, due to the large number of snRNAs and rRNAs annotated to these processes for which little or no high-throughput data is available. This analysis thus highlights an area for future exploration, even in an area as thoroughly studied as the ribosome. Other processes with relatively low coverage for their size (data not shown in Fig. 1) include protein complex assembly, ion homeostasis and transport and mitochondrion organization, all representing opportunities for future directed screens. Processes with low cohesiveness can either be particularly diverse (e.g. amino acid and derivative metabolism, protein processing) or not yet fully characterized, representing further opportunities for future experimental investigations.
3.1.1 Processes predicted to be enriched for uncharacterized genes
Networks of functional associations between processes represent a richly structured summarization of high-throughput data; they implicitly encode predicted details regarding pathway structure, association between gene sets and the functional diversity of currently available data. In addition to associating known processes and pathways, though, similar relationships can also be inferred to find areas of biology enriched for uncharacterized genes. These represent specific processes for which targeted genomic screens might uncover substantial new information.
A selection of processes that we find to be highly associated with uncharacterized genes is shown in Table 1, in addition to statistics describing the processes (see Supplementary Table 3 for complete results). The autophagy term, despite being the smallest and most cohesive process in this subset, still maintains a very strong association with uncharacterized genes. It is moderately well covered by available data, falling roughly in the middle of our 141 coverage estimates; it is thus possible that further information regarding autophagy could be gleaned from existing data, even though few experiments have specifically investigated the process in yeast. However, this predicted association with uncharacterized genes also suggests that substantial new functional assignments could be made by targeted screens for involvement in autophagy.
3.2 Similar functional activity in high-throughput datasets
While most high-throughput experiments are designed with fairly specific goals in mind, almost every dataset contains information about a variety of biological processes, and our analysis provides several ways of exploring these data. Our Bayesian learning process results in a probabilistic score indicating the activity of each biological process within each dataset. Collecting all such scores for a single dataset results in a functional profile for the dataset, and these numerical vectors can be compared between datasets to evaluate functional similarity. The network in Figure 2 contains a selection of datasets with similar functional activities (see Supplementary Table 2 for complete results).
Even in this small subset of analyzed datasets, several patterns are apparent. On the left, the first of the two main clusters contains primarily localization data from (Huh et al., 2003). Within the localization subsets, dataset similarity is correlated with cellular localization: the periphery and the bud are associated with the main body of data by way of actin, the Golgi stages are associated with each other, the endosome and peroxisome are related, and so forth. Three synthetic genetic array screens are also similar to the localization data. Davierwala et al. (2005) is associated primarily with the Golgi and ER, and one of the primary findings of this study was the characterization of PGA1, a gene essential for ER activity. Krogan et al. (2003) and Zhao et al. (2005) show similar functional activity to a variety of localization subsets (including several not shown in Fig. 2) and to Krogan et al. (2004), all of which are enriched for nuclear functions (DNA packaging, chromosome organization, transcription, RNA elongation, etc.) These functional similarities were generated solely by automatic data mining and call out important biological associations between disparate experimental results.
On the right, the cluster of microarray data is centered around a core of large datasets exploring very diverse conditions and thus enriched for many different biological processes (Brem and Kruglyak, 2005; Brem et al., 2002; Hughes et al., 2000; Yvert et al., 2003). The other main components of the cluster are stationary-phase growth and carbon metabolism (Brauer et al., 2005; Ideker et al., 2001; Martin et al., 2004; Pitkanen et al., 2004; Segal et al., 2003) and various stresses (Bro et al., 2003; Gasch et al., 2000; Jelinsky et al., 2000; O'Rourke and Herskowitz, 2004). Interestingly, (Bulik et al., 2003; Chitikila et al., 2002), and (Schawalder et al., 2004) are all likely included due to their use of galactose-inducible promoters while investigating other diverse processes; these datasets all share a carbohydrate metabolism enrichment in addition to their more specific targets [e.g. biopolymer biosynthesis, a parent of chitin biosynthesis, in Bulik et al. (2003)]. This demonstrates the power of associative functional analysis to uncover both primary and secondary enrichments, a consideration essential to getting the most out of any experimental result.
3.3 Simultaneous association of datasets and biological processes
Because our method assesses functional activity within datasets, functional similarities between datasets and associations between biological functions, it provides a means of coclustering datasets and processes in a biologically meaningful way. This raises the possibility of exploring complex data, potentially summarizing millions of individual measurements, in an intuitive manner. Each predicted weight between two datasets, two processes or a dataset and a process represents a measure of similar biological function, and thus an investigation of heavily weighted subgraphs in this space provides a way of exploring groups of related data and processes.
An example of such a cluster appears in Figure 3, which highlights one of the densest functional areas and the datasets in which these functions are most active. This consists of metabolic processes including alcohol, aldehyde and carbohydrate metabolism, cellular respiration, hydrogen and electron transport, and mitochondrion biogenesis; while they have been removed for visual clarity, several other related processes are also members of this cluster, including cofactor metabolism, autophagy and aging. The group of associated microarrays again represent a combination of broad genomic response (Brem and Kruglyak, 2005; Yvert et al., 2003), carbon metabolism (Schawalder et al., 2004; Segal et al., 2003) and stresses (Gasch et al., 2000), the latter likely included due to the relationship between stress response and growth rate (Brauer et al., 2008). These are linked into the cluster of biological processes primarily through carbohydrate metabolism, but also through the biclustering modules (PISA). These biclustering results incorporate all of the available microarray conditions, in contrast to the normalized correlation scores used to analyze individual datasets. Biclustering thus represents a view of expression data orthogonal to pairwise correlations and tends to be more sensitive to metabolic functions in general (phosphorus, amino acid and nitrogen compound metabolism in addition to those appearing in Fig. 3).
The non-microarray datasets associated with this functional cluster are diverse, including mitochondrial localization (in association with several mitochondrial and respiratory functions), cytoplasmic localization (in association with more general metabolism), two sequence-based analyses [downstream sequence similarity and shared transcription factor binding sites from Harbison et al. (2004)] and synthetic lethality interaction profiles from GRID (Stark et al., 2006) and BIND (Alfarano et al., 2005). Synthetic lethality profiles and shared binding sites both provide good coverage of many biological processes and are included largely due to moderate association with many of the functions within the cluster (most edges are not shown in Fig. 3); this is reflected in their relative isolation in the network. Broad downstream (and upstream) sequence similarity tends to capture structural features of the genome, in this case the close positional association of the GAL genes.
3.4 A case study: detecting a specific biological response in diverse data
At a more specific level, these interprocess associations and functional descriptions of datasets can be used to uncover detailed biological responses in high-throughput data. We were struck by the correlation in functional activities between three seemingly diverse datasets: Chitikila et al., (2002), an investigation of TBP inhibitors, Martin et al. (2004), an analysis of tor2 mutants described in Helliwell et al. (1998), and Pitkanen et al. (2004), a pmi40 deletion assayed over varying mannose concentrations. These three microarray collections share functional enrichments with other datasets assaying similar conditions [e.g. the nutritional cluster discussed above including Martin et al. (2004) and Pitkanen et al. (2004)], and no one pair of the three correlations is unusually high. They also represent two different experimental platforms: Martin et al. (2004) and Pitkanen et al. (2004) both employ single channel microarrays, while Chitikila et al. (2002) uses a two-color array. However, the average functional correlation between the three datasets is highly significant 
, P<10–3) for arrays under such apparently diverse conditions.
All three datasets are enriched for activity in distinct biological processes, and all three present unique biological conclusions that are in no way undermined by this unexpected similarity. Upon inspection of the three datasets’ experimental protocols, however, the common factor appears to be the use of a specific plasmid shuffle transformation employing a strain background of the form ura3 trp1 leu2 his3 or his4. We have confirmed this similarity in a fourth dataset we are currently developing investigating temperature-sensitive dbf4 mutants (Myers et al., 2005). Although the overarching biological conditions of our dataset share little in common with Chitikila et al. (2002), Martin et al. (2004) and Pitkanen et al. (2004), our mutants were also constructed using a similar plasmid transformation, and the resulting microarrays produce highly correlated functional profiles. Even when strain background and reference channels (when applicable) are all properly controlled, the plasmid shuffle process and associated auxotrophies result in subtle changes in global transcription detectable by large-scale functional analysis.
This effect is quite subtle, a fact which we stress for two reasons. First, it is a secondary effect within the more prominent biological features assayed by these three datasets, and it is only by large-scale analysis of their functional content in the context of many other datasets that the similarity was discovered. Second, we emphasize that it in no way diminishes these datasets’ primary results, and instead provides additional functional insight into their coexpression measurements. Most previous computational data integration has focused on associating genes with functions or genes with genes. As more high-throughput data becomes available, it opens up opportunities for associating entire datasets with broad functional activity and with other datasets, allowing the detection of biological signals and similarities that would remain undetectable at smaller scales.
|
4 DISCUSSION |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
We present a high-level functional analysis of very large compendia of genomic data and apply it to S.cerevisiae. By computationally summarizing thousands of whole-genome experimental conditions, we elucidate the current data coverage of S.cerevisiae biological processes, the cohesiveness of its functional annotations, and associations among these processes based on high-throughput experimental results. We also determine the functional activity in high-throughput datasets, allowing us to discover subtle relationships such as shared strain backgrounds in otherwise diverse microarray conditions. This analysis begins with specific functional relationships between individual genes predicted from large-scale data integration, and it extends into high-level information including functional associations between datasets, uncharacterized genes and biological processes.
A primary application of this system lies in directing future experimental efforts. In particular, high-throughput screens of any sort can be costly to implement and assay fairly general conditions; for example, if two proteins bind only during fermentation, their interaction will not be observed in a genomic screen during respiratory growth. A high-level functional analysis serves to call out underrepresented biological processes and those with increased likelihoods of novel discovery, which can in turn provide focus for experimental screens. This is analogous to candidate gene selection at a whole-genome level, a form of ‘candidate process’ selection, just as our predicted associations between biological processes represent functional relationships at a larger scale.
High-level functional analysis also provides very specific information on individual experimental results, in addition to its larger scale applications. This is exemplified by the functional signature of the plasmid shuffle strain discussed above; given any new high-throughput dataset, microarray or otherwise, we provide a means for establishing its functional activity in the context of existing data. Both this post hoc analysis and the a priori predictions of underrepresented functions are of particular use in less well-studied organisms. By designing experiments to explore processes shown to lack functional coverage and by leveraging all available data to interpret new results, laboratory work can be quickly guided to areas of biological interest and potential.
Finally, the functional information summarized by our system can also be employed in the continuous process of functional cataloging. While we have used examples from the GO, any sets of functionally related genes could drive analyses such as this, and the results can guide annotators in cataloging existing data much as they can guide experimenters in generating new data. By providing a means of directing annotators to potentially under-annotated functions and the datasets associated with them, our analysis simplifies a curation and cataloging task that grows with each new publication. By analyzing and presenting the large-scale functional structure of genome-scale data, we hope to guide annotators and experimenters alike in exploring the potential of the ongoing genomic revolution.
|
ACKNOWLEDGEMENTS |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
The authors would like to thank Chad Myers, Matthew Hibbs, Florian Markowetz and David Hess for insightful comments and conversations and Camelia Chiriac for experimental assistance.
Funding: This research is partially supported by NSF CAREER award DBI-0546275, NIH grant R01 GM071966, NIH grant T32 HG003284 and NIGMS Center of Excellence grant P50 GM071508. O.G.T. is an Alfred P. Sloan Research Fellow.
Conflict of Interest: none declared.
|
REFERENCES |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
Alfarano C, et al. The Biomolecular Interaction Network Database and related tools 2005 update. Nucleic Acids Res (2005) 33:D418–D424.
Ashburner M, et al. Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat. Genet (2000) 25:25–29.[CrossRef][Web of Science][Medline]
Brauer MJ, et al. Coordination of growth rate, cell cycle, stress response, and metabolic activity in yeast. Mol. Biol. Cell (2008) 19:352–367.
Brauer MJ, et al. Homeostatic adjustment and metabolic remodeling in glucose-limited yeast cultures. Mol. Biol. Cell (2005) 16:2503–2517.
Brem RB, Kruglyak L. The landscape of genetic complexity across 5700 gene expression traits in yeast. Proc. Natl Acad. Sci. USA (2005) 102:1572–1577.
Brem RB, et al. Genetic dissection of transcriptional regulation in budding yeast. Science (2002) 296:752–755.
Bro C, et al. Transcriptional, proteomic, and metabolic responses to lithium in galactose-grown yeast cells. J. Biol. Chem (2003) 278:32141–32149.
Bulik DA, et al. Chitin synthesis in Saccharomyces cerevisiae in response to supplementation of growth medium with glucosamine and cell wall stress. Eukaryot. Cell (2003) 2:886–900.
Charikar M. Greedy approximation algorithms for finding dense components in a graph. In: Third International Workshop on Approximation Algorithms for Combinatorial Optimization (2000) Germany: Springer, Saarbrücken.
Chitikila C, et al. Interplay of TBP inhibitors in global transcriptional control. Mol. Cell (2002) 10:871–882.[CrossRef][Web of Science][Medline]
David FN. The moments of the Z and F distributions. Biometrika (1949) 36:394–403.
Davierwala AP, et al. The synthetic genetic interaction spectrum of essential genes. Nat. Genet (2005) 37:1147–1152.[CrossRef][Web of Science][Medline]
Druzdzel MJ. SMILE: Structural Modeling, Inference, and Learning Engine and GeNIe: a development environment for graphical decision-theoretic models. In: Sixteenth National Conference on Artificial Intelligence (1999) Orlando, FL: American Association for Artificial Intelligence.
Eisen MB, et al. Cluster analysis and display of genome-wide expression patterns. Proc. Natl Acad. Sci. USA (1998) 95:14863–14868.
Franke L, et al. Reconstruction of a functional human gene network, with an application for prioritizing positional candidate genes. Am. J. Hum. Genet (2006) 78:1011–1025.[CrossRef][Web of Science][Medline]
Gansner ER, North SC. An open graph visualization system and its applications to software engineering. Software Pract. Exper (2000) 30:1203–1233.[CrossRef]
Gasch AP, et al. Genomic expression programs in the response of yeast cells to environmental changes. Mol. Biol. Cell (2000) 11:4241–4257.
Gavin AC, et al. Functional organization of the yeast proteome by systematic analysis of protein complexes. Nature (2002) 415:141–147.[CrossRef][Medline]
Giaever G, et al. Functional profiling of the Saccharomyces cerevisiae genome. Nature (2002) 418:387–391.[CrossRef][Medline]
Harbison CT, et al. Transcriptional regulatory code of a eukaryotic genome. Nature (2004) 431:99–104.[CrossRef][Medline]
Helliwell SB, et al. TOR2 is part of two related signaling pathways coordinating cell growth in Saccharomyces cerevisiae. Genetics (1998) 148:99–112.
Hibbs MA, et al. Exploring the functional landscape of gene expression: directed search of large microarray compendia. Bioinformatics (2007) 23:2692–2699.
Ho Y, et al. Systematic identification of protein complexes in Saccharomyces cerevisiae by mass spectrometry. Nature (2002) 415:180–183.[CrossRef][Medline]
Hughes TR, et al. Functional discovery via a compendium of expression profiles. Cell (2000) 102:109–126.[CrossRef][Web of Science][Medline]
Huh WK, et al. Global analysis of protein localization in budding yeast. Nature (2003) 425:686–691.[CrossRef][Medline]
Huttenhower C, et al. A scalable method for integration and functional analysis of multiple microarray datasets. Bioinformatics (2006) 22:2890–2897.
Huttenhower C, Troyanskaya OG. Bayesian data integration: a functional perspective. Computational Syst. Bioinform. Life Sci. Soc (2006) 5:341–351.
Ideker T, et al. Integrated genomic and proteomic analyses of a systematically perturbed metabolic network. Science (2001) 292:929–934.
Jansen R, et al. A Bayesian networks approach for predicting protein-protein interactions from genomic data. Science (2003) 302:449–453.
Jelinsky SA, et al. Regulatory networks revealed by transcriptional profiling of damaged Saccharomyces cerevisiae cells: Rpn4 links base excision repair with proteasomes. Mol. Cell Biol (2000) 20:8157–8167.
Karaoz U, et al. Whole-genome annotation by using evidence integration in functional-linkage networks. Proc. Natl Acad. Sci. USA (2004) 101:2888–2893.
Kloster M, et al. Finding regulatory modules through large-scale gene-expression data analysis. Bioinformatics (2005) 21:1172–1179.
Krogan NJ, et al. Methylation of histone H3 by Set2 in Saccharomyces cerevisiae is linked to transcriptional elongation by RNA polymerase II. Mol. Cell Biol (2003) 23:4207–4218.
Krogan NJ, et al. High-definition macromolecular composition of yeast RNA-processing complexes. Mol. Cell (2004) 13:225–239.[CrossRef][Web of Science][Medline]
Lee I, et al. A probabilistic functional network of yeast genes. Science (2004) 306:1555–1558.
Martin DE, et al. Rank Difference Analysis of Microarrays (RDAM), a novel approach to statistical analysis of microarray expression profiling data. BMC Bioinformatics (2004) 5:148.[CrossRef][Medline]
Myers CL, et al. Finding function: evaluation methods for functional genomic data. BMC Genomics (2006) 7:187.[CrossRef][Medline]
Myers CL, et al. Discovery of biological networks from diverse functional genomic data. Genome Biol (2005) 6:R114.[CrossRef][Medline]
Myers CL, Troyanskaya OG. Context-sensitive data integration and prediction of biological networks. Bioinformatics (2007) 23:2322–2330.
Neapolitan RE. Learning Bayesian Networks (2004) Chicago, IL: Prentice Hall.
O'Rourke SM, Herskowitz I. Unique and redundant roles for HOG MAPK pathway components as revealed by whole-genome expression analysis. Mol. Biol. Cell (2004) 15:532–542.
Pitkanen JP, et al. Excess mannose limits the growth of phosphomannose isomerase PMI40 deletion strain of Saccharomyces cerevisiae. J. Biol. Chem (2004) 279:55737–55743.
Schawalder SB, et al. Growth-regulated recruitment of the essential yeast ribosomal protein gene activator Ifh1. Nature (2004) 432:1058–1061.[CrossRef][Medline]
Segal E, et al. Module networks: identifying regulatory modules and their condition-specific regulators from gene expression data. Nat. Genet (2003) 34:166–176.[CrossRef][Web of Science][Medline]
SGD. Saccharomyces Genome Database. (2006) Available at http://www.yeastgenome.org.
Spellman PT, et al. Comprehensive identification of cell cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarray hybridization. Mol. Biol. Cell (1998) 9:3273–3297.
Stark C, et al. BioGRID: a general repository for interaction datasets. Nucleic Acids Res (2006) 34:D535–D539.
Tong AH, et al. Global mapping of the yeast genetic interaction network. Science (2004) 303:808–813.
Troyanskaya OG, et al. A Bayesian framework for combining heterogeneous data sources for gene function prediction (in Saccharomyces cerevisiae). Proc. Natl Acad. Sci. USA (2003) 100:8348–8353.
Yvert G, et al. Trans-acting regulatory variation in Saccharomyces cerevisiae and the role of transcription factors. Nat. Genet (2003) 35:57–64.[Web of Science][Medline]
Zhao R, et al. Navigating the chaperone network: an integrative map of physical and genetic interactions mediated by the hsp90 chaperone. Cell (2005) 120:715–727.[CrossRef][Web of Science][Medline]
Zhu J, Zhang MQ. SCPD: a promoter database of the yeast Saccharomyces cerevisiae. Bioinformatics (1999) 15:607–611.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  C. Huttenhower, E. M. Haley, M. A. Hibbs, V. Dumeaux, D. R. Barrett, H. A. Coller, and O. G. Troyanskaya Exploring the human genome with functional maps Genome Res., June 1, 2009; 19(6): 1093 - 1106. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||