GeneMANIA: a real-time multiple association network integration algorithm for predicting gene function.
Mostafavi S, Ray D, Warde-Farley D, Grouios C, Morris Q. · 2008
Read GeneMANIA: a real-time multiple association network integration algorithm for predicting gene function. by Mostafavi S, Ray D, Warde-Farley D, Grouios C, Morris Q. (2008) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc2447538.txt.
Introduction
Mostafavi and colleagues built a gene function predictor fast enough to run live in 2008. Most genes still have no known job, and computer methods guess jobs by merging many data sets linking genes that work together. The most accurate methods took so long that their predictions sat in ageing static databases.
Their algorithm, GeneMANIA, blends several association networks with a fast method derived from ridge regression, a standard statistical fit. It then spreads known labels across a process-specific network. On a mouse benchmark and a new yeast benchmark it matched or beat the leading methods. It also shrugged off redundant or irrelevant input data. Each prediction took under ten seconds on average in their tests. A prototype web server let users query it in real time rather than downloading frozen results.
Catalog record
| CREATOR | Mostafavi S, Ray D, Warde-Farley D, Grouios C, Morris Q. |
|---|---|
| YEAR | 2008 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 62 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
GeneMANIA: a real-time multiple association network integration algorithm for predicting gene function By Sara Mostafavi, Debajyoti Ray, David Warde-Farley, Chris Grouios, Quaid Morris Published in: Genome Biology DOI: 10.1186/gb-2008-9-s1-s4 ABSTRACT Abstract Background: Most successful computational approaches for protein function prediction integrate multiple genomics and proteomics data sources to make inferences about the function of unknown proteins. The most accurate of these algorithms have long running times, making them unsuitable for real-time protein function prediction in large genomes. As a result, the predictions of these algorithms are stored in static databases that can easily become outdated. We propose a new algorithm, GeneMANIA, that is as accurate as the leading methods, while capable of predicting protein function in real-time. Results: We use a fast heuristic algorithm, derived from ridge regression, to integrate multiple functional association networks and predict gene function from a single process-specific network using label propagation. Our algorithm is efficient enough to be deployed on a modern webserver and is as accurate as, or more so than, the leading methods on the MouseFunc I benchmark and a new yeast function prediction benchmark; it is robust to redundant and irrelevant data and requires, on average, less than ten seconds of computation time on tasks from these benchmarks. Conclusion: GeneMANIA is fast enough to predict gene function on-the-fly while achieving state-of-the-art accuracy. A prototype version of a GeneMANIA-based webserver is available at http://morrislab.med.utoronto.ca/prototype. INTRODUCTION The vast amount and increasing variety of genomic and proteomic data generated for model organisms creates an opportunity for in silico prediction of gene function through extrapolation of the functional properties of known genes. Genes with similar patterns of expression [1], synthetic lethality [2], or chemical sensitivity [3] often have similar functions. Additionally, function tends to be shared among genes whose gene products interact physically [4], are part of the same complex [5], or have similar three-dimensional structures [6]. Computational analyses have also revealed shared function among genes with similar phylogenetic profiles [7] or with shared protein domains [8]. More accurate predictions can be made by combining multiple heterogeneous sources of genomic and proteomic data [9]. Collectively, these observations have led to functional categorization of a number of previously uncharacterized genes using the so-called 'guilt-by-association' principle [10-12]. Algorithms that predict gene function using the guilt-by-association principle do so by extending a 'seed list' of genes known to have the given function by adding other genes highly associated with the seed list in one or more genomic and proteomic data sources. These algorithms typically compute a 'functional association network' to represent each dataset; in this network the nodes correspond to genes or proteins and the undirected links (or edges) are weighted according to evidence of co-functionality implied by the data source. Types of functional association networks include kernels used by support vector machines (SVMs) [9], functional linkage networks [13], and protein-protein linkage maps [14]. Individual functional association networks are often combined to generate a composite functional association network that summarizes all of the evidence of co-functionality. This network is then used as input to an algorithm that scores each gene based on its proximity to the genes in the seed list. When employed on multiple complementary data sources, these algorithms can accurately predict previously annotated gene functions in blind tests [15], suggesting that their predictions for unannotated genes are also quite accurate. Despite these successes, guilt-by-association algorithms have yet to achieve widespread use in gene annotation or as sources of new hypotheses about gene function; to do so, their predictions need to become more accessible, more accurate, and more regularly updated. In principle, all available data should be used when generating hypotheses about gene function; however, compiling a large number of heterogeneous data sources, generating functional association networks to represent these sources, and then mapping gene identifiers among the networks is a complex and onerous task that is best handled by specialists. Centrally managed web-based 'prediction servers' are an efficient strategy to ensure that casual users have access to the best available predictions. However, maintaining accurate and up-to-date prediction servers can be computationally prohibitive. Though a large number of algorithms have been developed to predict the function of unannotated genes by combining heterogeneous data sources (see [16] for a recent review), the most accurate of these algorithms have long running times, which can range from minutes [17] to hours [9] on yeast. Larger mammalian genomes increase the run time of these algorithms even more. As such, these algorithms cannot feasibly be run online and instead their predictions are made offline based on sets of pre-defined seed lists derived, for example, from Gene Ontology (GO) annotations [18]. However, because new data and annotations are being generated at a rapid rate, maintaining an up-to-date database of the best available predictions for all possible functions requires substantial and potentially unavailable computational resources. Due to this limitation, most prediction servers sacrifice accuracy for speed by relying on a single, or a small number of, pre-computed composite functional association networks and using simple heuristics to score genes based on a given seed list (for example, see [13,14,19]). While the scoring heuristics are fast enough to provide online predictions for arbitrary seed lists, we will show that their predictions are much less accurate than more advanced methods. Furthermore, by using a single pre-computed network, these servers do not take advantage of the fact that different data sources are more relevant for different categories of gene function [2,9] and are not extensible to new or user-supplied data sources. Here we demonstrate that it is not necessary to surrender either accuracy nor flexibility when building a prediction server by showing that GeneMANIA (Multiple Association Network Integration Algorithm) can, in seconds, generate genome-wide predictions that achieve state-of-the-art accuracy on arbitrary seed gene lists without relying on a pre-specified association network. We have achieved this goal through a series of algorithmic and technical advances that we have encapsulated in a new software package. With GeneMANIA, it is no longer necessary to maintain lists of in silico predictions of gene function because they can be recomputed as needed. BACKGROUND Automated methods for predicting gene and protein function Many algorithms for predicting gene function use a functional association network as a common representation of evidence of co-functionality for heterogeneous datasets. In this representation, nodes represent genes or proteins and the undirected edges (or links) between these nodes are weighted according to the evidence for co-functionality of the genes implied by the particular datasets. There are various techniques for deriving these weights depending on the source of data and it is still unclear how best to extract evidence of co-functionality (see, for example, [20,21]); as such we here use a single method to calculate these weights (see Materials and methods for details). Here, we assume that all of the association weights are positive. Predicting gene function from functional association networks Following Lanckriet and coworkers [9], we pose the problem of predicting gene function as a binary classification problem and attempt to solve the problem of integrating multiple heterogeneous input data sources by assigning each functional association network derived from these data sources a positive weight that reflects its usefulness in predicting a given function of interest. Once these weights are calculated, we