Search

Integrative systems biology approaches build, evaluate, and combine data from thousands of diverse experiments. These strategies rely on methods that effectively identify and summarize gene-gene relationships within individual experiments. For gene-expression datasets, the Pearson correlation is often applied to build coexpression networks because it is both easily interpretable and quick to calculate. Here we develop and evaluate weighted Pearson correlation approaches that better summarize gene expression data into coexpression networks for synchronized cell cycle time-course experiments. These methods use experimental measurements of cell cycle synchrony to estimate appropriate weights through either sliding window or linear regression approaches. We show that these weights improve our ability to build coexpression networks capable of identifying phase-specific functional relationships between genes. We evaluate our method on diverse experiments and find that both weighted strategies outperform the traditional method. This weighted correlation approach is implemented in the Sleipnir library, an open source library used for integrative systems biology. Integrative approaches using properly weighted time-course experiments will provide a more detailed understanding of the processes studied in such experiments.

In human genetics a primary goal is the discovery of genetic factors that predict individual susceptibility to common human diseases, but this has proven difficult to achieve because these diseases are likely to result from the joint failure of two or more interacting components. Currently geneticists measure genetic variations from across the genomes of individuals with and without the disease. The association of single variants with disease is then assessed. Our goal is to develop methods capable of identifying combinations of genetic variations predictive of discrete measures of health in human population data. "Artificial evolution" approaches loosely based on real biological processes have been developed and applied, but it has recently been suggested that "computational evolution" approaches will be more likely to solve problems of interest to biomedical researchers. Here we introduce a method to evolve parsimonious solutions in an open-ended computational evolution framework that more closely mimics the complexity of biological systems. In ecological systems a highly specialized organism can fail to thrive as the environment changes. By introducing numerous small changes into training data, i.e. the environment, during evolution we drive evolution towards general solutions. We show that this method leads to smaller solutions and does not reduce the power of an open-ended computational evolution system. This method of environmental perturbation fits within the computational evolution framework and is an effective method of evolving parsimonious solutions.

High-throughput genotyping has made genome-wide data on human genetic variation commonly available, however, finding associations between specific variations and common diseases has proven difficult. Individual susceptibility to common diseases likely depends on gene-gene interactions, i.e. epistasis, and not merely on independent genes. Furthermore, genome-wide datasets present an informatic challenge because exhaustive searching within them for even pair-wise interactions is computationally infeasible. Instead, search methods must be used which efficiently and effectively mine these datasets. To meet these challenges, we turn to a biologically inspired ant colony optimization strategy. We have previously developed an ant system which allows the incorporation of expert knowledge as heuristic information. One method of scaling expert knowledge to probabilities usable in the algorithm, an exponential distribution function which respects intervals between raw expert knowledge scores, has been previously examined. Here, we develop and evaluate three additional expert knowledge scaling methods and find parameter sets for each which maximize power.

High throughput sequencing technologies now routinely measure over one million DNA sequence variations on the human genome. Analyses of these data have demonstrated that single sequence variants predictive of common human disease are rare. Instead, disease risk is thought to be the result of a confluence of many genes acting in concert, often with no statistically significant individual effects. The detection and characterization of such gene-gene interactions that predispose for human disease is a computationally daunting task, since the search space grows exponentially with the number of measured genetic variations. Traditional artificial evolution methods have offered some promise in this problem domain, but they are plagued by the lack of marginal effects of individual sequence variants. To address this problem, we have developed a computational evolution system that allows for the evolution of solutions and solution operators of arbitrary complexity. In this study, we incorporate a linkage learning technique into the population initialization method of the computational evolution system and investigate its influence on the ability to detect and characterize gene-gene interactions in synthetic data sets. These data sets are generated to exhibit characteristics of real genomewide association studies for purely epistatic diseases with various heritabilities. Our results demonstrate that incorporating linkage learning in population initialization via expert knowledge sources improves classification accuracy, enhancing our ability to automate the discovery and characterization of the genetic causes of common human diseases.

Modern genotyping techniques have allowed the field of human genetics to generate vast amounts of data, but analysis methodologies have not been able to keep pace with this increase. In order to allow personal genomics to play a vital role in modern health care, analysis methods capable of discovering high order interactions that contribute to an individual’s risk of disease must be developed. An artificial immune system (AIS) is a method which maps well to this problem and has a number of appealing properties. By considering many attributes simultaneously, it may be able to effectively and efficiently detect epistasis, that is non-additive gene-gene interactions. This situation of interacting genes is currently very difficult to detect without biological insight or statistical heuristics. Even with these approaches, at low heritability (i.e. where there is only a small genetic signal), these approaches have trouble distinguishing genetic signal from noise. The AIS also has a compact solution representation which can be rapidly evaluated. Finally the AIS approach, by iteratively developing an antibody which ignores irrelevant genotypes, may be better able to differentiate signal from noise than machine learning approaches like ReliefF which struggle at small heritabilities. Here we develop a basic AIS and evaluate it on very low heritability datasets. We find that the basic AIS is not robust to parameter settings but that, at some parameter settings, it performs very effectively. We use the settings where the strategy succeeds to suggest a path towards a robust AIS for human genetics. Developing an AIS which succeeds across many parameter settings will be critical to prepare this method for widespread use.

A primary goal of human genetics is the discovery of genetic factors that influence individual susceptibility to common human diseases. This problem is difficult because common diseases are likely the result of joint failure of two or more interacting components instead of single component failures. Efficient algorithms that can detect interacting attributes are needed. The Relief family of machine learning algorithms, which use nearest neighbors to weight attributes, are a promising approach. Recently an improved Relief algorithm called Spatially Uniform ReliefF (SURF) has been developed that significantly increases the ability of these algorithms to detect interacting attributes. Here we introduce an algorithm called SURF* which uses distant instances along with the usual nearby ones to weight attributes. The weighting depends on whether the instances are are nearby or distant. We show this new algorithm significantly outperforms both ReliefF and SURF for genetic analysis in the presence of attribute interactions. We make SURF* freely available in the open source MDR software package. MDR is a cross-platform Java application which features a user friendly graphical interface.

A goal of human genetics is to discover genetic factors that influence individuals’ susceptibility to common diseases. Most common diseases are thought to result from the joint failure of two or more interacting components instead of single component failures. This greatly complicates both the task of selecting informative genetic variations and the task of modeling interactions between them. We and others have previously developed algorithms to detect and model the relationships between these genetic factors and disease. Previously these methods have been evaluated with datasets simulated according to pre-defined genetic models. Here we develop and evaluate a model free evolution strategy to generate datasets which display a complex relationship between individual genotype and disease susceptibility. We show that this model free approach is capable of generating a diverse array of datasets with distinct gene-disease relationships for an arbitrary interaction order and sample size. We specifically generate six-hundred pareto fronts; one for each independent run of our algorithm. In each run the predictiveness of single genetic variation and pairs of genetic variations have been minimized, while the predictiveness of third, fourth, or fifth order combinations is maximized. This method and the resulting datasets will allow the capabilities of novel methods to be tested without pre-specified genetic models. This could improve our ability to evaluate which methods will succeed on human genetics problems where the model is not known in advance. We further make freely available to the community the entire pareto-optimal front of datasets from each run so that novel methods may be rigorously evaluated. These 56,600 datasets are available from http://discovery.dartmouth.edu/model_free_data/.

The availability of chip-based technology has transformed human genetics and made routine the measurement of thousands of DNA sequence variations giving rise to an informatics challenge. This challenge is the identification of combinations of interacting DNA sequence variations predictive of common diseases. We have previously developed Multifactor Dimensionality Reduction (MDR), a method capable of detecting these interactions, but an exhaustive MDR analysis is exponential in time complexity and thus unsuitable for an interaction analysis of genome-wide datasets. Therefore we look to stochastic search approaches to find a suitable wrapper for the analysis of these data. We have previously shown that an ant colony optimization (ACO) framework can be successfully applied to human genetics when expert knowledge is included. We have integrated an ACO stochastic search wrapper into the open source MDR software package. In this wrapper we also introduce a scaling method based on an exponential distribution function with a single user-adjustable parameter. Here we obtain expert knowledge from Tuned ReliefF (TuRF), a method capable of detecting attribute interactions in the absence of main effects, and perform a power analysis at different parameter settings. We show that the expert knowledge distribution parameter, the retention factor, and the weighting of expert knowledge significantly affect the power of the method.

In human genetics it is now feasible to measure large numbers of DNA sequence variations across the human genome. Given current knowledge about biological networks and disease processes it seems likely that disease risk can best be modeled by interactions between biological components, which can be examined as interacting DNA sequence variations. The machine learning challenge is to effectively explore interactions in these datasets to identify combinations of variations which are predictive of common human diseases. Ant colony optimization (ACO) is a promising approach to this problem. The goal of this study is to examine the usefulness of ACO for problems in this domain and to develop a prototype of an expert knowledge guided probabilistic search wrapper. We show that an ACO approach is not successful in the absence of expert knowledge but is successful when expert knowledge is supplied through the pheromone updating rule.

Human genetics is undergoing a data explosion. Methods are available to measure DNA sequence variation throughout the human genome. Given current knowledge it seems likely that common human diseases are best predicted by interactions between biological components, which can be examined as interacting DNA sequence variations. The challenge is thus to examine these high-dimensional datasets to identify combinations of variations likely to predict common diseases. The goal of this paper was to develop and evaluate a genetic programming (GP) mutator suited to this task by exploiting expert knowledge in the form of Tuned ReliefF (TuRF) scores during mutation. We show that using expert knowledge guided mutation performs similarly to expert knowledge guided selection. This study demonstrates that in the context of an expert knowledge aware GP, mutation may be an appropriate component of the GP used to search for interacting predictors in this domain.