Background

Machine learning approaches for classification learn the pattern of the feature space of different classes, or learn a boundary that separates the feature space into different classes. The features of the data instances are usually available, and it is only the class-labels of the instances that are unavailable. For example, to classify text documents into different topic categories, the words in the documents are features and they are readily available, whereas the topic is what is predicted. However, in some domains obtaining features may be resource-intensive because of which not all features may be available. An example is that of protein-protein interaction prediction, where not only are the labels ('interacting' or 'non-interacting') unavailable, but so are some of the features. It may be possible to obtain at least some of the missing features by carrying out a few experiments as permitted by the available resources. If only a few experiments can be carried out to acquire missing features, which proteins should be studied and which features of those proteins should be determined? From the perspective of machine learning for PPI prediction, it would be desirable that those features be acquired which when used in training the classifier, the accuracy of the classifier is improved the most. That is, the utility of the feature-acquisition is measured in terms of how much acquired features contribute to improving the accuracy of the classifier. Active feature acquisition (AFA) is a strategy to preselect such instance-feature combinations (i.e. protein and experiment combinations) for maximum utility. The goal of AFA is the creation of optimal training set that would result in the best classifier, and not in determining the best classification model itself.

Results

We present a heuristic method for active feature acquisition to calculate the utility of acquiring a missing feature. This heuristic takes into account the change in belief of the classification model induced by the acquisition of the feature under consideration. As compared to random selection of proteins on which the experiments are performed and the type of experiment that is performed, the heuristic method reduces the number of experiments to as few as 40%. Most notable characteristic of this method is that it does not require re-training of the classification model on every possible combination of instance, feature and feature-value tuples. For this reason, our method is far less computationally expensive as compared with previous AFA strategies.

Conclusions

The results show that our heuristic method for AFA creates an optimal training set with far less features acquired as compared to random acquisition. This shows the value of active feature acquisition to aid in protein-protein interaction prediction where feature acquisition is costly. Compared to previous methods, the proposed method reduces computational cost while also achieving a better F-score. The proposed method is valuable as it presents a direction to AFA with a far lesser computational expense by removing the need for the first time, of training a classifier for every combination of instance, feature and feature-value tuples which would be impractical for several domains.

Background

Macromolecular protein complexes play important roles in a cell and their tertiary structure can help understand key biological processes of their functions. Multiple protein docking is a valuable computational tool for providing structure information of multimeric protein complexes. In a previous study we developed and implemented an algorithm for this purpose, named Multi-LZerD. This method represents a conformation of a multimeric protein complex as a graph, where nodes denote subunits and each edge connecting nodes denotes a pairwise docking conformation of the two subunits. Multi-LZerD employs a genetic algorithm to sample different topologies of the graph and pairwise transformations between subunits, seeking for the conformation of the optimal (lowest) energy. In this study we explore different configurations of the genetic algorithm, namely, the population size, whether to include a crossover operation, as well as the threshold for structural clustering, to find the optimal experimental setup.

Methods

Multi-LZerD was executed to predict the structures of three multimeric protein complexes, using different population sizes, clustering thresholds, and configurations of mutation and crossover. We analyzed the impact of varying these parameters on the computational time and the prediction accuracy.

Results and conclusions

Given that computational resources is a key for handling complexes with a large number of subunits and also for computing a large number of protein complexes in a genome-scale study, finding a proper setting for sampling the conformation space is of the utmost importance. Our results show that an excessive sampling of the conformational space by increasing the population size or by introducing the crossover operation is not necessary for improving accuracy for predicting structures of small complexes. The clustering is effective in reducing redundant pairwise predictions, which leads to successful identification of near-native conformations.

Background

Decades of genome-wide association studies (GWAS) have accumulated large volumes of genomic data that can potentially be reused to increase statistical power of new studies, but different genotyping platforms with different marker sets have been used as biotechnology has evolved, preventing pooling and comparability of old and new data. For example, to pool together data collected by 550K chips with newer data collected by 900K chips, we will need to impute missing loci. Many imputation algorithms have been developed, but the posteriori probabilities estimated by those algorithms are not a reliable measure the quality of the imputation. Recently, many studies have used an imputation quality score (IQS) to measure the quality of imputation. The IQS requires to know true alleles to estimate. Only when the population and the imputation loci are identical can we reuse the estimated IQS when the true alleles are unknown.

Methods

Here, we present a regression model to estimate IQS that learns from imputation of loci with known alleles. We designed a small set of features, such as minor allele frequencies, distance to the nearest known cross-over hotspot, etc., for the prediction of IQS. We evaluated our regression models by estimating IQS of imputations by BEAGLE for a set of GWAS data from the NCBI GEO database collected from samples from different ethnic populations.

Results

We construct a ν-SVR based approach as our regression model. Our evaluation shows that this regression model can accomplish mean square errors of less than 0.02 and a correlation coefficient close to 0.75 in different imputation scenarios. We also show how the regression results can help remove false positives in association studies.

Conclusion

Reliable estimation of IQS will facilitate integration and reuse of existing genomic data for meta-analysis and secondary analysis. Experiments show that it is possible to use a small number of features to regress the IQS by learning from different training examples of imputation and IQS pairs.

Background

Advancements in function prediction algorithms are enabling large scale computational annotation for newly sequenced genomes. With the increase in the number of functionally well characterized proteins it has been observed that there are many proteins involved in more than one function. These proteins characterized as moonlighting proteins show varied functional behavior depending on the cell type, localization in the cell, oligomerization, multiple binding sites, etc. The functional diversity shown by moonlighting proteins may have significant impact on the traditional sequence based function prediction methods. Here we investigate how well diverse functions of moonlighting proteins can be predicted by some existing function prediction methods.

Results

We have analyzed the performances of three major sequence based function prediction methods, PSI-BLAST, the Protein Function Prediction (PFP), and the Extended Similarity Group (ESG) on predicting diverse functions of moonlighting proteins. In predicting discrete functions of a set of 19 experimentally identified moonlighting proteins, PFP showed overall highest recall among the three methods. Although ESG showed the highest precision, its recall was lower than PSI-BLAST. Recall by PSI-BLAST greatly improved when BLOSUM45 was used instead of BLOSUM62.

Conclusion

We have analyzed the performances of PFP, ESG, and PSI-BLAST in predicting the functional diversity of moonlighting proteins. PFP shows overall better performance in predicting diverse moonlighting functions as compared with PSI-BLAST and ESG. Recall by PSI-BLAST greatly improved when BLOSUM45 was used. This analysis indicates that considering weakly similar sequences in prediction enhances the performance of sequence based AFP methods in predicting functional diversity of moonlighting proteins. The current study will also motivate development of novel computational frameworks for automatic identification of such proteins.

Background

In recent years, many algorithms have been developed for network-based analysis of differential gene expression in complex diseases. These algorithms use protein-protein interaction (PPI) networks as an integrative framework and identify subnetworks that are coordinately dysregulated in the phenotype of interest.

Motivation

While such dysregulated subnetworks have demonstrated significant improvement over individual gene markers for classifying phenotype, the current state-of-the-art in dysregulated subnetwork discovery is almost exclusively limited to binary phenotype classes. However, many clinical applications require identification of molecular markers for multiple classes.

Approach

We consider the problem of discovering groups of genes whose expression signatures can discriminate multiple phenotype classes. We consider two alternate formulations of this problem (i) an all-vs-all approach that aims to discover subnetworks distinguishing all classes, (ii) a one-vs-all approach that aims to discover subnetworks distinguishing each class from the rest of the classes. For the one-vs-all formulation, we develop a set-cover based algorithm, which aims to identify groups of genes such that at least one gene in the group exhibits differential expression in the target class.

Results

We test the proposed algorithms in the context of predicting stages of colorectal cancer. Our results show that the set-cover based algorithm identifying "stage-specific" subnetworks outperforms the all-vs-all approaches in classification. We also investigate the merits of utilizing PPI networks in the search for multiple markers, and show that, with correct parameter settings, network-guided search improves performance. Furthermore, we show that assessing statistical significance when selecting features greatly improves classification performance.

Background

Our aim was to simulate the data for the QTLMAS2011 workshop following a pig-type family structure under an oligogenic model, each QTL being specific.

Results

The population comprised 3000 individuals issued from 20 sires and 200 dams. Within each family, 10 progenies belonged to the experimental population and were assigned phenotypes and marker genotypes and 5 belonged to the selection population, only known on their marker genotypes. A total of 10,000 SNPs carried by 5 chromosomes of 1 Morgan each were simulated. Eight QTL were created (1 quadri-allelic, 2 linked in phase, 2 linked in repulsion, 1 imprinted and 2 epistatic). Random noise was added giving an heritability of 0.30. The marker density, LD and MAF were similar to real life parameters.

Background

The QTLMAS XVth dataset consisted of the pedigrees, marker genotypes and quantitative trait performances of 2,000 phenotyped animals with a half-sib family structure. The trait was regulated by 8 QTL which display additive, imprinting or epistatic effects. This paper aims at comparing the QTL mapping results obtained by six participants of the workshop.

Methods

Different regression, GBLUP, LASSO and Bayesian methods were applied for QTL detection. The results of these methods are compared based on the number of correctly mapped QTL, the number of false positives, the accuracy of the QTL location and the estimation of the QTL effect.

Results

All the simulated QTL, except the interacting QTL on Chr5, were identified by the participants. Depending on the method, 3 to 7 out of the 8 QTL were identified. The distance to the real location and the accuracy of the QTL effect varied to a large extent depending on the methods and complexity of the simulated QTL.

Conclusions

While all methods were fairly efficient in detecting QTL with additive effects, it was clear that for non-additive situations, such as parent-of-origin effects or interactions, the BayesC method gave the best results by detecting 7 out of the 8 simulated QTL, with only two false positives and a good precision (less than 1 cM away on average). Indeed, if LASSO could detect QTL even in complex situations, it was associated with too many false positive results to allow for efficient GWAS. GENMIX, a method based on the phylogenies of local haplotypes, also appeared as a promising approach, which however showed a few more false positives when compared with the BayesC method.

Background

Genomic selection (GS) is emerging as an efficient and cost-effective method for estimating breeding values using molecular markers distributed over the entire genome. In essence, it involves estimating the simultaneous effects of all genes or chromosomal segments and combining the estimates to predict the total genomic breeding value (GEBV). Accurate prediction of GEBVs is a central and recurring challenge in plant and animal breeding. The existence of a bewildering array of approaches for predicting breeding values using markers underscores the importance of identifying approaches able to efficiently and accurately predict breeding values. Here, we comparatively evaluate the predictive performance of six regularized linear regression methods-- ridge regression, ridge regression BLUP, lasso, adaptive lasso, elastic net and adaptive elastic net-- for predicting GEBV using dense SNP markers.

Methods

We predicted GEBVs for a quantitative trait using a dataset on 3000 progenies of 20 sires and 200 dams and an accompanying genome consisting of five chromosomes with 9990 biallelic SNP-marker loci simulated for the QTL-MAS 2011 workshop. We applied all the six methods that use penalty-based (regularization) shrinkage to handle datasets with far more predictors than observations. The lasso, elastic net and their adaptive extensions further possess the desirable property that they simultaneously select relevant predictive markers and optimally estimate their effects. The regression models were trained with a subset of 2000 phenotyped and genotyped individuals and used to predict GEBVs for the remaining 1000 progenies without phenotypes. Predictive accuracy was assessed using the root mean squared error, the Pearson correlation between predicted GEBVs and (1) the true genomic value (TGV), (2) the true breeding value (TBV) and (3) the simulated phenotypic values based on fivefold cross-validation (CV).

Results

The elastic net, lasso, adaptive lasso and the adaptive elastic net all had similar accuracies but outperformed ridge regression and ridge regression BLUP in terms of the Pearson correlation between predicted GEBVs and the true genomic value as well as the root mean squared error. The performance of RR-BLUP was also somewhat better than that of ridge regression. This pattern was replicated by the Pearson correlation between predicted GEBVs and the true breeding values (TBV) and the root mean squared error calculated with respect to TBV, except that accuracy was lower for all models, most especially for the adaptive elastic net. The correlation between the predicted GEBV and simulated phenotypic values based on the fivefold CV also revealed a similar pattern except that the adaptive elastic net had lower accuracy than both the ridge regression methods.

Conclusions

All the six models had relatively high prediction accuracies for the simulated data set. Accuracy was higher for the lasso type methods than for ridge regression and ridge regression BLUP.

Background

The aim of this study was to estimate haplotype effects and then to predict breeding values using linear models. The haplotype based analysis enables avoidance of loosing information due to linkage disequilibrium between single markers. There are also less explanatory variables in the linear model which makes the estimation more reliable.

Methods

Different methods and criteria for marker and haplotype selection were considered. First, markers with MAF lower than 5% where excluded from the data set. Then, SNPs in complete linkage disequilibrium where selected. Next step was to construct haplotypes and to estimate their frequencies basing on selected SNPs. The haplotypes with a frequency lower than 1% were not considered in further analysis. Chosen haplotypes were used as the explanatory variables in the linear models for breeding values prediction. Linear models with fixed and random haplotype effects as well as animal model were tested.

Results

The number of markers was limited to 1206, 1189, 1249, 1288 and 1167 for chromosome 1, 2, 3, 4 and 5, respectively due to MAF criterion. In total 409 subsets of SNPs with r2=1 were found. 1476 haplotypes with different lengths were inferred. The frequencies of 817 haplotypes were higher than 1% - 184 for the first chromosome, 172 for the second, 131 for the third, 146 for the forth and 184 haplotypes for the fifth chromosome. The haplotype effects estimated using random models were comparable and more precise in prediction for individuals with unknown phenotypes. A few haplotypes with large effects were found when their effects were defined as fixed in the linear model . The correlations of the predicted breeding values with true breeding values were not that high. This could be brought about by selection criteria imposed on the genotype data which led to substantial reduction of number of markers.

Conclusions

Although not many markers were considered in the study, the results obtained show that the implemented approach can be considered as quite promising. The haplotype approach let to avoid high dimensional models as compared with single SNPs models.

Background

In genomic models that assign an individual variance to each marker, the contribution of one marker to the posterior distribution of the marker variance is only one degree of freedom (df), which introduces many variance parameters with only little information per variance parameter. A better alternative could be to form clusters of markers with similar effects where markers in a cluster have a common variance. Therefore, the influence of each marker group of size p on the posterior distribution of the marker variances will be p df.

Methods

The simulated data from the 15th QTL-MAS workshop were analyzed such that SNP markers were ranked based on their effects and markers with similar estimated effects were grouped together. In step 1, all markers with minor allele frequency more than 0.01 were included in a SNP-BLUP prediction model. In step 2, markers were ranked based on their estimated variance on the trait in step 1 and each 150 markers were assigned to one group with a common variance. In further analyses, subsets of 1500 and 450 markers with largest effects in step 2 were kept in the prediction model.

Results

Grouping markers outperformed SNP-BLUP model in terms of accuracy of predicted breeding values. However, the accuracies of predicted breeding values were lower than Bayesian methods with marker specific variances.

Conclusions

Grouping markers is less flexible than allowing each marker to have a specific marker variance but, by grouping, the power to estimate marker variances increases. A prior knowledge of the genetic architecture of the trait is necessary for clustering markers and appropriate prior parameterization.

Background

Genomic breeding value estimation is the key step in genomic selection. Among many approaches, BLUP methods and Bayesian methods are most commonly used for estimating genomic breeding values. Here, we applied two BLUP methods, TABLUP and GBLUP, and three Bayesian methods, BayesA, BayesB and BayesCπ, to the common dataset provided by the 15th QTL-MAS Workshop to evaluate and compare their predictive performances.

Results

For the 1000 progenies without phenotypic values, the correlations between GEBVs by different methods ranged from 0.812 (GBLUP and BayesCπ) to 0.997 (TABLUP and BayesB). The accuracies of GEBVs (measured as correlations between true breeding values (TBVs) and GEBVs) were from 0.774 (GBLUP) to 0.938 (BayesCπ) and the biases of GEBVs (measure as regressions of TBVs on GEBVs) were from 1.033 (TABLUP) to 1.648 (GBLUP). The three Bayesian methods and TABLUP had similar accuracy and bias.

Conclusions

BayesA, BayesB, BayesCπ and TABLUP performed similarly and satisfactorily and remarkably outperformed GBLUP for genomic breeding value estimation in this dataset. TABLUP is a promising method for genomic breeding value estimation because of its easy computation of reliabilities of GEBVs and its easy extension to real life conditions such as multiple traits and consideration of individuals without genotypes.

Background

The QTLMAS XVth dataset consisted of pedigree, marker genotypes and quantitative trait performances of animals with a sib family structure. Pedigree and genotypes concerned 3,000 progenies among those 2,000 were phenotyped. The trait was regulated by 8 QTLs which displayed additive, imprinting or epistatic effects. The 1,000 unphenotyped progenies were considered as candidates to selection and their Genomic Estimated Breeding Values (GEBV) were evaluated by participants of the XVth QTLMAS workshop. This paper aims at comparing the GEBV estimation results obtained by seven participants to the workshop.

Methods

From the known QTL genotypes of each candidate, two "true" genomic values (TV) were estimated by organizers: the genotypic value of the candidate (TGV) and the expectation of its progeny genotypic values (TBV). GEBV were computed by the participants following different statistical methods: random linear models (including BLUP and Ridge Regression), selection variable techniques (LASSO, Elastic Net) and Bayesian methods. Accuracy was evaluated by the correlation between TV (TGV or TBV) and GEBV presented by participants. Rank correlation of the best 10% of individuals and error in predictions were also evaluated. Bias was tested by regression of TV on GEBV.

Results

Large differences between methods were found for all criteria and type of genetic values (TGV, TBV). In general, the criteria ranked consistently methods belonging to the same family.

Conclusions

Bayesian methods - A