Should genetic groups be fitted in BLUP evaluation? Practical answer for the French AI beef sire evaluation

Some analytical and simulated criteria were used to determine whether a priori genetic differences among groups, which are not accounted for by the relationship matrix, ought to be fitted in models for genetic evaluation, depending on the data structure and the accuracy of the evaluation. These criteria were the mean square error of some extreme contrasts between animals, the true genetic superiority of animals selected across groups, i.e. the selection response, and the magnitude of selection bias (difference between true and predicted selection responses). The different statistical models studied considered either fixed or random genetic groups (based on six different years of birth) versus ignoring the genetic group effects in a sire model. Including fixed genetic groups led to an overestimation of selection response under BLUP selection across groups despite the unbiasedness of the estimation, i.e. despite the correct estimation of differences between genetic groups. This overestimation was extremely important in numerical applications which considered two kinds of within-station progeny test designs for French purebred beef cattle AI sire evaluation across years: the reference sire design and the repeater sire design. When assuming a priori genetic differences due to the existence of a genetic trend of around 20% of genetic standard deviation for a trait with h² = 0.4, in a repeater sire design, the overestimation of the genetic superiority of bulls selected across groups varied from about 10% for an across-year selection rate p = 1/6 and an accurate selection index (100 progeny records per sire) to 75% for p = 1/2 and a less accurate selection index (20 progeny records per sire). This overestimation decreased when the genetic trend, the heritability of the trait, the accuracy of the evaluation or the connectedness of the design increased. Whatever the data design, a model of genetic evaluation without groups was preferred to a model with genetic groups when the genetic trend was in the range of likely values in cattle breeding programs (0 to 20% of genetic standard deviation). In such a case, including random groups was pointless and including fixed groups led to a large overestimation of selection response, smaller true selection response across groups and larger variance of estimation of the differences between groups. Although the genetic trend was correctly predicted by a model fitting fixed genetic groups, important errors in predicting individual breeding values led to incorrect ranking of animals across groups and, consequently, led to lower selection response.     

Simulation analysis to test the influence of model adequacy and data structure on the estimation of genetic parameters for traits with direct and maternal effects

Simulations were used to study the influence of model adequacy and data structure on the estimation of genetic parameters for traits governed by direct and maternal effects. To test model adequacy, several data sets were simulated according to different underlying genetic assumptions and analysed by comparing the correct and incorrect models. Results showed that omission of one of the random effects leads to an incorrect decomposition of the other components. If maternal genetic effects exist but are neglected, direct heritability is overestimated, and sometimes more than double. The bias depends on the value of the genetic correlation between direct and maternal effects. To study the influence of data structure on the estimation of genetic parameters, several populations were simulated, with different degrees of known paternity and different levels of genetic connectedness between flocks. Results showed that the lack of connectedness affects estimates when flocks have different genetic means because no distinction can be made between genetic and environmental differences between flocks. In this case, direct and maternal heritabilities are under-estimated, whereas maternal environmental effects are overestimated. The insufficiency of pedigree leads to biased estimates of genetic parameters.     

基于高密度SNP标记估计群体间遗传关联

联合育种的准确性受到群体间遗传关联程度的影响。本研究通过比较基于系谱数据和基因组数据计算的群体遗传关联,探究高密度SNP标记在遗传关联估计中的应用前景。本研究同时使用了模拟数据和真实数据,采用6种不同的遗传关联计算方法,包括PEVD(prediction error variance of differences)、PEVD(x)、VED(variance of estimated difference)、CD(generalized coefficient of determination)、r(prediction error correlation)和CR(connectedness rating),比较基于构建不同的关系矩阵(A、G、G_s、G_0.5和H矩阵)的群体间遗传关联。模拟数据和实际数据结果表明,除PEVD(x)和VED方法外,PEVD、CD、r和CR基于基因组信息的G、G_s和G_0.5阵计算的遗传关联程度均高于基于系谱信息的A阵,基于同时利用系谱和基因组信息的H阵遗传关联结果一般介于A阵与G阵之间。当CR和r为0时,CD都较高,高估了群体遗传关联。用r度量3个遗传分化程度不同的猪场间遗传关联时,基于G阵的r值均为0.01,不能准确反映群体真实遗传关联。随着遗传力的提高,所有群体遗传关联评估方法都有所改善,但遗传力为0.1时,PEVD基于A阵结果优于G阵,中高遗传力性状用于估计遗传关联优于低遗传力性状。本研究证明高密度SNP标记比系谱信息估计群体间遗传关联更有优势,CR是衡量遗传关联稳健而可靠的评价指标,计算简单,受性状遗传力影响较小。PEVD可以作为补充,量化具体群体遗传关联下的育种值预测误差情况。G矩阵比G_s、G_0.5阵能更好反映群体遗传关联。     

Bias in the prediction of genetic gain due to mass and half-sib selection in random mating populations

The prediction of gains from selection allows the comparison of breeding methods and selection strategies, although these estimates may be biased. The objective of this study was to investigate the extent of such bias in predicting genetic gain. For this, we simulated 10 cycles of a hypothetical breeding program that involved seven traits, three population classes, three experimental conditions and two breeding methods (mass and half-sib selection). Each combination of trait, population, heritability, method and cycle was repeated 10 times. The predicted gains were biased, even when the genetic parameters were estimated without error. Gain from selection in both genders is twice the gain from selection in a single gender only in the absence of dominance. The use of genotypic variance or broad sense heritability in the predictions represented an additional source of bias. Predictions based on additive variance and narrow sense heritability were equivalent, as were predictions based on genotypic variance and broad sense heritability. The predictions based on mass and family selection were suitable for comparing selection strategies, whereas those based on selection within progenies showed the largest bias and lower association with the realized gain.     

Design-based mapping of tree attributes by 3P sampling

The estimation of individual values (marks) in a finite population of units (e.g., trees) scattered onto a survey region is considered under 3P sampling. For each unit, the mark is estimated by means of an inverse distance weighting interpolator. Conditions ensuring the design-based consistency of maps are considered under 3P sampling. A computationally simple mean squared error estimator is adopted. Because 3P sampling involves the prediction of marks for each unit in the population, prediction errors rather than marks can be interpolated. Then, marks are estimated by the predictions plus the interpolated errors. If predictions are good, prediction errors are more smoothed than raw marks so that the procedure is likely to better meet consistency requirements. The purpose of this paper is to provide theoretical and empirical evidence on the effectiveness of the interpolation based on prediction errors to prove that the proposed strategy is a tool of general validity for mapping forest stands.     

Extinction,recolonization, and the genetic structure of tidepool copepod populations

Summary Extinction and recolonization in metapopulations may either increase or decrease genetic differentiation among populations, but recent genetic models predict increased differentiation under most circumstances of recolonization. I examine this prediction empirically using tidepool populations of the marine tidepool copepodTigriopus californicus. The probability of extinction of tidepool populations was sufficiently high to invoke the model's predictions, but varied among populations. Nearly 75% of colonizing groups consisted of 10 or fewer individuals. The genetic effective size of colonizing groups might be as high as 18, depending on assumptions, but colonists probably originated from a subset of local populations. In contrast to my predictions, genetic differentiation was smaller among younger tidepool populations than among older populations on each of three rock outcrops, suggesting that genetic differentiation was reduced by metapopulation dynamics. The discrepancy between the prediction and the results may be explained by the unmet assumptions of classical metapopulation structure underlying the genetic models.     

Tackling intraspecific genetic structure in distribution models better reflects species geographical range

Genetic diversity provides insight into heterogeneous demographic and adaptive history across organisms’ distribution ranges. For this reason, decomposing single species into genetic units may represent a powerful tool to better understand biogeographical patterns as well as improve predictions of the effects of GCC (global climate change) on biodiversity loss. Using 279 georeferenced Iberian accessions, we used classes of three intraspecific genetic units of the annual plant Arabidopsis thaliana obtained from the genetic analyses of nuclear SNPs (single nucleotide polymorphisms), chloroplast SNPs, and the vernalization requirement for flowering. We used SDM (species distribution models), including climate, vegetation, and soil data, at the whole‐species and genetic‐unit levels. We compared model outputs for present environmental conditions and with a particularly severe GCC scenario. SDM accuracy was high for genetic units with smaller distribution ranges. Kernel density plots identified the environmental variables underpinning potential distribution ranges of genetic units. Combinations of environmental variables accounted for potential distribution ranges of genetic units, which shrank dramatically with GCC at almost all levels. Only two genetic clusters increased their potential distribution ranges with GCC. The application of SDM to intraspecific genetic units provides a detailed picture on the biogeographical patterns of distinct genetic groups based on different genetic criteria. Our approach also allowed us to pinpoint the genetic changes, in terms of genetic background and physiological requirements for flowering, that Iberian A. thaliana may experience with a GCC scenario applying SDM to intraspecific genetic units.     

Measuring connectedness among herds in mixed linear models: From theory to practice in large-sized genetic evaluations

A procedure to measure connectedness among groups in large-sized genetic evaluations is presented. It consists of two steps: (a) computing coefficients of determination (CD) of comparisons among groups of animals; and (b) building sets of connected groups. The CD of comparisons were estimated using a sampling-based method that estimates empirical variances of true and predicted breeding values from a simulated n-sample. A clustering method that may handle a large number of comparisons and build compact clusters of connected groups was developed. An aggregation criterion (Caco) that reflects the level of connectedness of each herd was computed. This procedure was validated using a small beef data set. It was applied to the French genetic evaluation of the beef breed with most records and to the genetic evaluation of goats. Caco was more related to the type of service of sires used in the herds than to herd size. It was very sensitive to the percentage of missing sires. Disconnected herds were reliably identified by low values of Caco. In France, this procedure is the reference method for evaluating connectedness among the herds involved in on-farm genetic evaluation of beef cattle (IBOVAL) since 2002 and for genetic evaluation of goats from 2007 onwards.     

Genetic robustness and selection at the protein level for synonymous codons

Synonymous codons are neutral at the protein level, therefore natural selection at the protein level should have no effect on their frequencies. Synonymous codons, however, differ in their capacity to reduce the effects of errors: after mutation, certain codons keep on coding for the same amino acid or for amino acids with similar properties, while other synonymous codons produce very different amino acids. Therefore, the impact of errors on a coding sequence (genetic robustness) can be measured by analysing its codon usage. I analyse the codon usage of sequenced nuclear and cytoplasmic genomes and I show that there is an extensive variation in genetic robustness at the DNA sequence level, both among genomes and among genes of the same genome. I also show theoretically that robustness can be adaptive, that is natural selection may lead to a preference for codons that reduce the impact of errors. If selection occurs only among the mutants of a codon (e.g. among the progeny before the adult phase), however, the codons that are more sensitive to the effects of mutations may increase in frequency because they manage to get rid more easily of deleterious mutations. I also suggest other possible explanations for the evolution of genetic robustness at the codon level.     

The evolution of male genitalia: patterns of genetic variation and covariation in the genital sclerites of the dung beetle Onthophagus taurus

Three main hypotheses, have been invoked to explain divergent genital evolution, the lock and key, pleiotropy, and sexual selection hypotheses, each of which make different predictions about how genital traits are inherited. Here we used a half-sib breeding design to examine the patterns of genetic variation and covariation between male genital sclerites, and their covariance with general body morphology in the dung beetle Onthophagus taurus. We found CV(A)'s and CV(P)'s were similar for both genital and general morphological traits and that CV(R)'s were large for both trait types. We found that male genital sclerites were negatively genetically correlated with general morphological traits. Variation in male genital morphology has direct implications for a male's fertilization success and the resulting sexual selection acting on male genitalia is predicted to maintain high levels of additive genetic variance. Contrary to this prediction, we found that individual genital sclerites all had low levels of additive genetic variance and large maternal and environmental sources of variation. Our data suggest that the genital sclerites in O. taurus are not inherited independently but as a genetically integrated unit. More importantly, the way the different sclerites function to influence male fertilization success reflects this genetic integration. Even though levels of V(A) in individual genital sclerites may be low, there may still be sufficient V(A) in multivariate trait space for selection to generate evolutionary change in the overall morphology of male genitalia.     

Benchmarking secondary structure prediction for fold recognition

If secondary structure predictions are to be incorporated into fold recognition methods, an assessment of the effect of specific types of errors in predicted secondary structures on the sensitivity of fold recognition should be carried out. Here, we present a systematic comparison of different secondary structure prediction methods by measuring frequencies of specific types of error. We carry out an evaluation of the effect of specific types of error on secondary structure element alignment (SSEA), a baseline fold recognition method. The results of this evaluation indicate that missing out whole helix or strand elements, or predicting the wrong type of element, is more detrimental than predicting the wrong lengths of elements or overpredicting helix or strand. We also suggest that SSEA scoring is an effective method for assessing accuracy of secondary structure prediction and perhaps may also provide a more appropriate assessment of the "usefulness" and quality of predicted secondary structure, if secondary structure alignments are to be used in fold recognition.     

Foraging in a landscape mosaic: selection for energy and minerals in free-ranging cattle

Several studies have indicated the potential importance of nutrients, other than energy, in determining foraging decisions. A model was developed to test this idea, on the assumption of an intake maximization for different nutrients (energy, sodium and phosphorus). The model predictions were tested using field data from cattle grazing in a landscape mosaic of Pleistocene cover-sand and riverine grassland. Observations on foraging behaviour, food intake and diet composition were collected in thirteen 4-day-periods over 2 years. Habitat selection was determined by comparing the proportion of grazing time in different vegetation units with the available area proportion of the units. Two levels of habitat selection were examined: a micro-level (fine-scale, where vegetation units were considered separately) and a macro-level (coarse-scale, where vegetation units were combined to give selection at the landscape level). At the micro-level of habitat selection, no selection was apparent between the vegetation units of the riverine landscape, but the Deschampsia flexuosa unit was significantly selected for in the cover-sand landscape. At the landscape (macro-) level, the animals preferred the riverine landscape. The model revealed poor predictions of habitat occupancy on a micro-level. A much better prediction was obtained when vegetation units were combined at a macro-level. The D. flexuosa unit provided a higher energy intake, whereas the intake of sodium was higher in riverine grassland. Phosphorus proved relatively significant in determining habitat occupancy. Based on energy maximization alone, the model was a very poor predictor of habitat occupancy. It is argued that selection occurred mainly at the macro-level. The incorporation of different nutrient constraints in foraging models can then prove fruitful when seeking explanations of habitat occupancy. At the micro-level, difficulties for the animal in assessing nutrient availability may result in a less selective foraging pattern. However, the costs of increased selectivity may be greater than the benefits.     

Discovery of hidden pedigree errors combining genomic information with the genomic relationship matrix in Texel sheep

Genomic variants such as Single Nucleotide Polymorphisms and animal pedigree are now used widely in routine genetic evaluations of livestock in many countries. The use of genomic information not only can be used to enhance the accuracy of prediction but also to verify pedigrees for animals that are extensively managed using natural mating and enabling multiple-sire mating groups to be used. By so doing, the rate of genetic gain is enhanced, and any bias associated with incorrect pedigrees is removed. This study used a set of 8 764 sheep genotypes to verify the pedigree based on both the conventional opposing homozygote method as well as a novel method when combined with the inclusion of the genomic relationship matrix (GRM). The genomic relationship coefficients between verified pairs of animals showed on average a relationship of 0.50 with parent, 0.25 with grandparent, 0.13 with great grandparent, 0.50 with full-sibling and 0.27 with half-sibling. Minimum obtained values from these verified pairs were then used as thresholds to determine the pedigree for unverified pairs of animals, to detect potential errors in the pedigree. Using a case study from a population partially genotyped UK sheep, the results from this study illustrate a powerful way to resolve parentage inconsistencies, when combining the conventional ‘opposing homozygote’ method using genomic information together with GRM for pedigree checking. In this way, previously undetected pedigree errors can be resolved.     

Predicting evolutionary responses to selection on polyandry in the wild: additive genetic covariances with female extra-pair reproduction

The evolutionary forces that underlie polyandry, including extra-pair reproduction (EPR) by socially monogamous females, remain unclear. Selection on EPR and resulting evolution have rarely been explicitly estimated or predicted in wild populations, and evolutionary predictions are vulnerable to bias due to environmental covariances and correlated selection through unmeasured traits. However, evolutionary responses to (correlated) selection on any trait can be directly predicted as additive genetic covariances (cov_A) with appropriate components of relative fitness. I used comprehensive life-history, paternity and pedigree data from song sparrows (Melospiza melodia) to estimate cov_A between a female''s liability to produce extra-pair offspring and two specific fitness components: relative annual reproductive success (ARS) and survival to recruitment. All three traits showed non-zero additive genetic variance. Estimates of cov_A were positive, predicting evolution towards increased EPR, but 95% credible intervals overlapped zero. There was therefore no conclusive prediction of evolutionary change in EPR due to (correlated) selection through female ARS or recruitment. Negative environmental covariance between EPR and ARS would have impeded evolutionary prediction from phenotypic selection differentials. These analyses demonstrate an explicit quantitative genetic approach to predicting evolutionary responses to components of (correlated) selection on EPR that should be unbiased by environmental covariances and unmeasured traits.     

Genetic differences based on a beef terminal index are reflected in future phenotypic performance differences in commercial beef cattle

The increased demand for animal-derived protein and energy for human consumption will have to be achieved through a combination of improved animal genetic merit and better management strategies. The objective of the present study was to quantify whether differences in genetic merit among animals materialised into phenotypic differences in commercial herds. Carcass phenotypes on 156 864 animals from 7301 finishing herds were used, which included carcass weight (kg), carcass conformation score (scale 1 to 15), carcass fat score (scale 1 to 15) at slaughter as well as carcass price. The price per kilogram and the total carcass value that the producer received for the animal at slaughter was also used. A terminal index, calculated in the national genetic evaluations, was obtained for each animal. The index was based on pedigree index for calving performance, feed intake and carcass traits from the national genetic evaluations. Animals were categorised into four terminal index groups on the basis of genetic merit estimates that were derived before the expression of the phenotypic information by the validation animals. The association between terminal index and phenotypic performance at slaughter was undertaken using mixed models; whether the association differed by gender (i.e. young bulls, steers and heifers) or by early life experiences (animals born in a dairy herd or beef herd) was also investigated. The regression coefficient of phenotypic carcass weight, carcass conformation and carcass fat on their respective estimated breeding values (EBVs) was 0.92 kg, 1.08 units and 0.79 units, respectively, which is close to the expectation of one. Relative to animals in the lowest genetic merit group, animals in the highest genetic merit group had, on average, a 38.7 kg heavier carcass, with 2.21 units greater carcass conformation, and 0.82 units less fat. The superior genetic merit animals were, on average, slaughtered 6 days younger than their inferior genetic merit contemporaries. The superior carcass characteristics of the genetically elite animals materialised in carcasses worth €187 more than those of the lowest genetic merit animals. Although the phenotypic difference in carcass traits of animals divergent in terminal index differed statistically by animal gender and early life experience, the detected interactions were generally biologically small. This study clearly indicates that selection on an appropriate terminal index will produce higher performing animals and this was consistent across all production systems investigated.     

The socio‐genetics of a complex society: female gelada relatedness patterns mirror association patterns in a multilevel society

Multilevel societies with fission–fusion dynamics—arguably the most complex animal societies—are defined by two or more nested levels of organization. The core of these societies are modular social units that regularly fission and fuse with one another. Despite convergent evolution in disparate taxa, we know strikingly little about how such societies form and how fitness benefits operate. Understanding the kinship structure of complex societies could inform us about the origins of the social structure as well as about the potential for individuals in these societies to accrue indirect fitness benefits. Here, we combined genetic and behavioural data on geladas (Theropithecus gelada), an Old World Monkey, to complete the most comprehensive socio‐genetic analysis of a multilevel society to date. In geladas, individuals in the core social ‘units’, associate at different frequencies to form ‘teams’, ‘bands’ and, the largest aggregations, ‘communities’. Units were composed of closely related females, and females remained with their close kin during permanent fissions of units. Interestingly, female–female relatedness also significantly predicted between‐unit, between‐team and between‐band association patterns, while male–male relatedness did not. Thus, it is likely that the socio‐genetic structure of gelada society results from females maintaining associations with their female relatives during successive unit fissions—possibly in an attempt to balance the direct and indirect fitness benefits of group living. Overall, the persistence of associations among related females across generations appears to drive the formation of higher levels of gelada society, suggesting that females seek kin for inclusive fitness benefits at multiple levels of gelada society.     

A model for the adaptive significance of partial reproductive synchrony within social units

Summary Synchronized breeding in social units of animals (like colonies and herds) is often interpreted as a strategy against predators (predator satiation), and one might expect to find little variation in the relative date of starting reproduction. However, data on many colonies or herds show only partial synchrony. Our hypothesis is that this is not simply unavoidable variance but may be an adaption against predators. We consider the case where the members of a social unit can avoid predation actively by cooperation. If the contribution to this predator avoidance is different for individuals engaged in different phases of reproduction, our model shows that partial synchrony is of adaptive significance. Data on the black-headed gull (Larus ridibundus) are used to test the model's predictions.     

Genetic variation and population structure in a threatened species,the Utah prairie dog Cynomys parvidens: the use of genetic data to inform conservation actions

The Utah prairie dog (Cynomys parvidens), listed as threatened under the United States Endangered Species Act, was the subject of an extensive eradication program throughout its range during the 20th century. Eradication campaigns, habitat destruction/fragmentation/conversion, and epizootic outbreaks (e.g., sylvatic plague) have reduced prairie dog numbers from an estimated 95,000 individuals in the 1920s to approximately 14,000 (estimated adult spring count) today. As a result of these anthropogenic actions, the species is now found in small isolated sets of subpopulations. We characterized the levels of genetic diversity and population genetic structure using 10 neutral nuclear microsatellite loci for twelve populations (native and transplanted) representative of the three management designated “recovery units,” found in three distinct biogeographic regions, sampled across the species' range. The results indicate (1) low levels of genetic diversity within colonies (H_e = 0.109–0.357; H_o = 0.106‐ 0.313), (2) high levels of genetic differentiation among colonies (global F_ST = 0.296), (3) very small genetic effective population sizes, and (4) evidence of genetic bottlenecks. The genetic data reveal additional subdivision such that colonies within recovery units do not form single genotype clusters consistent with recovery unit boundaries. Genotype cluster membership support historical gene flow among colonies in the easternmost West Desert Recovery Unit with the westernmost Pausaugunt colonies and among the eastern Pausaugunt colonies and the Awapa Recovery unit to the north. In order to maintain the long‐term viability of the species, there needs to be an increased focus on maintaining suitable habitat between groups of existing populations that can act as connective corridors. The location of future translocation sites should be located in areas that will maximize connectivity, leading to maintenance of genetic variation and evolutionary potential.     

The prediction of C and N content and their potential mineralisation in heterogeneous soil samples using Vis–NIR spectroscopy and comparative methods

The development of a rapid, accurate and cost-effective method for the prediction of constituents related to soil nitrogen (N) supply is considered important. The potential of using visible (Vis) and near infrared reflectance (NIR) spectroscopy (400–2500 nm) as such a method was investigated. Vis–NIR calibrations were performed for organic carbon (C_org) and total N (N_tot) content and their potential mineralisation using 80 grassland soil samples of rather heterogeneous origin. Prediction accuracy was tested using a 'take-out-four' validation strategy (48 samples). Within investigated variables a ratio of standard deviation of reference data to standard error of bias corrected prediction (RPD) within 1.7 (r²=0.65) and 2.7 (r²=0.87) were achieved. Apparent differences in Vis–NIR prediction accuracy among the variables were partly due to errors in the reference values. Thawed moist samples tend to be more accurately predicted than dried samples, and no benefit was derived from the grinding of sieved (4 mm) and dried samples. Prediction accuracy did not differ using two different systems for sample presentation to the Vis–NIR analyses. Comparative predictions of C_org and N_tot and their potential mineralisations were performed using the take-out-four validation strategy and simple linear regression to loss on ignition (LOI) values and hot KCl extracted NH₄-N (N_hotKCl) values as predictors. Likewise, the reference values of C_org and N_tot were also used as predictors for each other and for the potential C and N mineralisation constituents. Accuracy obtained for the Vis–NIR predictions of investigated constituents was in general equal or better than prediction accuracy obtained by these comparative methods. The Vis–NIR method provided promising predictions of variables important for the soil N supply.     

Ensemble Positive Unlabeled Learning for Disease Gene Identification

An increasing number of genes have been experimentally confirmed in recent years as causative genes to various human diseases. The newly available knowledge can be exploited by machine learning methods to discover additional unknown genes that are likely to be associated with diseases. In particular, positive unlabeled learning (PU learning) methods, which require only a positive training set P (confirmed disease genes) and an unlabeled set U (the unknown candidate genes) instead of a negative training set N, have been shown to be effective in uncovering new disease genes in the current scenario. Using only a single source of data for prediction can be susceptible to bias due to incompleteness and noise in the genomic data and a single machine learning predictor prone to bias caused by inherent limitations of individual methods. In this paper, we propose an effective PU learning framework that integrates multiple biological data sources and an ensemble of powerful machine learning classifiers for disease gene identification. Our proposed method integrates data from multiple biological sources for training PU learning classifiers. A novel ensemble-based PU learning method EPU is then used to integrate multiple PU learning classifiers to achieve accurate and robust disease gene predictions. Our evaluation experiments across six disease groups showed that EPU achieved significantly better results compared with various state-of-the-art prediction methods as well as ensemble learning classifiers. Through integrating multiple biological data sources for training and the outputs of an ensemble of PU learning classifiers for prediction, we are able to minimize the potential bias and errors in individual data sources and machine learning algorithms to achieve more accurate and robust disease gene predictions. In the future, our EPU method provides an effective framework to integrate the additional biological and computational resources for better disease gene predictions.