Time filter

Source Type

Stuttgart Mühlhausen, Germany

Pernici B.,Polytechnic of Milan | Salomie I.,Technical University of Cluj Napoca | Wesner S.,High Performance Computing Center
Lecture Notes in Business Information Processing

The vision of the recently started GAMES European Research project is a new generation of energy efficient IT Service Centres, designed taking into account both the characteristics of the applications running in the centre and context-aware adaptivity features that can be enabled both at the application level and within the IT and utility infrastructure. Adaptivity at the application level is based on the service-oriented paradigm, which allows a dynamic composition and recomposition of services to guarantee Quality of Service levels that have been established with the users. At the infrastructure level, adaptivity is being sought with the capacity of switching on and off dynamically the systems components, based on the state of the service centre. However, these two perspectives are usually considered separately, managing at different levels applications and infrastructure. In addition, while performance and cost are usually the main parameters being considered both during design and at run time, energy efficiency of the service centre is normally not an issue. However, given that the impact of service centres is becoming more and more important in the global energy consumption, and that energy resources, in particular in peak periods, are more and more constrained, an efficient use of energy in service centres has become an important goal. In the GAMES project, energy efficiency improvement goals are tackled based on exploiting adaptivity, on building a knowledge base for evaluating the impact of the applications on the service centre energy consumption, and exploiting the application characteristics for an improved use of resources. © 2011 Springer-Verlag Berlin Heidelberg. Source

Cong G.,IBM | Muzio P.,High Performance Computing Center
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

We study parallel connected components algorithms onGPUs in comparison with CPUs. Although straightforward implementation of PRAMalgorithms performs relatively better on GPUs than on CPUs, the GPU memory subsystem performance is poor due to non-coalesced random accesses.We argue that generic sort-based access coalescing is too costly on GPUs.We propose a new coalescing technique and a new meta algorithm to improve locality and performance. Our optimization achieves up to 2.7 times speedup over the straightforward implementation. Interestingly, our optimization also works well on CPUs.Comparing the best-performing algorithms on GPUs and CPUs, we find our new algorithm is the fastest on GPUs and the second fastest on CPUs, while the parallel Rem’s algorithm is the fastest on CPUs but does not perform well on GPUs due to path divergence. © Springer International Publishing Switzerland 2014. Source

Hammond K.,University of St. Andrews | Aldinucci M.,University of Turin | Brown C.,University of St. Andrews | Cesarini F.,Erlang Solutions Ltd. | And 6 more authors.
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

This paper describes the ParaPhrase project, a new 3-year targeted research project funded under EU Framework 7 Objective 3.4 (Computer Systems), starting in October 2011. ParaPhrase aims to follow a new approach to introducing parallelism using advanced refactoring techniques coupled with high-level parallel design patterns. The refactoring approach will use these design patterns to restructure programs defined as networks of software components into other forms that are more suited to parallel execution. The programmer will be aided by high-level cost information that will be integrated into the refactoring tools. The implementation of these patterns will then use a well-understood algorithmic skeleton approach to achieve good parallelism. A key ParaPhrase design goal is that parallel components are intended to match heterogeneous architectures, defined in terms of CPU/GPU combinations, for example. In order to achieve this, the ParaPhrase approach will map components at link time to the available hardware, and will then re-map them during program execution, taking account of multiple applications, changes in hardware resource availability, the desire to reduce communication costs etc. In this way, we aim to develop a new approach to programming that will be able to produce software that can adapt to dynamic changes in the system environment. Moreover, by using a strong component basis for parallelism, we can achieve potentially significant gains in terms of reducing sharing at a high level of abstraction, and so in reducing or even eliminating the costs that are usually associated with cache management, locking, and synchronisation. © 2013 Springer-Verlag Berlin Heidelberg. Source

News Article
Site: http://www.nature.com/nature/current_issue/

Anchor loci described in ref. 12 were extended such that each contained approximately 1,350 bp. In some cases neighbouring loci were joined to form a single locus. Also, loci that performed poorly in ref. 12 were removed from the locus set. This process produced 394 loci (referred to as the version 2 vertebrate loci). Genome coordinates corresponding to these regions in the Gallus gallus genome (galGal3, UCSC genome browser) were identified and sequences corresponding to this region were extracted (coordinates are available in the Zenodo archive (http://dx.doi.org/10.5281/zenodo.28343)). In order to improve the capture efficiency for passerines, we also obtained homologous sequences for Taeniopygia guttata. After aligning the Gallus and Taeniopygia sequences using MAFFT31, alignments were trimmed to produce the final probe region alignments (alignments available in the Zenodo archive), and probes were tiled at approximately 1.5 × tiling density (probe specification will be made available upon publication). Data were collected following the general methods of ref. 12 through the Center for Anchored Phylogenomics at Florida State University (http://www.anchoredphylogeny.com). Briefly, each genomic DNA sample was sonicated to a fragment size of ~150–350 bp using a Covaris E220 focused-ultrasonicator with Covaris microTUBES. Subsequently, library preparation and indexing were performed on a Beckman-Coulter Biomek FXp liquid-handling robot following a protocol modified from ref. 32. One important modification is a size-selection step after blunt-end repair using SPRIselect beads (Beckman-Coulter; 0.9 × ratio of bead to sample volume). Indexed samples were then pooled at equal quantities (typically 12–16 samples per pool), and enrichments were performed on each multi-sample pool using an Agilent Custom SureSelect kit (Agilent Technologies), designed as specified above. After enrichment, the 12 enrichment pools were pooled in groups of three in equal quantities for sequencing on four PE150 Illumina HiSeq2000 lanes (three enrichment pools per lane). Sequencing was performed in the Translational Science Laboratory in the College of Medicine at Florida State University. Paired-read merging (Merge.java). Typically, between 50% and 75% of sequenced library fragments had an insert size between 150 bp and 300 bp. As 150 bp paired-end sequencing was performed, this means that the majority of the paired reads overlap and thus should be merged before assembly. The overlapping reads were identified and merged following the methods of ref. 33. In short, for each degree of overlap for each read we computed the probability of obtaining the observed number of matches by chance, and selected degree of overlap that produced the lowest probability, with a P value less than 10−10 required to merge reads. When reads are merged, mismatches are reconciled using base-specific quality scores, which were combined to form the new quality scores for the merged read (see ref. 33 for details). Reads failing to meet the probability criterion were kept separate but still used in the assembly. The merging process produces three files: one containing merged reads and two containing the unmerged reads. Assembly (Assembler.java). The reads were assembled into contigs using an assembler that makes use of both a divergent reference assembly approach to map reads to the probe regions and a de novo assembly approach to extend the assembly into the flanks. The reference assembler uses a library of spaced 20-mers derived from the conserved sites of the alignments used during probe design. A preliminary match was called if at least 17 of 20 matches exist between a spaced kmer and the corresponding positions in a read. Reads obtaining a preliminary match were then compared to an appropriate reference sequence used for probe design to determine the maximum number of matches out of 100 consecutive bases (all possible gap-free alignments between the read and the reference ware considered). The read was considered mapped to the given locus if at least 55 matches were found. Once a read is mapped, an approximate alignment position was estimated using the position of the spaced 20-mer, and all 60-mers existing in the read were stored in a hash table used by the de novo assembler. The de novo assembler identifies exact matches between a read and one of the 60-mers found in the hash table. Simultaneously using the two levels of assembly described above, the three read files were traversed repeatedly until an entire pass through the reads produced no additional mapped reads. For each locus, mapped reads were then clustered into clusters using 60-mer pairs observed in the reads mapped to that locus. In short, a list of all 60-mers found in the mapped reads was compiled, and the 60-mers were clustered if found together in at least two reads. The 60-mer clusters were then used to separate the reads into clusters for contig estimation. Relative alignment positions of reads within each cluster were then refined in order to increase the agreement across the reads. Up to one gap was also inserted per read if needed to improve the alignment. Note that given sufficient coverage and an absence of contamination, each single-copy locus should produce a single assembly cluster. Low coverage (leading to a break in the assembly), contamination, and gene duplication, can all lead to an increased number of assembly clusters. A whole-genome duplication, for example, would increase the number of clusters to two per locus. Consensus bases were called from assembly clusters as follows. For each site an unambiguous base was called if the bases present were identical or if the polymorphism of that site could be explained as sequencing error, assuming a binomial probability model with the probability of error equal to 0.1 and alpha equal to 0.05. If the polymorphism could not be explained as sequencing error, the ambiguous base was called that corresponded to all of the observed bases at that site (for example, ‘R’ was used if ‘A’ and ‘G’ were observed). Called bases were soft-masked (made lowercase) for sites with coverage lower than five. A summary of the assembly results is presented in a spreadsheet in the electronic data archive (http://dx.doi.org/10.5281/zenodo.28343; Prum_AssemblySummary_Summary.xlsx). Contamination filtering (IdentifyGoodSeqsViaReadsMapped.r, GatherALLConSeqsWithOKCoverage.java). In order to filter out possible low-level contaminants, consensus sequences derived from very low coverage assembly clusters (<10 reads) were removed from further analysis. After filtering, consensus sequences were grouped by locus (across individuals) in order to produce sets of homologues. Orthology (GetPairwiseDistanceMeasures.java, plotMDS5.r). Orthology was then determined for each locus as follows. First, a pairwise distance measure was computed for pairs of homologues. To compute the pairwise distance between two sequences, we computed the percent of 20-mers observed in the two sequences that were found in both sequences. Note that the list of 20-mers was constructed from consecutive 20-mers as well as spaced 20-mers (every third base), in order to allow increased levels of sequence divergence. Using the distance matrix, we clustered the sequences using a neighbour-joining algorithm, but allowing at most one sequence per species to be in a given cluster. Clusters containing fewer than 50% of the species were removed from downstream processing. Alignment (MAFFT). Sequences in each orthologous set were aligned using MAFFT v7.023b31 with “–genafpair” and “–maxiterate 1000” flags. Alignment Trimming (TrimAndMaskRawAlignments3). The alignment for each locus was then trimmed/masked using the following procedure. First, each alignment site was identified as ‘good’ if the most common character observed was present in >40% of the sequences. Second, 20 bp regions of each sequence that contained <10 good sites were masked. Third, sites with fewer than 12 unmasked bases were removed from the alignment. Lastly, entire loci were removed if both outgroups or more than 40 taxa were missing. This filter yielded 259 trimmed loci containing fewer than 2.5% missing characters overall. To minimize the overall model complexity while accurately accounting for substitution processes, we performed a partition-model sensitivity analysis with the development version of PartitionFinder v2.0 (ref. 13), sensu14, and compared a complex partition-model (one partition per gene) to a heuristically optimized (relaxed clustering with the RAxML option for accelerated model selection) partition-model using BIC. Based on a candidate pool of potential partitioning strategies that spanned a single partition for the entire data set to a model allowing each locus to represent a unique partition, the latter approach suggested that 75 partitions of our data set represented the best-fitting partitioning scheme, which reduced the number of necessary model parameters by 71%, and hugely decreased computation time. We analysed each individual locus in RAxML v8.0.20 (ref. 18), and then the concatenated alignment, using the two partitioning strategies identified above with both maximum likelihood and Bayesian based approaches in RAxML v8.0.20, and ExaBayes v1.4.2 9 (ref. 34). For each RAxML analysis, we executed 100 rapid bootstrap inferences and thereafter a thorough ML search using a GTR+Γ model of nucleotide substitution for each data set partition. Although this may potentially over-parameterize a partition with respect to substitution model, the influence of this form of model over-parameterization has been found to be negligible in phylogenetic inference35. For the Bayesian analyses, we ran four Metropolis-coupled ExaBayes replicates for 10 million generations, each with three heated chains, and sampling every 1,000 generations (default tuning and branch swap parameters; branch lengths among partitions were linked). Convergence and proper sampling of the posterior distribution of parameter values were assessed by checking that the effective sample sizes of all estimated parameters and branch lengths were greater than 200 in the Tracer v1.6 software36 (most were greater than 1,000), and by using the ‘sdsf’ and ‘postProcParam’ tools included with the ExaBayes package to ensure the average standard deviation of split frequencies and potential scale reduction factors across runs were close to zero and one, respectively. Finally, to check for convergence in topology and clade posterior probabilities, we summarized a greedily refined majority-rule consensus tree (default) from 10,000 post burn-in trees using the ExaBayes ‘consense’ tool for each run independently and then together. Analyses of the reduced data set referenced in the main text were conducted using the same partition-model as the full data set. To explore variation in gene tree topology and to look for outliers that might influence combined analysis, we calculated pairwise Robinson-Foulds37 (RF) and Matching Splits (MS) tree distances implemented in TreeCmp38. We then visualized histograms of tree distances and multidimensional scaling plots in R, and estimated neighbour-joining ‘trees-of-trees’ in the Phangorn R package sensu lato39, 40. Using RF and MS distances, outlier loci were identified as those that occurred in the top 10% of pairwise distances for >30 comparisons to other loci (~10%) in the data set. We also identified putative outlier loci using the kdetrees.complete function of the kdetrees R package41. All three methods identified 13 of the same loci as potential outliers; however removal of these loci from the analysis had no effect on estimating topology or branch lengths. Although fully parametric estimation (for example, *BEAST, see ref. 42) of a coalescent species tree with hundreds of genes and hundreds of taxa is not currently possible, we estimated species trees using three gene-tree summation methods that have been shown to be statistically consistent under the multispecies coalescent model43. First, we used the STRAW web server44 to estimate bootstrapped species trees using the STAR19 and NJ-ST20 algorithms (also available through STRAW). The popular MP-EST45 method cannot currently work for more than ~50 taxa. STAR takes rooted gene trees and uses the average ranks of coalescence times19 to build a distance matrix from which a species tree is computed with the neighbour-joining method46. By contrast, NJst applies the neighbour-joining method to a distance matrix computed from average gene-tree internode distances, and relaxes the requirement for input gene trees to be rooted20. We also summarized a species tree with the ASTRAL 4.7.6 algorithm. With simulated data, ASTRAL has been shown to outperform concatenation or other summary methods under certain amounts of incomplete lineage sorting21. For very large numbers of taxa and genes, ASTRAL uses a heuristic search to find the species tree that agrees with the largest number of quartet trees induced by the set of input gene trees. For analysis with ASTRAL, we also attempted to increase the resolution of individual gene trees (Supplementary Fig. 2) by generating supergene alignments using the weighted statistical binning pipeline of refs 47, 48 with a bootstrap score of 0.75 as a bin threshold. STAR, NJst (not shown), and the binned ASTRAL (Supplementary Fig. 3) analysis produced virtually identical inferences when low support branches (<0.75) were collapsed, and differed only with respect to the resolution of a few branches. NJst resolved the Passeroidea (Fringilla plus Spizella) as the sister group to a paraphyletic sample of Sylvioidea (Calandrella, Pycnonotus, and Sylvia), while STAR does not resolve this branch. Comparing STAR/NJst to ASTRAL, we find five additional differences: (1) within tinamous, STAR/NJst resolves Crypturellus as sister to the rest of the tinamous, whereas ASTRAL resolves Crypturellus as sister to Tinamus (similar to ExaBayes/RAxML); (2) STAR/NJst resolves pigeons as sister to a clade containing Mesitornithiformes and Pteroclidiformes, while ASTRAL does not resolve these relationships; (3), STAR/NJst fails to resolve Oxyruncus and Myiobius as sister genera, while ASTRAL does (similar to RAxML/ExaBayes); (4), in STAR/NJst, bee-eaters (Merops) are resolved as the sister group to coraciiforms (congruent with ref. 4), while ASTRAL resolves bee-eaters as sister to the rollers (Coracias) (similar to RAxML/ExaBayes); (5) lastly, in STAR/NJst, buttonquail (Turnix) is resolved as sister to the most inclusive clade of Charadriiformes not including Burhinus, Charadrius, Haematopus, and Recurvirostra, while in ASTRAL, buttonquail is resolved as sister to a clade containing Glareola, Uria, Rynchops, Sterna, and Chroicocephalus (similar to RAxML/ExaBayes). Although lower level relationships detected with concatenation are generally recapitulated in the species trees, few of the higher level, or interordinal, relationships are resolved. This lack of resolution of the gene-tree species-tree based inferences relative to the inferences based on concatenation are not surprising, as it is increasingly recognized that the phylogenetic information content required to resolve the gene-tree histories of individual loci becomes scant at deep timescales47. Despite our extensive taxon sampling and the slow rate of nucleotide substitution that characterizes loci captured using anchored enrichment12, no single locus was able to fully resolve a topology, and this lack of information will challenge the accuracy of any coalescent-based summary approach relative to concatenation49, 50, 51, 52, 53, 54. Finally, all summation methods tested here assume a priori that the only source of discordance among gene trees is deep coalescence, and violations of this assumption may introduce systematic error in phylogeny estimation54. Site-specific evolutionary rates, λ , were calculated for each locus using the program HyPhy55 in the PhyDesign web interface56 in conjunction with a guide chronogram generated by a nonparametric rate smoothing algorithm57 applied to our concatenated RAxML tree. Using these rates to predict whether an alignment will yield correct, incorrect, or no resolution of a given node, we quantified the probability of phylogenetically informative changes (ψ)16 contributing to the resolution of the earliest divergences in Neoaves. Estimates generated under a three character state model58 reveal that the majority of loci have a strong probability of ψ, and suggest a high potential for most loci and partitions containing multiple loci (assigned by PartitionFinder) to correctly resolve this internode. The potential for resolution as a consequence of phylogenetic signal is therefore high relative to the potential for saturation and misleading inference induced by stochastic changes along the subtending lineages (Supplementary Fig. 4a). To assess the information content of the loci across the entire topology, we profiled their phylogenetic informativeness (PI)15, (Supplementary Fig. 4b). There was considerable variation in PI across loci (Supplementary Fig. 4). In all cases, the loci with the lowest values of ψ are categorized by substantially lower (60–90%) values of PI, rather than sharp declines in their PI profiles. The absence of a sharp decline in the PI profile suggests that a lack of phylogenetic information, rather than rapid increases in homoplasious sites, underlie low values of the probability of signal ψ59. Because declines in PI can be attributed to increases in homoplasious site patterns59, we further assessed the phylogenetic utility of data set partitions by quantifying the ratio of PI at the most recent common ancestor of Neoaves to the PI at the most recent common ancestor of Aves (Supplementary Fig. 4c). Values of this ratio that are less than 1 correspond to a rise in PI towards the root. Values close to 1 correspond to fairly uniform PI. Values greater than 1 correspond to a decline in PI towards the root. Sixty-six out of 75 partitions demonstrated less than a 50% percent decline in PI, and only six partitions demonstrated a decline of PI greater than 75% (Supplementary Fig. 4c). As all but a few nodes in this study represent divergences younger than the crown of Neoaves, these ratios of PI suggest that the predicted impact of homoplasy on our topological inferences should be minimal. As PI profiles do not directly predict the impact of homoplasious site patterns on topological resolution16, 60, we evaluated probabilities of ψ for focal nodes using both the concatenated data set as well as individual loci that span the variance in locus lengths. Concordant with expectations from the PI profiles, all quantifications strongly support the prediction that homoplasy will have a minimal impact on topological resolution for the concatenated data set across a range of tree depths and internode distances (ψ = 1.0 for all nodes), while individual loci vary in their predicted utility (Supplementary Fig. 4d). As the guide tree does not represent a true known tree, we additionally quantified ψ across a range of tree depths and internode distances to test if our predictions of utility are in line with general trends in the data. Concordant with our results above, the concatenated data set is predicted to be of high phylogenetic utility at all timescales (ψ = 1.0 for all nodes), while the utility of individual loci begins to decline for small internodes at deep tree depths (Supplementary Fig. 5). We estimated a time-calibrated tree with a node dating approach in BEAST 1.8.1 (ref. 42) that used 19 well justified fossil calibrations phylogenetically placed by rigorous, apomorphy-based diagnoses (see the descriptions of avian calibration fossils in the Supplementary Information). We used a starting tree topology based on the ExaBayes inference (Fig. 1), and prior node age calibrations that followed a lognormal parametric distribution based on occurrences of fossil taxa. To prevent BEAST from exploring topology space and only allow estimation of branch lengths, we turned off the subtree-slide, Wilson–Balding, and narrow and wide exchange operators61, 62. Finally, we applied a birth–death speciation model with default priors. As rates of molecular evolution are significantly variable across certain bird lineages63, 64, 65, we applied an uncorrelated relaxed clock (UCLN) to each partition of the data set where rates among branches are distributed according to a lognormal distribution66. All dating analyses were performed without crocodilian outgroups to reduce the potential of extreme substitution rate heterogeneity to bias rate and consequent divergence time estimates of the UCLN model67. All calibrations were modelled using soft maximum age bounds to allow for the potential of our data to overwhelm our user-specified priors68. Soft maximum bounds are the preferred method for assigning upper limits on the age of phylogenetic divergences69. As effective priors necessarily reflect interactions between user specified priors, topology, and the branching-model, they may not precisely reflect the user-specified priors70. To correct for this potential source of error, we carefully examined the effective calibration priors by first running the prepared BEAST XML without any nucleotide data (until all ESS values were above 200). We then iteratively adjusted our user-defined priors until all of the effective priors (as examined in the Tracer software) reflected the intended calibration densities. Finally, using the compare.phylo function in the Phyloch R package, we examined how the inclusion of molecular data influenced the divergence time estimates relative to the effective prior (Supplementary Fig. 9; see below). Our initial approach was to set a prior’s offset to the age of its associated fossil; the mean was then manually adjusted such that 95% of the calibration density fell more recently than the K–Pg boundary at 65 Ma (million years ago) (the standard deviation was fixed at 1 Ma). In general, priors constructed this way generated calibration densities that specified their highest density peak (their mode) about 3–5 million years older than the age of the offset. We applied a loose gamma prior to the node reflecting the most recent common ancestor of crown birds—we used an offset of 60.5 Ma (the age of the oldest known definitive, uncontroversial crown bird fossil; the stem penguin Waimanu), and adjusted the scale and shape of the prior such that 97.5% of the calibration density fell more recently than 86.5 Ma71 (see below and Supplementary Information for discussion of the >65 Ma putative crown avian Vegavis). This date (86.5 Ma) reflects the upper bound age estimate of the Niobrara Formation—one of many richly fossiliferous Mesozoic deposits exhibiting many crownward Mesozoic stem birds, without any trace of avian crown group representatives. The Niobrara, in particular, has produced hundreds of stem birds and other fragile skeletons, without yielding a single crown bird fossil, and therefore represents a robust choice for a soft upper bound for the root divergence of the avian crown71, 72, 73. Previous soft maxima employed for this divergence have arbitrarily selected the age of other Mesozoic stem avians (that is, Gansus yumenensis, 110 Ma) that are phylogenetically stemward of the Niobrara taxa28. Although the implementation of very ancient soft maxima such as the age of Gansus are often done in the name of conservatism, the extremely ancient divergence dates yielded by such analyses illustrate the misleading influence of assigning soft maxima that are vastly too old to be of relevance to the divergence of crown group birds74. However, this problem has been eliminated in some more recent analyses75. All of the fossil calibrations employed in our analysis represent neognaths; rootward divergences within Aves (for example the divergence between Palaeognathae and Neognathae, and Galloanserae and Neoaves) cannot be confidently calibrated due to a present lack of fossils representing the palaeognath, neognath, galloanserine, and neoavian stem groups. As such, the K–Pg soft bound was only applied to comparatively apical divergences within neognaths. Although the question of whether major neognath divergences occurred during the Mesozoic has been the source of controversy76, 77, 78, renewed surveys of Mesozoic sediments for definitive crown avians or even possible crown neoavians have been unsuccessful (with the possible exception of Vegavis; see Supplementary Information), and together with recent divergence dating analyses have cast doubt on the presence of neoavian subclades before the K–Pg mass extinction1, 74, 79. Further, recent work has demonstrated the tendency of avian divergence estimates to greatly exceed uninformative priors, resulting in spuriously ancient divergence dating results (for example, refs 28, 75, 76, 80). These results motivated our implementation of the 65 Ma soft bound for our neoavian calibrations. Contrary to expectation, when we compared the effective prior on the entire tree to the final summary derived from the posterior distribution of divergence times (Supplementary Fig. 9), we found no overall trend of posterior estimated ages post-dating prior calibrations. In fact, the inclusion of our molecular data decreases the inferred ages of almost all of the deepest nodes in our tree. A similar result has been obtained for mammals by using large amounts of nuclear DNA sequences81. Future work investigating the interplay of the density of genomic sampling and the application of various calibration age priors will be indispensible for sensitivity analyses to help us further develop a robust timescale of avian evolution. However, the pattern of posterior versus prior age estimates observed in our study raises the prospect that the new class of data used in this study (that is, semi-conserved anchor regions) may exhibit some immunity to longstanding problems associated with inferring avian divergence times, such as systematically over-estimating the antiquity of extant avian clades. In addition to making predictions about the phylogenetic utility of a locus or partition towards topological resolution, PI profiles have recently also been used to mitigate the influence of substitution saturation on divergence time estimates82. Given the variance in PI profile shapes for captured loci and their subsequent partition assignments (Supplementary Fig. 4c), and observations that alignments and subsets of data alignments characterized by high levels of homoplasy can mislead branch length estimation83, 84, we limited our divergence time estimates to 36 partitions that did not exhibit a decline in informativeness towards the root of the tree. We ran BEAST on each partition separately until parameter ESS values were greater than 200 (most were greater than 1,000) to ensure adequate posterior sampling of each parameter value. After concatenating 10,000 randomly sampled post burn-in trees from each of these completed analyses, we summarized a final MCC tree with median node heights in TreeAnnotator v1.8.1 (ref. 42). Supplementary Fig. 6 shows the full, calibrated Bayesian tree (Fig. 1) with 95% HPD confidence intervals on the node ages, and Supplementary Fig. 7 shows the distribution of estimated branching times, ranked by median age (using clade numbers from Fig. 1). All computations were carried out on 64-core PowerEdge M915 nodes on the Louise Linux cluster at the Yale University Biomedical High Performance Computing Center. No statistical methods were used to predetermine sample size.

News Article
Site: http://www.cemag.us/rss-feeds/all/rss.xml/all

An international group of physicists has discovered a phenomenon of large magnitude in an unexpected class of materials that can lead to a variety of devices used in optical systems. That phenomenon — the elasto-optic effect — characterizes the formation of a periodic variance of light refraction when an acoustic wave propagates in optical materials, says Yurong Yang, a research assistant professor at the University of Arkansas who led the research. “We found a significantly large elasto-optic effect in thin films made of materials called ferroelectrics,” Yang says, “which are usually considered for their changes in mechanical energy into electrical energy and vice versa, as well in multiferroelectric thin films, which are commonly investigated because of the possible control of their magnetic response by electric input, as well as of their electric response by magnetic input.” The research group published its findings in a paper in Physical Review Letters, the journal of the American Physical Society. A second paper describing the research was published in Nature Communications, an online journal published by the journal Nature. “Those discoveries of a large elasto-optic effect in ferroelectrics and multiferroelectrics therefore broaden the potential of these materials since they can now be put in use to also control their optical responses by elastic property,” says Laurent Bellaiche, Distinguished Professor of physics at the U of A, “which suggests exciting device opportunities arising from this overlooked coupling in these classes of materials.” Yang and Bellaiche, who holds the Twenty-First Century Endowed Professorship in Nanotechnology and Science Education, both conduct research in the Institute for Nanoscience and Engineering and physics department at the U of A. The researchers performed calculations on supercomputers at the Arkansas High Performance Computing Center and a U.S. Department of Defense supercomputing resource. The results published in Physical Review Letters were obtained through a collaborative effort with Zhigang Gui, a U of A physics graduate who is now a postdoctoral research associate at the University of Delaware; Lan Chen and X.K. Meng at Nanjing University in China, and Daniel Sando and Manuel Bibes at University of Paris-Sud in France. The results published in Nature Communications were obtained through a collaborative effort with Daniel Sando and Manuel Bibes and Cecile Carretero, Vincent Garcia, Stephane Fusil, and Agnes Barthelemy at the University of Paris-Sud; Eric Bousquet and Philippe Ghosez at the University of Liege in Belgium; and Daniel Dolfi of Thales Research and Technology in France. Source: University of Arkansas

Discover hidden collaborations