Learning in the machine: Recirculation is random backpropagation.

Learning in physical neural systems must rely on learning rules that are local in both space and time. Optimal learning in deep neural architectures requires that non-local information be available to the deep synapses. Thus, in general, optimal learning in physical neural systems requires the presence of a deep learning channel to communicate non-local information to deep synapses, in a direction opposite to the forward propagation of the activities. Theoretical arguments suggest that for circular autoencoders, an important class of neural architectures where the output layer is identical to the input layer, alternative algorithms may exist that enable local learning without the need for additional learning channels, by using the forward activation channel as the deep learning channel. Here we systematically identify, classify, and study several such local learning algorithms, based on the general idea of recirculating information from the output layer to the hidden layers. We show through simulations and mathematical derivations that these algorithms are robust and converge to critical points of the global error function. In most cases, we show that these recirculation algorithms are very similar to an adaptive form of random backpropagation, where each hidden layer receives a linearly transformed, slowly-varying, version of the output error.

Clinical veterinary proteomics: Techniques and approaches to decipher the animal plasma proteome.

Over the last two decades, technological advancements in the field of proteomics have advanced our understanding of the complex biological systems of living organisms. Techniques based on mass spectrometry (MS) have emerged as powerful tools to contextualise existing genomic information and to create quantitative protein profiles from plasma, tissues or cell lines of various species. Proteomic approaches have been used increasingly in veterinary science to investigate biological processes responsible for growth, reproduction and pathological events. However, the adoption of proteomic approaches by veterinary investigators lags behind that of researchers in the human medical field. Furthermore, in contrast to human proteomics studies, interpretation of veterinary proteomic data is difficult due to the limited protein databases available for many animal species. This review article examines the current use of advanced proteomics techniques for evaluation of animal health and welfare and covers the current status of clinical veterinary proteomics research, including successful protein identification and data interpretation studies. It includes a description of an emerging tool, sequential window acquisition of all theoretical fragment ion mass spectra (SWATH-MS), available on selected mass spectrometry instruments. This newly developed data acquisition technique combines advantages of discovery and targeted proteomics approaches, and thus has the potential to advance the veterinary proteomics field by enhancing identification and reproducibility of proteomics data.

Enhanced Higgs boson to τ(+)τ(-) search with deep learning.

The Higgs boson is thought to provide the interaction that imparts mass to the fundamental fermions, but while measurements at the Large Hadron Collider (LHC) are consistent with this hypothesis, current analysis techniques lack the statistical power to cross the traditional 5σ significance barrier without more data. Deep learning techniques have the potential to increase the statistical power of this analysis by automatically learning complex, high-level data representations. In this work, deep neural networks are used to detect the decay of the Higgs boson to a pair of tau leptons. A Bayesian optimization algorithm is used to tune the network architecture and training algorithm hyperparameters, resulting in a deep network of eight nonlinear processing layers that improves upon the performance of shallow classifiers even without the use of features specifically engineered by physicists for this application. The improvement in discovery significance is equivalent to an increase in the accumulated data set of 25%.

Searching for exotic particles in high-energy physics with deep learning.

Collisions at high-energy particle colliders are a traditionally fruitful source of exotic particle discoveries. Finding these rare particles requires solving difficult signal-versus-background classification problems, hence machine-learning approaches are often used. Standard approaches have relied on 'shallow' machine-learning models that have a limited capacity to learn complex nonlinear functions of the inputs, and rely on a painstaking search through manually constructed nonlinear features. Progress on this problem has slowed, as a variety of techniques have shown equivalent performance. Recent advances in the field of deep learning make it possible to learn more complex functions and better discriminate between signal and background classes. Here, using benchmark data sets, we show that deep-learning methods need no manually constructed inputs and yet improve the classification metric by as much as 8% over the best current approaches. This demonstrates that deep-learning approaches can improve the power of collider searches for exotic particles.

Tumor development in the Beckwith-Wiedemann syndrome is associated with a variety of constitutional molecular 11p15 alterations including imprinting defects of KCNQ1OT1.

Dysregulation of imprinted genes on human chromosome 11p15 has been implicated in Beckwith-Wiedemann syndrome (BWS), an overgrowth syndrome associated with congenital malformations and tumor predisposition. The molecular basis of BWS is complex and heterogeneous. The syndrome is associated with alterations in two distinct imprinting domains on 11p15: a telomeric domain containing the H19 and IGF2 genes and a centromeric domain including the KCNQ1OT1 and CDKNIC genes. It has been postulated that disorders of imprinting in the telomeric domain are associated with overgrowth and cancer predisposition, whereas those in the centromeric domain involve malformations but not tumor development. In this study of 125 BWS cases, we confirm the association of tumors with constitutional defects in the 11p15 telomeric domain; six of 21 BWS cases with uniparental disomy (UPD) of 11p15 developed tumors and one of three of the rare BWS subtype with hypermethylation of the H19 gene developed tumors. Most importantly, we find that five of 32 individuals with BWS and imprinting defects in the centromeric domain developed embryonal tumors. Furthermore, the type of tumors observed in BWS cases with telomeric defects are different from those seen in BWS cases with defects limited to the centromeric domain. Whereas Wilms' tumor was the most frequent tumor seen in BWS cases with UPD for 11p15 or H19 hypermethylation, none of the embryonal tumors with imprinting defects at KCNQ1OT1 was a Wilms' tumor. This suggests that distinct tumor predisposition profiles result from dysregulation of the telomeric domain versus the centromeric domain and that these imprinting defects activate distinct genetic pathways for embryonal tumorigenesis.

Directional resolution of synthetic holliday structures by the Cre recombinase.

The Cre recombinase of bacteriophage P1 cleaves its target site, loxP, in a defined order. Recombination is initiated on one pair of strands to form a Holliday intermediate, which is then resolved by cleavage and exchange of the other pair of strands to yield recombinant products. To investigate the influence of the loxP sequence on the directionality of resolution, we constructed synthetic Holliday (chi) structures containing either wild-type or mutant lox sites. We found that Cre preferentially resolved the synthetic wild-type chi structures on a particular pair of strands. The bias in the direction of resolution was dictated by the asymmetric loxP sequence since the resolution bias was abolished with symmetric lox sites. Systematic substitutions of the loxP site revealed that the bases immediately 5' to the scissile phosphodiester bonds were primarily responsible for the directionality of resolution. Interchanging these base pairs was sufficient to reverse the resolution bias. The Cre-lox co-crystal structures show that Lys(86) makes a base-specific contact with guanine immediately 5' to one of the scissile phosphates. Substituting Lys(86) with alanine resulted in a reduction of the resolution bias, indicating that this amino acid is important for establishing the bias in resolution.

Chimeras of the Flp and Cre recombinases: tests of the mode of cleavage by Flp and Cre.

The Flp and Cre recombinases are members of the integrase family of tyrosine recombinases. Each protein consists of a 13 kDa NH(2)-terminal domain and a larger COOH-terminal domain that contains the active site of the enzyme. The COOH-terminal domain also contains the major determinants for the binding specificity of the recombinase to its cognate DNA binding site. All family members cleave the DNA by the attachment of a conserved nucleophilic tyrosine residue to the 3'-phosphate group at the sites of cleavage. In order to gain further insights into the determinants of the binding specificity and modes of cleavage of Flp and Cre, we have made chimeric proteins in which we have fused the NH(2)-terminal domain of Flp to the COOH-terminal domain of Cre ("Fre") and the NH(2)-terminal domain of Cre to the COOH-terminal domain of Flp ("Clp"). These chimeras have novel binding specificities in that they bind strongly to hybrid sites containing elements from both the Flp and Cre DNA targets but poorly to the native target sites. In this study we have taken advantage of the unique binding specificities of Fre and Clp to examine the mode of cleavage by Cre, Flp, Fre and Clp. We find that the COOH-terminal domain of the recombinases determines their mode of cleavage. Thus Flp and Clp cleave in trans whereas Cre and Fre cleave in cis. These results agree with the studies of Flp and with the cocrystal structure of Cre bound to its DNA target site. They disagree with our previous findings that Cre could carry out trans cleavage. We discuss the variations in the experimental approaches in order to reconcile the different results.

Trans complementation of variant Cre proteins for defects in cleavage and synapsis.

The Cre recombinase is a member of the integrase family of conservative site-specific recombinases. These proteins share five conserved catalytic residues, one of which is a tyrosine that acts as the nucleophile to attack the scissile phosphodiester bond in the DNA target. Recombination by the Cre recombinase takes place in a supramolecular structure called a synapse that consists of four molecules of Cre bound to two DNA target sequences called lox sites. The synapse is held together by an intricate network of protein-protein interactions. They bend the two sites into square planar structure that resembles a Holliday intermediate. We have studied three mutant Cre proteins that appear to have defects in synapsis (Cre A36V, Cre T41F, and Cre G314R). We found that they were unable to carry out strand cleavage but that cleavage occurred if they were mixed with a cleavage-defective Cre protein that lacks the catalytic nucleophilic tyrosine residue. The three variant proteins could also be complemented for the formation of a novel structure ("complexV"), which may be a cleaved synaptic intermediate. We suggest that these three mutant proteins have a defect in DNA bending and discuss the relationship between bending, synapsis, and cleavage.

DNA sequence determinant for FIp-induced DNA bending.

The Flp site-specific recombinase from Saccharomyces cerevisiae induces DNA bending upon interaction with the Flp recognition target (FRT) site. The minimal FRT site comprises the inverted a and b binding elements, which flank a central 8 bp core region. The DNA bend in a complex of two Flp monomers bound to the FRT site is located in the middle of the core region. When the central AT basepair was replaced with a CG, the DNA bend was positioned at the outside end of the core region adjacent to the a binding element. The other basepairs surrounding the central AT basepair were not important to the position of Flp-induced bends. The change also decreased Flp-mediated cleavage of the top strand of the FRT site and increased Flp-mediated cleavage of the bottom strand. The overall recombination proficiency of the site was impaired. We conclude that the central AT basepair provides a point of flexure in the FRT site, which Flp uses to position the bend in dimeric Flp-DNA complexes, and that the structure of the core DNA influences the functionality of the site.

Determinants of the position of a Flp-induced DNA bend.

The Flp site-specific recombinase from Saccharomyces cerevisiae induces DNA bending upon interaction with the Flp recognition target (FRT) site. The minimal FRT site is comprised of two inverted binding elements which flank a central core region. Binding of a single monomer of Flp to DNA induces a DNA bend of 60 degrees. The position of this bend differed depending on whether the substrate contained a single binding element or a two-element FRT site. In the present work we tested and disproved a model in which a single Flp monomer interacts with both symmetry elements of a single FRT site. Likewise, we showed that a model in which a Flp monomer dissociates from a singly occupied FRT site and reassociates with the unbound element of another singly occupied FRT site during electrophoresis, does not account for the apparent shift in the position of the bend centre. It seems that the movement of a Flp monomer between the a and b elements of one FRT site during electrophoresis accounts for this anomaly. The position of the DNA bend resulting from the association of a Flp monomer with the FRT site is also influenced by the DNA sequences flanking the site. We conclude that attempts to measure the bend centre of a complex of one Flp molecule bound to a DNA containing two binding elements give misleading results. The position of the bend is more accurately measured in the presence of a single binding element.

Selection of novel, specific single-stranded DNA sequences by Flp, a duplex-specific DNA binding protein.

Flp is a member of the integrase family of site-specific recombinases. Flp is known to be a double-stranded (ds)DNA binding protein that binds sequence specifically to the 13 bp binding elements in the FRT site (Flprecognitiontarget). We subjected a random pool of oligonucleotides to the in vitro binding site selection method and have unexpectedly recovered a series of single-stranded oligonucleotides to which Flp binds with high affinity. These single-stranded oligonucleotides differ in sequence from the duplex FRT site. The minimal length of the oligonucleotides which is active is 29 nt. This single strand-specific DNA binding activity is located in the same C-terminal 32 kDa domain of Flp in which the site-specific dsDNA binding activity resides. Competition studies suggest that the apparent affinity of Flp for single-stranded oligonucleotide is somewhat less than for a complete duplex FRT site but greater than for a single duplex 13 bp binding element. We have also shown that Cre, another member of the integrase family of site-specific recombinases, also exhibits single-stranded DNA binding similar to that of Flp.

The role of single-stranded DNA in Flp-mediated strand exchange.

The Flp recognition target site contains two inverted 13-base pair (bp) Flp binding sequences that surround an 8-bp core region. Flp recombinase has been shown to carry out strand ligation independently of its ability to execute strand cleavage. Using a synthetic activated DNA substrate bearing a 3'-phosphotyrosine group, we have developed an assay to measure strand exchange by Flp proteins. We have shown that wild-type Flp protein was able to catalyze strand exchange using DNA substrates containing 8-bp duplex core sequences. Mutant Flp proteins that are defective in either DNA bending or DNA cleavage were also impaired in their abilities to carry out strand exchange. The inability of these mutant proteins to execute strand exchange could be overcome by providing a DNA substrate containing a single-stranded core sequence. This single-stranded core sequence could be as small as 3 nucleotides. Full activity of mutant Flp proteins in strand exchange was observed when both partner DNAs contained an 8-nucleotide single-stranded core region. Using suicide substrates, we showed that single-stranded DNA is also important for strand exchange reactions where Flp-mediated strand cleavage is required. These results suggest that the ability of Flp to induce DNA bending and strand cleavage may be crucial for strand exchange. We propose that both DNA bending and strand cleavage may be required to separate the strands of the core region and that single-stranded DNA in the core region might be an intermediate in Flp-mediated DNA recombination.

Asymmetry in Flp-mediated cleavage.

Flp is a member of the integrase family of site-specific recombinases. Members of the integrase family mediate DNA strand cleavage via a transesterification reaction involving an active site tyrosine residue. The first step of the reaction results in covalent linkage of the protein to the 3'-phosphoryl DNA terminus, leaving a 5'-hydroxyl group at the site of the nick. We have used Flp recognition target (FRT) sites containing a 5'-bridging phosphorothioate linkage at the site of Flp cleavage to accumulate intermediates in which Flp is covalently bound at a cleavage site. We have probed these intermediates with dimethylsulfate using methylation protection and find that Flp-mediated cleavage is associated with protection of two adenine residues that are opposite the sites of cleavage and covalent attachment by Flp. Methylation interference studies showed that cleavage and covalent attachment are also accompanied by differences in the contacts of Flp with each of the two cleavage sites and with the surrounding symmetry elements. Therefore, we provide evidence that Flp-mediated cleavage and covalent attachment result in changes to the conformation of the Flp-FRT complex. These changes may be required for Flp-mediated strand exchange activity.

Topological analysis of the role of homology in Flp-mediated recombination.

Recombination by the Flp recombinase of Saccharomyces cerevisiae is known to be inhibited by heterology of the overlap regions of the two recombining DNA targets (FRT sites). We have used topological analysis to show that Flp can promote two rounds of intramolecular recombination between heterologous FRT sites contained within the same supercoiled plasmid. The products are in parental nonrecombinant configuration. Thus, heterology may appear to "block" recombination by rendering the heteroduplex recombinant products unstable, thus favoring a second round of recombination to homoduplex (but parental) products. Hence, homology in the core region is not a requirement for the recombination reaction by Flp but for the formation of recombinant products.

The Cre recombinase cleaves the lox site in trans.

The Cre protein is a conservative site-specific recombinase that is encoded by bacteriophage P1. Its function in vivo is to resolve dimeric lysogenic P1 plasmids that arise by general recombination. In this way Cre facilitates effective partition of the P1 prophage. Cre is a member of the integrase family of conservative site-specific recombinases. Cleavage of the DNA by the integrases involves covalent attachment of a conserved nucleophilic tyrosine to the 3'-phosphoryl end at the site of the break. We have used in vitro complementation tests to show that the Cre protein, like the Flp protein of the 2-microm plasmid of Saccharomyces cerevisiae, cleaves its target lox site in trans. Moreover, the data are compatible with two modes of cleavage; one requires the reconstitution of a pseudo full-site from half-sites and the other requires the assembly of a higher order complex that resembles a synaptic complex.

The Flp recombinase cleaves Holliday junctions in trans.

The Flp site-specific recombinase is encoded by the 2 micrometers plasmid Saccharomyces cerevisiae and is a member of the integrase family of recombinases. Like all members of the integrase family studied, Flp mediates recombination in two steps. First, a pair of strand exchanges creates a Holliday-like intermediate; second, this intermediate is resolved to recombinant products by a second pair of strand exchanges. Evidence derived from experiments using linear substrates indicates that Flp's active site is composed of two Flp protomers. One binds to the Flp recognition target site (FRT site) and activates the scissile phosphodiester bond for cleavage. Another molecule of Flp bound elsewhere in the synaptic complex (in trans) donates the nucleophilic tyrosine that executes cleavage and thereby becomes covalently attached to the 3' phosphoryl group at the cleavage site. It has previously been shown that Flp efficiently resolves synthetic, Holliday-like (chi) structures to linear products. In this paper, we examined whether resolution of chi structures by Flp also occurs via the trans cleavage mechanism. We used in vitro complementation studies of mutant Flp proteins as well as nicked chi structures to show that Flp resolves chi structures by trans cleavage. We propose a model for Flp-mediated recombination that incorporates trans cleavage at both the initial and resolution steps of strand exchange.

Cleavage-dependent ligation by the FLP recombinase. Characterization of a mutant FLP protein with an alteration in a catalytic amino acid.

The FLP recombinase of the 2 microM plasmid of Saccharomyces cerevisiae belongs to the integrase family of recombinases whose members have in common four absolutely conserved residues (Arg-191, His-305, Arg-308, and Tyr-343). We have studied the mutant protein FLP R308K in which the arginine residue at position 308 has been replaced by lysine. Although FLP R308K was previously reported to be defective in ligation of certain substrates (Pan, G., Luetke, K., and Sadowski, P.D., Mol. Cell. Biol. 13, 3167-3175, 1993b), we show in this work that the protein is able to ligate those substrates that it can cleave (cleavage-dependent ligation activity). FLP R308K is defective in in vitro recombination and in strand exchange. It is able to carry out strand exchange at one of the two cleavage sites of the FLP recognition target site (FRT site), but is defective in strand exchange at the other cleavage site. These results are consistent with a model in which wild-type FLP initiates recombination only at one of the two cleavage sites. FLP R308K may be defective in the initiation of recombination.

The role of DNA bending in Flp-mediated site-specific recombination.

The Flp recombinase of the 2 micron plasmid of Saccharomyces cerevisiae binds to a recognition target site, induces DNA bending and catalyses DNA cleavage and strand exchange to bring about recombination. The minimal Flp recognition target site contains two Flp binding sequences flanking an 8 bp core region; binding of Flp results in the formation of two Flp:DNA complexes (complexes I and II). Binding of a Flp monomer to a single symmetry element generates a DNA bend of about 60 degrees (a type I bend), whereas binding of two Flp monomers to the FRT site generates a DNA bend of > 144 degrees (a type II bend). We have used circular permutation analysis to locate the centre of the type I and type II DNA bends induced by Flp, and the Flp peptides P27 (27 kDa; amino acid residues 124 to 346) and P32 (32 kDa; amino acid residues 124 to 423). The location of the centre of the type I bend depends upon whether the substrate contains one or two Flp binding elements. When the substrate contains one symmetry element, the centre of the type I bend induced by Flp is located at the core-distal end of the b element. However, it is located at the core-proximal end of the b element when the substrate contains two Flp-binding elements. The P27 and P32 peptides, which lack the NH2-terminal 13 kDa region of Flp, do not show this behaviour. We deduce that the 13 kDa region of Flp is critical for the positioning of the type I bend centre on a minimal Flp recognition site. We propose a model in which a single molecule of Flp interacts with two symmetry elements to account for these results. The centre of the type II bend induced by Flp is in the middle of the core region. We used ligation-defective Flp proteins to determine the location of the type II bend centres in complexes where either the top or bottom strand was cleaved. The bend centres of such complexes depend upon which strand is cleaved. We propose a model which associates the position of Flp-induced type II bends with a defined order of strand exchanges in the recombination reaction.

Functional analysis of the ABF1-binding sites within the Ya regions of the MATa and HMRa loci of Saccharomyces cerevisiae.

Cell type in the yeast Saccharomyces cerevisiae is determined by information present at the MAT locus. Cells can switch mating types when cell-type information located at a silent locus, HML or HMR, is transposed to the MAT locus. The HML and HMR loci are kept silent through the action of a number of proteins, one of which is the DNA-binding protein, ABF1. We have identified a binding site for ABF1 within the Ya region of MATa and HMRa. In order to examine the function of this ABF1-binding site, we have constructed strains that lack the site in the MATa or HMRa loci. Consistent with the idea that ABF1 plays a redundant role in silencing, it was found that a triple deletion of the ABF1-binding sites at HMRE, Ya and I did not permit the expression of HMRa. We have also shown that chromosomal deletion of the binding site at MATYa had no effect on the level of cutting by the HO endonuclease nor on the amount of mating-type switching observed. Similarly, chromosomal deletion of all three ABF1-binding sites at HMRa had no effect on the directionality of mating-type switching.

Homology requirements for ligation and strand exchange by the FLP recombinase.

The FLP recombinase of the 2-microns plasmid of Saccharomyces cerevisiae belongs to the integrase family whose members form a covalent bond between a conserved tyrosine of the recombinase and the 3'-phosphoryl group at the site of cleavage. Ligation takes place when the 5'-OH generated during the cleavage step attacks the phosphotyrosine bond and reforms a phosphodiester bond. When the incoming 5'-OH is from the partner duplex, strand exchange occurs. The FLP recognition target (FRT) contains two inverted 13-base pair (bp) FLP binding sequences that surround an 8-bp core region. It has been shown that heterology in the core regions of the recombinase FLP recognition target sites can dramatically impair recombination. Therefore, it was of interest to study the homology requirements of the core sequence for FLP-mediated ligation. Using nicked duplex substrates containing mismatches in the core sequence, we have demonstrated that the FLP ligation reaction can tolerate mismatches at all positions in the 8-bp core except the position immediately adjacent to the cleavage site. Using half-FRT substrates that contain a single-stranded core sequence, we showed that 4 base pairs adjacent to the cleavage site in the core are required for FLP to execute ligation with a single-stranded oligonucleotide. FLP is also able to ligate the protruding single strand on a half-FRT site to the opposite strand to form a hairpin. We have studied the effect of the base composition of the protruding 8-nucleotide single strand upon the efficiency of hairpin ligation. These studies revealed the importance of intrastrand complementarity in the formation of hairpin by FLP. Hence we conclude that the homology in the position adjacent to the cleavage site is most important, and the degree of the homology required is dependent on the nature of the ligation assay.