unedited notes taking during talks. David Arnosti Thermodynamic and really dynamics modeling of transcription. how transcription factors talk to eachother on enhancers short-range and long-range repressors – global deacytilation (focus on short). Conservation of eve stripe enhancers despite limited conservation of core sequence. ‘Billboard’ enhancer model – same message conveyed by symbols in 500 bp region
DNA binding motifs are informative. What grammar matters? Spacing, stoichiometry, site arrangement binding site affinity, abundance of factors. From sequence to expression pattern. Emphasized population characteristics.Knockout single and double activators and different combinations of single binding site knockouts (fromm 300bp enhancer) drop 40-50%. Double knockouts have between nothing and a very strong effect. (are these physically clustered?)Furlong worried about relative intensity of stains. Stark worried about ‘neutral DNA’, definition of sites. Do you make new sites ? Total sequences vary by only a couple of base-pairs. What if the sequences are very different?“thermodynamic model”: specify all states (permutations of bound and unbound).
Overfitting constraint (having separate coupling for each interaction vs. assuming they are the same).
Penalize with AIC. :(. mutual information between model and data set? Is RMSE best? Are RMSE errors inclusive of AIC corrections? Model doesn’t require long range interactions from snail. Test model on other species rho and mel vnd and twi enhancers.
Dynamic interactions. (twi falls off too fast relative to snail in these plots)? Sinah (U Ill. Urbane-Champagne Comp sci)
Function and evolution of developmental enhancers in drosophila. GEMSTAT model: eq thermodynamic model. Weight all possible configurations of protein binding. + basal Tx machinery (BTM) present or not? Gene expression proption to bound BTM. Activators and repressors can interact with BTM (independently). Statistical signficance of difference still needs development. Test effects of mechanism: such as TF cooperative binding. Are the input distributions static? Model fits better when allowing ‘synergistic’ / multiplicative interaction. > come from independence integrated in time. Short-range repressor, block activator, don’t interact with BTM. Take 20-50 kb and precit expression pattern. Doesn’t work with existing model framework. Alternative approach. Run model on windowed sections of original sequence. Aggregate enhancer outputs. Model succesfully predicts locations and behavior of different eve enhancers within the eve locus. John points out Hairy 2 is eve dependent (pair-rule genes not in model).
No difference between long and short range (only a result of using small enhancer).Pieces outside of known enhancers model predicts would give expression in inappropriate stripes. These correspond to less accessable regions in genome data. Eileen Furlong quite concerned about utility of virtual fly: about averaging and tiramide amplification etc of virtual fly. Non-quantitative data set. → skipped rest or that section.
Modeling TF occupancy. Take top 1000 chip peaks 500 bp arround summit. + 1000 random non-peaks. Compare model predicted accuracy and chip measured occupancy. ‘Occupancy’ not well predicted by number of sites. Allowing cooperative binding helps somewhat. Adding competitive interactions helps more. Facilitative effects of other TFs. Substantial effects from having Zelda and GAGA for predicting binding. (‘chromatin opening’ effect?). Yogi Jaeger Aim: infer constraints on evolutionary variability Repression between overlapping gap genes (substantially weaker). Scutle fly (Megaselia). No caudal. Same gap domains. Clogmia (moth midge). No bcd gradient, yes caudal. 6 eve stripes, no early posterior hb. Relative timing and positions sufficient to infer network. Investigated gap gene regulation by maternal gradients using RNAi. How hard is it to add more insect species to the gap network? Stark Regulatory code: go from DNA sequence to tissue specific expression and timing. Use TF binding peaks as a proxy for enhancers to start. Sequence determinants of TF binding peaks. Very few Twist sequence motifs are bound in vivo. ~40% of binding is stage specific. Zelda, Ttk motif enriched in early twi binding sites (chip peaks). Tin zeste gaga motifs characteristic of late bound peaks. Cact, ths, and wntD enhancers drop substantially when zelda sites mutated show reduced DNAse I accessibility to enhancer with mutated zelda sites from 1.4 to 1.0 Vienna-Tile library of putative Drosophila enhancers 807 flies of 2kb tiles covering 21 Mb. Gal4 reporter ector with minimal promoter. Stain in 96 well plates 4007 lines stained and annotated 2,000 enhancers. Increasing number of active enhancers during development, from 10% to 35% stage 5 to stage 16 use fluorescent gut and DIC image to make calls on stage and orientation for clustering. Considerable redundancy: 3-5 overlapping. Bias to TFs hard to make ectopic calls – might be enhancer of neighboring gene. Compare motif-content of active sequences, can identify motifs characteristic of particular patterns (like dorsal for early mesoderm). motif-content of inactive sequences. (Analayze inactive ) enhancer classes: early mesoderm, procepahlic ectoderm, CNS, midgut, control shuffle assignment of reads to their classes (should give 50%). If most non-active enhancers have repressive marks, the entire non-coding genome has transcriptional effects. Use lines from N Carolina wild population collection. Analyze insertations deletions and SNPs. 873 insertions, 452 deletions 824 complex variants are common among all 39 lines 70-90% of these insertions are ancestral. Rh6 is a null in berkeley genome line (not in w-line or in OreR). Minor deletion bias believed to be a method issue rather than a reality. 80% of protein coding genes are affected by nonSNP variants. SNP density increases near indels (insertions/deletions). True for coding and non-coding SNPs. Old indels don’t have more new SNPs. Revised interpretation. Indels are located in regions prone to mutation? Locations that get indels in some lines get more SNPs in others. Doesn’t appear to be true. Revised: Evidence of admixture. Single event created both indels and SNPs at once. 85% of gene expression variation explainable by separating males and females (principle component 1). Many expression based QTL – eQTLs are sex specific. 53% of cis-eQTLs have both SNP and indels. Not biased to X linked genes. 20% of sex-biased changes are exclusive to somatic tissue. 20% of male specific loci are male tissue specific. And only 5% of female specific QTLs are genes in emale specific tissue. Many eQTL cause changes in only one gender. Number of variants highest just before the TSS, lower just after TSS. Effect on expression highest at TSS and just after TSS. Number of variants drops just before end site. Significance of variance increases at endsite. Weak enrichment in K4me3 at eQTL, under-enriched in K27me3. Conserved non-coding regions also have slightly fewer than expected eQTLs. Yeast 1-hybrid screen for all TFs that bind particular DNA elements. Scalable mobility shift assay by microfluidics. Use intensity of fluorescently labeled DNA. Angela Stathopolous How can dorsal set sog boundary? Plots again the ‘flat’ dorsal. Vnd transiently derepressed at start of cc14. Time for snail to rebind to enhancers slower than for dorsal. Distinguish accumulation from continuous transcription vs. levels of dorsal increasing. Simulate expression boundaries pt 2, size of DV axis varies by 15%, some of the patterns seem to scale. Standard deviation is about 5%. Claim Dorsal gradient changes correspondingly. Plotted as a function of nuclear number. Assume toll signaling is independent of L. claim sog and ind don’t scale and vnd do scale (not very convincing distinction). Dorsal width no longer scales in AP-toll in wind- embryos. Fernando Casares en represses Ci in the ocelli two make two ocelli instead of one. en must be self activating to remain expressed after repressing Hh signalling (since en is initially downstream of En). This autoregulation is requires delta Notch signaling. Wnt inhibits ocellus to eye transformation.
out of battery…
[ML update Stark presented enhancer tiling screen. Now 6,000 2kb regions tested. 50% are enhancers. Enriched motifs indicative of tissue specific pattern.] Denis Thieffry Abibatou Mbodj et al. Dynamical logical modeling. Peak motif finder Nicholas (Carthew and Amaral labs at Northwestern) – Yan network. Omitidia cels specified sequentially during development. Sequentially: R8, R2/5, R3/4, R1/6, R7, clone How differentiation is robustly maintained despite variation in the network components. Make Yan-YFP fusion gene, rescue Yan null transheterozygote. Yan expression is transient, not consistent with bistable model. R8 neurons never gain high level of Yan. Different cell types have different amounts of Yan. Yan also always decays in time. Variance is increasing in progenitor cells, final differentiation is still robust. Predict network is robust to Yan. Yan copy number increase does increase fluorescence intensity but does not affect eye patterning mir7 mutants don’t seem to affect Yan expression. Yan and pnt coexpressed (though opposing functions) maybe post-translational modifications knock-down yan function Sven Bergmann Precision and scaling bcd and dpp Adult bodyplan set up in early development. Propose scaling coefficient under/overcompensate . dlog(x)/dlog(L) = dx/dL L/x shifts of even stripes in bcd copy number variants are less than expected (4-8% vs expected 12%) and position dependent. Sigma (x/L) as measure of precision. Lowest in mid-embryo. 1st 2 eve stripes overcompensate. Rest have almost perfect scaling. Thus position dependent but not gene dependent.
Dpp signaling requires pentagone to scale with tissue size in the wing disk.
Response to dpp signals scale well in the wing-disk. In the absense of pent scaling is lost for spalt.
Fisun Hamaratoglu dpp scaling. Starved flies smaller. Normal well fed growth requires chico (insulin signaling) does dpp gradient scale during growth? Appears to in terms of gradient and target genes. pMAD gradients are sharp in pentagon mutants. Propose dpp inhibit pentagon, pentagon activate dpp. Do the other supposed readouts of the dpp gradient move.? Omb scales in only the posterior compartment, Absence of pent L5 and the scaling on Omb lost. Spalt still scales in pentagon mutant? Range of dpp morphogen itself scales with length.
But the wing is the same except for missing L5. L2 marks splt boundary. Anil Aswani Nonparametric modeling of spatioal and temporal transcription in Drosophila. Postdoc with Claire Tomlin. Using virtual embryo data. ‘non-parametric’ model / ‘network level’ model. Fit independent linear models in each cell of the embryo. i.e. deve/dt = a_bcd*[Bcd] + a_gt[Gt] … don’t preassign sign of factor / activator repressor. Dynamic model (using changes in concentration of factors) works better than spatial correlation model. Reinitz Transcriptional control (with Ah-Ram Kim) Are enhancers physiological units or experimental abstractions? Stripe 2-3 fusions with without ‘spacer’ (actually part of st2). Doesn’t like terminology thermodynamic for transcription (out of equilibrium problem)Model components; steric competition for bound sites. Cooperative binding only for bcd. Short range repression 100-150 bp (critical). Coactivation. Hb activator of st2. Repressor of posterior str 3. Direct repression / quenching near basal promoter. Arreneus rate law activation ~ threshold like activation. Fractional occupancy enhancer models. Quenching: distance factor (based on Arnosti data), quenching strength, ‘fractional occupancy’ (requires multiple. Train model on str-2 3 with different spacers. Without further training predict location of other stripes. ‘whole locus’ model, not devided into enhancers. Can connect previous model as a Taylor expansion to first order of site occupancy model. pt. 2. Waddington: tissue types are discrete, not continuous. Mutants are variable, wt is reproducible. Developmental reactions are ‘canalized’ to bring about one definite end result. Variance of hb in Kr kni double mutants doubles. Bcd cuticle phenotype is discrete. Corresponds to saddle node bifurcation separating anterior lost
Furlong see no pattern in the motifs organization from large collection of enhancers binding 5 different TFs. Revise motifs based on ChIP data. Bin motifs alone give viseral muscle pattern, follows bin expression pattern very closely. pMAD sites 6, 4 or 2 multimerized sites follow dpp expression well. 2 base pair tin / pMAD spacing is optimal. Additional expression in the cardiac tissue. Trainable segmentation program “llastic” pMAD and tin cardiac expression shows partial pentetrance and partial expresivity. Test positional effects – lose expression in other sites by p-element integration.Screening for shadow enhancers identify potential enhancers from ChIP data (300-500 bp). 91% function as enhancers predict based on binding profiles which are likely to be shadows. 80% of activity patterns are correct. Mef2, cadN putative shadows. Use 200 sequenced isogenic strains to find structural variants ID’d 20,000 unique deletions. Median deletion size 180 bp. Consider only CRMs with at least 25% deleted. Find a Mef2 deleted shadow in this. Gasper Tkacik – physical basis of ‘positional information’ in early development with Thomas and Bill Basis for positional information. We want to understand how we go from this “pre-blastoderm embryo” to this “embryo with cephalic furrow”. (Alfonso comment: elegance of a physicist, just a stripe, not a fly). How much info is there in these spatial patterns of gene expression. Are some patterns more infomative than others (in bits). Information as a quantification of precision. Log2 number of distinguishable states in output, accessible by changing c. adventage of information: function of noise, input output shape, input distribution, scalar. Unambiguoiously correct on average. For 60 cells, need log2(6) ->5 bits of info to determine single row precision. 1.5 +/- 0.1 bits I[Bcd;Hb] about 25% of nuclei have intermediate levels of hb. Hb is a morphogen. What is the the number you get for bcd using position as input. Given noise, what is optimal input output distriution of bcd to hb → 1.7 bits. What sort of structure would the information optimal network have? Setup: c activates several downstream gap genes. Input output relations are all sigmoidal + noise. Noise: output noise due to making limited number of outputs. Regulator noise upstream (diffusion limited noise) Let C be ratio input vs output noise. If input noise is dominant all 5 gap genes should do the same thing. If output noise is dominant, gap gene response should tile the space. What if we allow cross-interaction between genes. Allow intergap gene activation, optimal solution suggests positive interaction should be zero. Allow inter-gap repression: optimal solution is to e repressed. (Walzak, tkacik Phys Rev E) globally the best gives stripes (anterior dominance expression. Current model does not allow cross-repression). Neurons encode stimuli Knowning info in all 4 gap genes 1% error. In individual gap low information where it is high. Positional error is constant along AP axis. This a property of optimally encoded positional information. Information theory more precise specification of positional information than sharpness of the boundary. Graded boundaries of gap genes are critical for specification of positional info need not be all in gap genes. Why use gap genes to position segments? Data processing inequality Role of time? Well original signal isn’t improving.
Pavel Tomancak (prev. Rubin lab) Stephan Sallfeld – stack registration of EM data using image analysis to processes high throughput screening data. Many developmental genes are localized sub-cellularly (basal, apical, components of mitotic apparatus) (maybe why sog counts are low?) in situ libraries in 2D of limited utility. 3D berkeley transcription factor data project not being used (few users cut out 2D segments) FlyFos transgenome.mpi-cbg.de Rabome database – display 3D confocal image data from GFP knockins rablibrary.mpi-cbg.de SPIM (selective plane illumination microscope). CCD camera (fast imaging), light sheet through different parts of embryo. Rotate sample in 3D on microscope. Register using fiducials in media or using object in sample (e.g. nuclei). 40s time registration from 5 viewpoints. Can therefore track the dynamic behavior of each nucleus. Project staining data from in situ database onto embryo. Trace the motion of these cells forward Open SPIM in a briefcase. With open source software. Addressing data size issue? updates Furlong use chip-seq data to predict shadows. Test. screening 200 genome project for deleted shadows. Numerous hits. e.g. Mef2. Diachete has strong segmentation phenotype, embryos highly variable. Expression of eve, runt, hairy becomes highly variable. Effect is strongest pre-cellularization. Heat shock pulse of Diachete leads to reliable but distorted eve runt hairy pattern. Ftz is weakened but reproducibly so from embryo to embryo. Claus Schertel Make miRNA library. Pool plasmids together. inject together. Sequence transformants. Redundant families of miRNAs may explain part of the no null phenotype but frequent over-expression phenotype. Express library ubiquitous, in wing and eye. Conserved miRNAs have more obvious phenotypic affects. Size match bulk injections. 100-400 at once. Around 5 rounds of injection for 180 plasmids. Joaquin de Navascues (Cambridge) Adult midgut homoestasis intestinal stem cell (ISC), enterblast (EB), differentated epithelial endocrine cell EC, and EE. 5x asymmetric divisions of ISC during larval development (?) life (?) Is it actually asymmetric division? If so, expect: stem cell lineage is immortal. Clones (descedents of a common stem cell) grow to typical patch size. Typical size reflects stem cell fraction (1:6 in this tissue). Clone size goes up to almost 18. Some clones don’t have stem cells. Sometimes both daughter cells are differentiated or are both stem cells. Follows expected distribution from birth death process. (clone fraction % vs clone size is exponential). Heat shock itself induces tissues turnover. 30 min heat shock increases turnover rate for 7 days. Resting turn over is 12days, stressed is 3 days, can recover 90% tissue loss in 2 days. Are their more symmetric, self renewed divisions . Bas van Steensel Chromatin organization in Drosophila hundreds of non-histone proteins associated with genome. Maybe 25 protein molecules per nucleosome. DamID (fuse portein of interest to DAM, any site bound by the protein will be A methylated (doesn’t normally happen in Drosophila. PCR amplify methyl specific sites. PCA suggests 5 major states of chromatin (based on 50+ TFs). Type 1 HP1 / K9me3 moderate expression (?) type 2 Pc / H3K27me3 silent type 3 (‘black’) chromatin also silent. 50% of genome. 2/3rd of silent genes in this type. Inserting reporter genes into black chromatin are most frequently silenced. (more often than in ‘Pc’ type). Su(Hw), Lam, suUR. Resembles Pc chromatin type except that PcG members are absent. But other commonly associated ones are gone. Type 4 (‘yellow’). Housekeeping genes enriched. strong K36me3. Not in red chromatin Type 5 (‘red’). Tissue specific genes enriched. No K36me2/3 enrichment. Very chromatin protein dense. Type 5 and type 1 expression incompatable.Finding new chromatin proteins. Analyze ~50 new chromatin proteins, plot PCA, see if old colors still cluster one separate branches. True for all but PcGs (no new polycomb members). Bayesian network inference on 112 mapped chromatin components. No clear pattern in insulator protein / protein types and chromatin type boundaries. To keep in mind: Physics of living matter conference in UK.