1. 23. Lecture WS 2007/08Bioinformatics III1 V23 Transcriptional Control in Halobacter salinarum Leroy HoodNitin Baliga

2. 23. Lecture WS 2007/08Bioinformatics III2 What is the goal? Important goal of systems biology: understand how a simple genetic change or environmental perturbation influences the behavior of an organism at the molecular level and ultimately its phenotype. High-throughput technologies to interrogate the transcriptome, proteome, protein- protein, protein-DNA interactions etc present a powerful toolkit to accomplish this goal. However, each of these individual data types captures an incomplete picture of global cellular dynamics. Therefore, these data need to be integrated appropriately to formulate a model that can quantitatively predict how the environment interacts with cellular networks to effect changes in behavior. Ultimate test of our understanding of a given system that will enable re-engineering of cellular circuits: accurate prediction of its quantitative behavior. Here: integrate experimental and computational approaches to construct a predictive gene regulatory network model covering 80% of the transcriptome of Halobacterium salinarum NRC-1, a free-living cell.

3. 23. Lecture WS 2007/08Bioinformatics III3 Haliobacter salinarium NRC-1 H. salinarum NRC-1 belongs to Archaea and lives at an environment of 4.5 M salt. Therefore, it provides a unique window into molecular mechanisms underlying fascinating response physiologies in extreme environments such as above boiling temperatures and in deep sea ocean vents. Specifically, it can be expected to provide insights into evolutionary adaptation for survival in high-salinity-induced low-water activity, which precludes growth of most organisms. Like most organisms it is also subject to daily and seasonal changes in many environmental factors (EFs). One can expect it to have regulatory circuits that effectively negotiate these complex and often stressful conditions. From a practical standpoint, all these physiological capabilities are encoded in ca. 2400 nonredundant genes in a very compact and easily manipulable 2.6 Mbp genome. Task: discover and characterize a significant fraction of the gene regulatory network associated with the intercoordination of physiological processes in this organism in differing environmental and genetic backgrounds.

4. 23. Lecture WS 2007/08Bioinformatics III4 Halobacterium salinarum Pictures borrowed from http://www.biochem.mpg.de/en/rd/oesterhelt/web_page_list/Org_Hasal/index.html Halobacterium salinarum in its natural environment. The picture shows a salty pond in the Arabian desert, which is colored red due to the presence of Halobacterium salinarum. Massive growth of Halobacterium salinarum in a saline. An electron microscopic image of Halobacterium salinarum with ca 13.500-fold magnification. From the pole of the rod-shaped cell body extends the long flagellar bundle. Interesting model system to study chemotaxis. Dieter Oesterhelt, MPI Martinsried

5. 23. Lecture WS 2007/08Bioinformatics III5 Integrated approach Approach: - perturb the cells (genetically or environmentally), - characterize their growth and/or survival phenotype, - quantitatively measure steady-state and dynamic changes in mRNAs, - assimilate these changes into a network model that can recapitulate all observations, and, - finally, experimentally validate hypotheses formulated from the model. Realization: This approach required the integrated development and implementation of computational and experimental technologies and consisted of the following steps:

6. 23. Lecture WS 2007/08Bioinformatics III6 Integrated approach 1 Sequence the genome and assign functions to genes using protein sequence and structural similarities. 2 Perturb cells by changing relative concentrations of EFs and/or gene knockouts. 3 Measure the resulting dynamic and/or steady-state transcriptional changes in all genes using microarrays. 4 Integrate diverse data (mRNA levels, evolutionarily conserved associations among proteins, metabolic pathways, cis-regulatory motifs, etc.) with the cMonkey algorithm to reduce data complexity and identify subsets of genes that are coregulated in certain environments (biclusters). 5 Using the machine learning algorithm Inferelator construct a dynamic network model for influence of changes in EFs and TFs on the expression of coregulated genes. 6 Explore the network with Gaggle, a framework for data integration and software interoperability, to formulate and then experimentally test hypotheses to drive additional iterations of steps 2–6.

7. 23. Lecture WS 2007/08Bioinformatics III7 Predictive Modeling of Cellular Responses Subsequent to genome sequencing there were two major interconnected and iterative components: experimentation and computation followed by data visualization and analyses. Within the first component the major efforts included computational genomic analyses for discovering functional associations among proteins (black boxes); putative functional assignment to proteins using sequence- and structure-based methods (blue boxes); and high-throughput microarray, proteomic, and ChIP-chip assays on genetically and/or environmentally perturbed strains (red boxes). All data (with the exception of proteomic and ChIP-chip data) from these approaches along with associated records of experiment design (green boxes) were analyzed with regulatory network inference algorithms (purple box). The resulting EGRIN was explored along with underlying raw data using software visualization tools within Gaggle (yellow box), which enables seamless software interoperability and database integration. Gaggle also provides a cost-effective interface to third party tools and databases. This manual exploration and analysis enabled hypothesis formulation and provided feedback for additional iterations of systems analyses. Cell 131, 1354 (2007)

8. 23. Lecture WS 2007/08Bioinformatics III8 Halobacterium functional association network 6818 associations among proteins were inferred by comparative proteomics. 1. Domain fusion. Two or more genes that are individually translated in one genome but translated as a single fused protein in other genomes are predicted to functionally and physically interact in the former (Rosetta stone method). No. of domain fusion edges in the NRC-1 genome: 2460. 2. Phylogenetic pattern. This type of interaction is based on the premise that similar profiles of presence or absence of pairs of orthologs in fully sequenced genomes is often indicative of their close functional relationship. No. of phylogenetic pattern edges in the NRC-1 genome: 525. 3. Chromosomal proximity. Pairs of orthologs with evolutionarily conserved chromosomal proximity in multiple genomes are hypothesized to be maintained as such to retain functional association through evolution and events such as lateral gene transfer. No. of chromosomal proximity edges in the NRC-1 genome: 327.

9. 23. Lecture WS 2007/08Bioinformatics III9 Halobacterium functional association network 4. Yeast and H.pylori two-hybrid interactions. We have employed COGs (clusters of orthologous genes as the unit of protein structure and interaction). 1431 yeast and 178 H. pylori interaction were mapped onto halobacterial proteins. 5. SCOP interactions. Proteins in Halobacterium sp. were mapped to structural superfamilies (SCOP) via homology modeling. Protein pairs belonging to structural superfamilies known to interact are also likely to interact in Halobacterium sp. No. of SCOP interactions mapped onto halobacterial proteins: 562. 6. Operons. We have predicted Halobacterium sp. operons by analyzing chromosomal proximity on its genome alone (for genes without many orthologs in other organisms) and the conservation of chromosomal proximity across multiple genomes (for genes with sufficient numbers of orthologs). No. of operon edges in the Halobacterium NRC-1 genome: 1335. The modular architecture of the Halobacterium sp. network deciphered through hierarchical clustering of genes based on their shortest network paths to every gene in the network correlated well with gene functions with the modules often coinciding with sequential biochemical steps in metabolic pathways.

10. 23. Lecture WS 2007/08Bioinformatics III10 Assign protein functions Cell 131, 1354 (2007) 1. Sequence the genome and assign functions to genes using protein sequence and structural similarities. Using primary sequence similarity of the H. salinarum proteins to characterized orthologs in other organisms left a significant fraction (38%) of ~2,400 putative protein-coding genes that could not be assigned any function. To overcome this hurdle, we applied a more sensitive approach that incorporated functional relationships among proteins from comparative genomics as well as protein structure predictions to detect similarities at 3D level to proteins and protein domains in the protein data bank (PDB). This resulted in a comprehensive parts list for which nearly 90% of all predicted genes had some meaningful association with either a characterized protein, a protein family or a structural fold. Importantly, this re-annotation provided several putative regulators for designing targeted perturbations, as well as for use as key input parameters for regulatory network inference in subsequent steps. Specifically, through analysis of protein family signature or predicted structural matches we were able to catalogue a list of 130 putative TFs, of which at least 14 are general transcription factors (six TATA-binding proteins (TBPs), seven Transcription Factor B (TFB), and Transcription Factor E alpha-subunit orthologs), and the remainder have matches to sequence-specific DNA-binding proteins.

11. 23. Lecture WS 2007/08Bioinformatics III11 Annotating protein structures Flow chart for annotation. Sequence based methods are employed first (top), domains that elude primary sequence based methods are predicted by structure-prediction methods (bottom). For any given genome, data from all levels in this method hierarchy are integrated using SBEAMS (Systems Biology Experiment Analysis and Management System). Implicit in this annotation hierarchy is the idea that multi-domain proteins should be divided into domains as early as possible in the annotation process. Bonneau et al. Genome Biology 2004 5:R52

12. 23. Lecture WS 2007/08Bioinformatics III12 Collect transcriptional responses Collectively analyze transcriptional responses to individual and combinatorial perturbations in - 10 EFs including light, oxygen, UV radiation, gamma radiation, manganese (Mn), iron (Fe), cobalt (Co), nickel (Ni), copper (Cu), and zinc (Zn) and - 32 genes including TFs, signal transducers, and metabolic enzymes. Analyzing the microarray data classified 1929 of the total 2400 predicted genes into 300 biclusters that were often highly enriched in genes with known metabolic processes. Each of these biclusters represents a subset of genes that are potentially coregulated in a defined set of environmental conditions. We then constructed subcircuits that model expression changes in each of these biclusters as a function of corresponding changes in 72 TFs and 9 EFs (although Co was included as a potential predictor it did not make it into the final network).

13. 23. Lecture WS 2007/08Bioinformatics III13 Genetic perturbations (knockouts) used Cell 131, 1354 (2007)

14. 23. Lecture WS 2007/08Bioinformatics III14 Distribution of fluxes in E.coli Cell 131, 1354 (2007) Figure S1. Profiles of mRNA level changes for genes in operons provide metric of quality of microarray data. A significant fraction of genes in prokaryotic genomes are organized into operons of two or more genes that are each co-transcribed into a polycistronic transcript. This organization offers a unique opportunity to test the quality of microarray data by comparing profiles of change for the same polycistronic transcript as measured by different probes – each unique to a different gene of the operon. In the H. salinarum NRC-1 genome 2,141 genes are in operons with 1-25 other genes. We highlight the high degree of transcriptional coherrence among these co-trancribed genes over >400 experiments by showing profiles for 11 different operons encoded at different loci around the genome.

15. 23. Lecture WS 2007/08Bioinformatics III15 Co-clustering - cMonkey Cell 131, 1354 (2007) (B) A plot of transcriptional changes of genes in bc208 demonstrates their co-expression over ~160 conditions included in the bicluster (conditions to the left of the vertical red dashed line). (C) Three motifs were detected by cMonkey of which two had high statistical significance (see ST3 for details). A sequence logo 45 derived from the position specific scoring matrix (PSSM) of the motif that corresponds to the experimentally characterized UAS is shown. (D) The colored boxes (red = UAS) indicate the relative positioning of the three motifs detected upstream to all genes in bc208. Regulators included in the bicluster are indicated in red font (VNG1464G = Bat).

16. 23. Lecture WS 2007/08Bioinformatics III16 Data integration Cell 131, 1354 (2007) Figure S2. Data integration using cMonkey extends knowledge of known processes through the discovery of new metabolic and regulatory relationshsips. (A) In bc208 containing a total of 29 genes, 7 are characterized phototrophy-associated genes with five known to constitute a regulon 2,41; 5 genes are characterized 42 DMSO respiration genes; 22 are organized into eight operons; two pairs of genes have similar phylogenetic 19 profiles 43; 7 are interconnected by their evolutionarily conserved chromosomal proximity 34 into three groups; two are connected by gene fusion 33; and none were interconnected by the KEGG metabolic network. Some genes were included purely on basis of their coexpression with other bicluster genes. The computational detection of the conserved UAS cis-regulatory motif 36,44 upstream to seven genes in addition to the five known phototrophy genes in bc208 has also helped extend membership of the this regulon.

17. 23. Lecture WS 2007/08Bioinformatics III17 Inferelator algorithm for biclustering Cell 131, 1354 (2007) 5. Use the machine learning algorithm Inferelator to discover the dynamic influences of EFs and TFs on the expression of co-regulated genes within biclusters. Next, using the Inferelator algorithm, we discovered instances wherein individual or combinatorial changes in the concentrations of certain TFs1 and/or EFs (archived in the metainformation from step 3) temporally preceded average transcriptional changes within a given bicluster or a gene. Briefly, the Inferelator (a) selects parsimonious models (i.e. minimum number of regulatory influences for each bicluster) that are predictive; (b) explicitly includes the time dimension to discover causal influences; and (c) models combinatorial logic i.e. interactions between EFs and TFs and between pairs of TFs. All of these represent reasonable assumptions about how biological networks are constructed and operate and thus yield models that are more likely to encapsulate true biological properties. In this specific case, 72 TFs and 10 EFs were used as predictors (components that influence the expression of others).

18. 23. Lecture WS 2007/08Bioinformatics III18 Distribution of fluxes in E.coli Cell 131, 1354 (2007)

19. 23. Lecture WS 2007/08Bioinformatics III19 Distribution of fluxes in E.coli Cell 131, 1354 (2007)

20. 23. Lecture WS 2007/08Bioinformatics III20 Distribution of fluxes in E.coli Cell 131, 1354 (2007)

21. 23. Lecture WS 2007/08Bioinformatics III21 EGRIN: environment and gene regulatory influence model The resulting model is a set of equations that can take as input measured changes in a few TFs and/or EFs to predict kinetic and steady-state transcriptional changes in 80% of genes in this organism with an average (Pearson) correlation of 0.8 to their actual measured changes. Importantly, this predictive capability reduces significantly when the time component is removed from the model, strongly suggesting that a significant fraction of the influences have causal properties. Although we provide evidence that some of the regulatory influences are mediated directly via TF-DNA interactions, we expect that a large fraction, especially EF influences, act indirectly, for example, via interactions with signal-transducing environmental sensors.

22. 23. Lecture WS 2007/08Bioinformatics III22 Statistically learned gene regulatory influences Cell 131, 1354 (2007)

23. 23. Lecture WS 2007/08Bioinformatics III23 Examples of biclusters Cell 131, 1354 (2007)

24. 23. Lecture WS 2007/08Bioinformatics III24 Cell 131, 1354 (2007)

25. 23. Lecture WS 2007/08Bioinformatics III25 EGRIN predicts novel regulatory influences for known biological processes Bicluster bc66 contains 34 genes including cytochrome oxidase, ribosomal proteins, and RNA polymerase. It turns out that their transcriptional behavior is nearly perfectly modelled by corresponding changes of 2 EFs (oxygen and light) and 2 TFs (Cspd1 and TFBf). The influences from TFBf and light act through an AND logic gate (triangle). Cell 131, 1354 (2007)

26. 23. Lecture WS 2007/08Bioinformatics III26 mRNA profile of bc66 Cell 131, 1354 (2007) (B) The mRNA profile of bc66 recreated by the combined TFs and environmental influences is nearly identical to the actual (averaged) mRNA levels over 398 experiments.

27. 23. Lecture WS 2007/08Bioinformatics III27 Bc66 transcription at varying oxygen levels Cell 131, 1354 (2007) (C) The transcript levels of genes in bc66 changes proportionally with changes in oxygen tension in controlled experiments. The profile represents average transcription level changes of genes in bc66. The error bars indicate the standard deviation among mRNA level changes of genes in bc66.

28. 23. Lecture WS 2007/08Bioinformatics III28 ChlP-Chip experiments to detect protein-DNA interactions (D) Crosscorrelation of predicted influences in EGRIN with physically mapped binding sites suggests that the TFBf influence may be directly effected via binding of this GTF to promoters of 24 out of genes (and operons) in bc66 (p < 10 −10 ). This suggests that TFBf directly influences the expression of these genes. Cell 131, 1354 (2007)

29. 23. Lecture WS 2007/08Bioinformatics III29 How are connected cellular processes regulated? (A) Components of pyruvate metabolism, ATP synthesis, glutamate-glutamine metabolism, and accessory processes for transport of raw materials and synthesis of cofactors are distributed across 9 biclusters (boxes) containing altogether 162 genes. The expression of genes in these 9 biclusters is modeled by gene-regulatory influences (red: activate, green: repress, black: possible autoregulators coclustered with the regulated genes) from 27 TFs (circular nodes) that operate individually or in combination through AND gates (connected by blue edges). The assembly of the regulatory influence subcircuits for all biclusters into the complete EGRIN has reconstructed known relationships among cellular processes that are connected in metabolic networks and play complementary roles. More importantly, based on the confidence gained from recapitulating these known relationships, we can investigate the architecture of EGRIN to discover new experimentally testable relationships. We illustrate this point by selecting genes distributed across 9 biclusters (bc20, bc28, bc45, bc48, bc61, bc75, bc76, bc163, and bc174) that bring together components of pyruvate metabolism, glutamate-glutamine metabolism, and ATP synthesis as well as some accessory functions required for enzyme cofactor biosynthesis and raw material transport to support these metabolic processes. The predicted subnetwork controlling these biclusters is presented above.

30. 23. Lecture WS 2007/08Bioinformatics III30 Reconstruction of metabolic pathways (B) Metabolic pathways were reconstructed on the basis of known and putative functions of genes in the 9 biclusters. Memberships of various enzymes or enzyme subunits in each of the 9 biclusters in (A) are indicated with color-coded bars next to each step in the metabolic pathway (see key in panel A for interpreting this color code).

31. 23. Lecture WS 2007/08Bioinformatics III31 Relationships among biclusters (C) The dendogram represents relationships among the 9 biclusters based on the similarities among the averaged expression profiles of their member genes. The differences in how the biclusters (cellular processes) relate to one another in varying environments are illustrated by highlighting relationships between two bicluster groups: I (bc20, bc28, bc48) and II (bc76 and bc163).

32. 23. Lecture WS 2007/08Bioinformatics III32 Temporal changes of the network architecture (D) The incorporation of weighted regulatory influences with an associated time constant into EGRIN enables the architecture of the network to change with the environment. As a consequence of this, despite environment- specified differences in relationships among cellular processes (C) the same set of regulatory influences acting on each bicluster accurately models the averaged transcriptional changes of its constituent genes even for responses to new EF perturbations (for example, responses to EMS and H 2 O 2 ). Each of the nine graphs shows profiles of predicted versus measured transcript level changes in each individual bicluster in environmental responses that were part of the training set as well as 147 completely new experiments.

33. 23. Lecture WS 2007/08Bioinformatics III33 Correlation of predicted and measured mRNA levels Histogram of Pearson correlations of predicted and measured mRNA levels of individual biclusters over the 266 experiments in the training set (A) and the 131 newly collected experiments (B) are shown. (C) shows a comparison of correlations between predicted and measured mRNA levels for all 300 biclusters in training set and new data. (D) Transcription of the broad specificity metal ion efflux pump ZntA is upregulated under Cu stress in the ΔVNG1179C strain background in which the primary efflux pump is transcriptionally inactivated (Δura3 is the parent strain in which knockouts are constructed). This altered transcriptional response of ZntA to Cu was accurately modeled by the regulatory influences on bc189, which contains this gene along with 7 other genes. 147 new experiments: (1) New combinatorial perturbations of EFs already in training set (2) New EF perturbations: oxidative stress agent hydrogen peroxide, chemical mutagen ethyl methyl sulfonate (3) New combinations of TF and EF perturbations.

34. 23. Lecture WS 2007/08Bioinformatics III34 Relative abundance of 5 Na + /H + antiporters Cell 131, 1354 (2007) To withstand high salinity, H. salinarum maintains a high 4M K + / 1M Na + content in its cytoplasm which is in inverse proportion to the high 2.7 mM K + / 4.3 M Na + content in its environment.  To maintain this gradient, the genome encodes at least 5 Na + /H + antiporters. Which one is most important? ChlP-chip data to map protein-DNA binding suggests that NhaC3 is under the direct control of 5 different TFBs.

35. 23. Lecture WS 2007/08Bioinformatics III35 However, (B) according to cMonkey nhaC3 is coregulated with genes in five biclusters within EGRIN (bc2, bc3, bc12, bc16, bc50, and bc113). The average expression changes of genes in four of these biclusters are modeled by corresponding changes in TFBg transcript levels; the circuit diagram shows the Inferelator model for one of these biclusters (bc113). (C) nhaC3 transcript levels during different phases of growth in five strains, each carrying a plasmid-borne copy of the respective cmyc-tagged tfb gene.  Only the deletion of the TFBg gene results in a significant suppression of nhaC3 expression. Perturbing TFBg alters regulation of Na/H antiporter

36. 23. Lecture WS 2007/08Bioinformatics III36 Conclusions Our choice of H. salinarum NRC-1 has helped highlight the power of a systems approach for rapidly discovering new biology in largely uncharacterized organisms. By observing the consequences of systematically perturbing this organism with both genetic and environmental perturbations we were able to construct statistically significant and meaningful associations among most genes encoded in the genome of this organism. However, transcriptional control of 20% of all genes is not represented within the biclusters in the EGRIN model. While this could be due to technical limitations in measuring transcript level changes of these genes, or absence of their differential regulation in response to perturbations used in our studies, an important point to consider is that our model does not yet account for a plethora of regulatory mechanisms such as epigenetic modifications, small RNAs, posttranslational protein modifications, and metabolite-based feedback. The challenges associated with investigating these important control mechanisms at a global level are now being overcome through technological innovations. Our approach to regulatory network inference is extensible to incorporate these new data types and model their associated control mechanisms to eventually completely model the entire regulatory circuit in this archaeon.

37. 23. Lecture WS 2007/08Bioinformatics III37 Conclusions It took < 6 years to move from genome sequence to this level of understanding for a relatively poorly studied organism. It would now be significantly quicker to implement the same approach with a newly sequenced organism given that much of the scientific methods including experimental procedures, algorithms, and software have been delineated through our study. Will the potential for enormous complexity of a biological system will ever allow the construction of a complete model of a cell? In this regard it has been favorably suggested, at least in the context of metabolism, that despite this potential for complexity, a cell usually functions in one of few dominant modes or states. We speculate that this natural property of a biological system simplifies the problem to inferring gene regulatory models for its transitions among relatively few states. In addition, as discussed earlier, the extensive connectivity within EF and biological networks makes it tractable to effectively construct a comprehensive model of cellular responses to changes in multiple EFs from a modest number of well-designed systematic perturbation experiments. We believe that this type of a model will hold true for environmental responses of all organisms and, more importantly, that it should be possible to construct such models solely from EF perturbation experiments. This will be especially valuable in context of organisms that currently lack tools for genetic analysis.