SBC seminars 2009

Robustness, the ability to maintain functionality in the presence of internal and external perturbations, is a fundamental property of biological systems. Uncovering the system level principles and architectures underlying biological robustness can be a key to understand the design principles of complex biochemical networks. In this talk I will present a control theoretic approach to robustness analysis of biochemical network models that can serve to identify mechanisms underlying a given function and its robustness. Essentially, we consider the impact of adding general type dynamic perturbations to the direct network interactions and determine the smallest distance to a perturbed network with a qualitatively different behavior. Important advantages over more traditional approaches to robustness analysis, such as parametric sensitivity analysis, is that information on the importance of specific interactions is obtained and that the potential impact of unmodelled phenomena can be accounted for. Furthermore, specific network fragilities can be identified. The focus will be on functions related to bistable switches and sustained oscillations, and I will demonstrate the method through application to circadian clocks and MAPK signaling cascades.

Constraint satisfaction problems (CSPs) are the real-world, and often very large analogs of common leisure-time pursuits such sudoku puzzles. They have been studied intensively in computer science for their theoretical interest, and in industry and applied fields for the direct pay-off from solving them well, in e.g. planning, scheduling, and electronic circuit design (and other fields).

Constraint satisfaction has many connections to combinatorial optimization, i.e. one may talk of "hard constraints" (that need to be satsified) and "soft constraints" (solving them, as well as possible, is then an optimization problem).

While (many) CSPs are provably computationally hard in worst-case, it is a well-known empirical fact that many are also typically easy in practical applications. This observation can be formalized by considering ensembles of random CSPs, and asking whether an instance from such an ensemble can be solved in e.g. linear time, with high probabability, by some incomplete algorithm (heuristic).

I will describe what are to date the most effective local heuristics on the standard benchmarks of random 3-SAT, random 4-SAT, random 5-SAT and random 6-SAT.

I will try to convey two lessons: first, that the best local heuristics (on these problems) are rather different from e.g. simulated annealing, a well known heuristic, and, second, that expected performance can depend quite sensitively on details (parameters) in the heuristics.

As a motivating example, I will use now rather old work with Mats Carlsson and others at SICS to improve the global gene measurement technology of a no longer existing Swedish biotech company. The technology is now long obsolete, but that was in fact my first contact with the field of CSPs.

This is joint work with Mikko Alava, John Ardelius, Petteri Kaski, Supriya Krishnamurthy and Pekka Orponen

Alava M. et al (2008): Circumspect descent prevails in solving random constraint satisfaction problems. Proc. Nat. Acad. Sci. USA: 105, 15253-15257

Whereas DNA microarrays have for some time proven invaluable tools for biochemical analysis, for instance in the fields of SNP analysis, evolutionary genetics and mutation screening, protein microarrays have until recently lagged behind. Most protein microarrays are used for comprehensive mapping of protein-protein interactions. However, the use of protein microarrays for disease biomarker screening and diagnosis is also coming of age. In this talk, I will present several protein microarray platforms being employed and developed at KTH/Biotechnology that may find use in such health care areas as cancer, autoimmunity, allergy, and infection. Finally, I will briefly present the usual statistical framework employed by us for basic clustering and validation.

Online seminar found at http://www.hstalks.com/main/browse_talk_info.php?talk_id=639.

This seminar will briefly summarize the work I have done so far within the PhD programme. The projects I have worked on to date have been directly or indirectly tied to improving and evaluating methods for inferring protein function, including interaction properties, by means of generally applicable, scaleable and potentially automated approaches. This includes work on the evolution of protein domain architectures, the link between protein domain architectures and protein function, and an evaluation of the relationship between functional conservation as found between orthologous proteins and evolutionary conservation of protein domain architectures. I will also report on an auxiliary project, recently submitted, which evaluates low-complexity filters in protein sequence similarity searches. Following this, I will discuss some leads I have for upcoming research.

There is a vast need to develop better methods for diagnostics and prognostics in cancer therapy. The methods used today are dependent on experienced cytologists and pathologists and are very time consuming. Hence, there is a need to find clinically applicable protein biomarkers as support for diagnosis and tumour classification. The use of multivariate methods such as PLS, where the expression of several genes or proteins are studied simultaneously has earlier shown to be powerful in biomarker discovery.

The main aim of this study is to perform multivariate meta data analysis of 2D gel electrophoresis data originating from several different studies on different cancer types. By incorporating data from various tumour types numerous clinical questions can be addressed. For example potential biomarkers specific for a certain tumour type can be identified as well as those biomarkers that are general for all malign tumour types. Results from meta-analysis on prostate cancer (n=39) and colon cancer (n=43) epithelial tumour tissue profiling are presented in this study. The datasets are matched to each other using the PDQuest software and an expression database containing the intensities for the spots in all samples was established. The further data analysis work was performed in R.

Two different ways of treating missing values were run in parallel through the analysis. Spots with a large fraction of missing data were excluded prior to analysis and the remaining missing data points were exchanged for either the mean value of the spot or the 10% lowest value for the spot. PLS-DA (Partial Least Squares Discriminant Analysis) was utilized to build predictive models and to select the most important variables for distinguishing between the classes normal and tumour. The spots were ranked by the PLS dependent VIP (Variable Importance on Projection) score and the most important variables were selected for prediction. This was repeated with decreasing number of variables and the prediction success measures were evaluated. The modelling procedure was performed in two levels of validation to ascertain a stable variable selection and model optimization, and to measure the optimized model performance. The most stable variables from a bootstrap validation were selected for the final prediction of a test set.

Despite such different tissues in the datasets, there were around 50 variables selected in at least 50% of the bootstrap rounds. This reveals some stability in the dataset and a strong signal for those variables. When applied in a PLS model to predict a held-out test set the variables yielded a rather promising prediction success (geometric mean of sensitivity and specificity was 86%). Further analysis will aim at identifying the selected proteins and validate their use as biomarkers in cancer diagnostics and therapy.

Online seminar found at http://www.hstalks.com/main/browse_talk_info.php?talk_id=475&series_id=40.

Recently published structure of Adenosine Receptor A2a complexed with antagonist ZM2413853 (PDB: 3EML) is first protein of the subfamily A16, class A GPCR which structure have been solved with atomic resolution. Thanks to Critical Assessment of GPCR Modeling and Docking Project prediction of this structure could be done on fully blind manner and undergo independent assessment. This case study have been also motivation to construct automatic homology modeling and docking pipeline designed for predictions of TM7 family complexes. Adenosine receptor share no significant homology to Rhodopsin or Adrenergic receptors which where the only GPCR structures available at that time. Structural comparison reveals major differences in binding site composition and ligand binding characteristics making structure prediction and docking of Adenosine receptor challenging task. Presented results show good overall agreement between best ranked model and crystal structure with less than 2.0Å RMSD for TM helices correct pose of the ligand in the binding pocket. The major cavity was lack of disulfide bonds in extracellular loop and overestimated coulombic contribution in total free enery of binding. Even tough, the prediction is not faultless it shows positive perspective for this prediction routine.

While gene duplication and loss are generally acknowledged as major evolutionary forces, there is a growing awareness of the importance of lateral (horizontal) gene transfers, LGTs, for gene family evolution in particularly single-celled species. However, phylogenetic characterization of families where LGT events are prevalent is notoriously cumbersome due to the inherent incongruence between a family and its species tree.

We present ongoing work on DTLRS, a novel generative model where a gene family evolves embedded in a species tree by means of duplications, losses, and more interestingly, LGTs, in a birth-death process-like manner. The model has been applied in a Bayesian MCMC framework, enabling us to perform probabilistic orthology analysis, taking the species tree into account alongside sequence evolution with a relaxed molecular clock.

This opens up the possibility of gene tree inference even in LGT-prone cases, and may provide estimates on LGT rates and counts as opposed to duplications. Some initial results on synthetically generated data, verifying the soundness of our implementation, is presented, together with a discussion of what lies ahead.

I will talk about (mainly) two papers: "Membrane insertion of marginally hydrophobic transmembrane helices depends on sequence context" and "Repositioning of transmembrane alpha-helices during membrane protein folding".

Online seminar found at: http://www.hstalks.com/main/browse_talk_info.php?talk_id=478&series_id=40&c=252.

With the rapid advances in the collection of DNA sequence data, the limitation for biological understanding of these data has increasingly become a computational problem. This is especially true for the accurate determination of phylogenetic relationships, where likelihood calculation is the main bottleneck. Meanwhile, general-purpose computing on graphics processing units (GPUs) has quickly become an important area of research, and scientific computing applications in several areas have seen notable performance speedups from adaptation to this hardware.

I will present BEAGLE, an open API and fast GPU implementations of a library for evaluating phylogenetic likelihoods of biomolecular sequence evolution. BEAGLE uses novel algorithms and methods for evaluating phylogenies under arbitrary molecular evolutionary models on GPUs, making use of the large number of processing cores to efficiently parallelize calculations even for large state-size models. The objective is to provide high performance evaluation 'services' to a wide range of phylogenetic software, both Bayesian samplers and Maximum Likelihood optimizers. Current results show a near 90-fold speed increase over an optimized CPU-based computation for estimating the phylogeny of 62 complete mitochondrial genomes of carnivores under a 60-state codon model.

Alpha-helical membrane proteins consist of transmembrane (TM) segments that traverse the membrane with the N-terminal end at alternating sides of the membrane. This division of the chain into TM and non-TM segments along with the designations of the non-TM regions as either "inside" (cytosolic) or "outside" (extracellular/inside cellular organelles) is referred to as the "topology" of the membrane protein. The prediction of topology is a useful sub-problem in the overall structure prediction of alpha-helical membrane proteins.

In this work we attempt to add to the knowledge about the translocon and its workings' effects on membrane protein evolution. We attempt to shed light on some interesting features of membrane protein duplications, how they appear and how this knowledge is transferrable to other protein sequences.

Duplications in homologues to known structures of TM proteins were found using STRUCTAL (structural superposition) and SCAMPI (sequence-based profile-HMM alignment). Different scoring schemes of these methods were tried on a test set of known structures for their ability to find duplications. SCAMPI turns out to perform slightly worse, but acceptably, and was used for finding potential duplications in proteins of a number of genomes.

Results from the genome investigation show that the most common du- plication units are of 3- and 6-TM-regions respectively and that anti-parallel orientations - N-terminal end of the units on the same side of the membrane - of duplications in homologues are surprisingly common. In addition, a particu- larly interesting example, the acriflavine resistance protein (PDB 1oyeA), shows a distinct set of homologues of 3 and 6 TM-regions, which hints at a duplication and subsequent fusing in the distant past.

The abundance of messenger RNA molecules in the cell reflects a balance between transcription and degradation. What are the relationships between these two opposing forces? What could be the consequences downstream (in translation) if the balance is subject to heavy fluctuations? In this talk I will survey results from several projects done at the Pilpel Lab that shed some light on these questions. Specifically, I will talk about our experimental work in yeast that uncovered interesting coupling between mRNA degradation and transcription in response to stress. Next, relationships between the activity of transcription factors and microRNAs, that were found in a binformatics study in human, will be presented. Finally, a theoretical study of a small regulatory network will illustrate how fluctuations in mRNA abundance may significantly affect the kinetics of such a network.

Experimental methods have made great progress recently but they are still inadequate to achieve a complete mechanistic understanding of complex diseases. Computational methods can be useful to integrate data from different levels and formulate distinct complex biological processes, but they need yet to overcome three major and related challenges. The first is to move beyond a pure molecular scale and to incorporate several vertical scales in the models. Secondly, it is necessary to assure the robustness of the model against minor changes in parameters, in order to define models consistent with biological systems. Finally, it is necessary to develop clinically relevant models.

We develop a "clinically relevant" computational approach to atherogenesis in the arterial wall based on a multi-scale model. The term "atherosclerosis" was introduced by Marchand to describe the association of fatty degeneration and vessel stiffening (see [1] and [2]). Atherosclerosis is characterized by the accumulation of lipids and fibrous elements in the large arteries; and atherogenesis consists of sub-endothelial accumulations of cholesterol-engorged macrophages, called "foam cells". Atherosclerosis is considered a progressive disease whose most devastating consequences, caused by superimposed thrombosis, are heart attack and stroke (see [3]). Biological experiments reveal a critical period where the disease accelerates rapidly (see [4]). Moreover, experiments reveal that lowering of the lipids is effective in reducing the disease progression provided that such a lipid lowering is delivered before the disease onset but not after. It is yet not clear why there is such a rapid switch in the disease development and why there is a time-dependent effect of lipid lowering.

Our model, named A-model, represents some of the key players (molecules, different cell-types, blood) in the disease development active in the arterial wall. A-model is flexible in terms of the quantitative interactions between the components in the arterial wall. We are interested in those sets of parameters that are consistent consistent with observed experimental data in a mouse model prone for atherosclerosis (see [4]).

In order to search among the parameter space, we use the Particle Swarm Optimization algorithm (PSO) and simulation power. As described by Kennedy and Eberhart (see [5] and [6]), PSO is an adaptive algorithm based on a social environment where a set of particles, called population, are visiting possible "positions" (in our case, sets of parameters) of a given dominion. Each position has a fitness value (and it can be computed). At each iteration particles will move returning stochastically toward the population's best fitness position and its own previous best fitness position. Particles of the population are sharing information of the best areas to search. Fitness function is defined using experimental data and knowledge from experts in the field.

Despite the high-dimensionality of the parameter space and the potential combinatorial complexity in the number of solutions, we find that from SET there are a surprisingly small number of classes of solutions that are consistent with the experimental data; we also find that each class represents an isolated area in the parameter space. The reason is that there are strong implicit correlations within the model which appear to highly constrain the possible behaviors in the system. The small number of possible solutions provided us with an explanation as to why there is a rapid switch in the disease development. Since the model spans several vertical layers we could use available gene expression data for experimental validation at the gene level.

[1] Aschoff L., (1933), Introduction. In: Arteriosclerosis: A survey of problem. Cowdry E.V. (Ed). Macmillan. New York.

[2] Crowther M.A., (2005), Pathogenesis of atherosclerosis. Hematology. Am. Soc. Hematol. Educ. Program. 436-441.

[3] Falk E., (2006), Pathogenesis of Atherosclerosis J. Am. Coll. Cardiol. Vol. 47, N. 8, C7-C12.

[4] Skogsberg J, Lundström J, Kovacs A, Nilsson R, Noori P, et al., (2008), Transcriptional Profiling Uncovers a Network of Cholesterol-Responsive Atherosclerosis Target Genes, PLoS Genetics 4(3): e1000036 doi:10.1371/journal.pgen.1000036.

[5] Kennedy, J.,Eberhart, R.C., (1995), Particle Swarm Optimization, Proc. IEEE International Conference on Neural Networks, Piscataway, NJ, USA, pp. 1942-1948.

[6] Kennedy, J., Eberhart, R.C., (2001), Swarm Intelligence. Morgan Kaufmann Publishers.

[7] Tuomisto T.T., Binder B.R., Yla-Herttuala S. (2005) Genetics, genomics and proteomics in atherosclerosis research. Ann Med 37: 323-332.

Accumulating evidence from tumor biology studies emphasize the importance of the tumor microenvironment for tumor growth, progression and drug response. The cancer-associated fibroblast (CAF) is a major cell type of the tumor stroma. Novel pro-tumorigenic effects of CAFs include recruitment of bone-marrow-derived endothelial precursor cell and pro-metastatic effects. Different types of marker studies indicate the existence of different CAF-subsets, and translational studies have identified prognostic, and response-predicative, significance of different CAF-derived markers. Clinical studies aiming at CAF-targeting are envisioned based on findings from experimental intervention studies with agents targeting e.g. FAP, PDGF-, TGF-β- or hedgehog-signaling.

Recent studies in our own group have demonstrated prognostic, and response-predicative, roles of stromal PDGF receptors in breast cancer. These findings support our earlier animal studies which demonstrated that stromal PDGF receptors control tumor drug uptake and thereby affects the therapeutic efficacy of systemic cancer treatments. Ongoing studies show that PDGF-dependent paracrine signaling from fibroblasts also strongly influence the migration, proliferation and drug-sensitivity of co-cultured cancer cells. Molecular identification of these PDGF-induced paracrine signaling molecules are ongoing. Through analyses of prostate cancer fibroblasts we have also identified two novel secreted CAF-derived proteins, CXCL14 and GDF15, which stimulate prostate cancer growth. Finally, studies on lung cancer CAFs have identified the transcription factor FoxF1 as a clinically relevant inducer of tumor-stimulating CAF phenotypes.

Selected References:

Heldin, C-H., Rubin, K., Pietras, K., and Östman A. High interstitial fluid pressure - an obstacle in cancer therapy. Nat Rev Cancer 4, 806-813 (2004).

Östman, A. and Heldin, C-H. PDGF receptors as targets in tumor treatment. Adv. Cancer Res. 97, 247-274 (2007).

Östman, A. and Augsten, M. Cancer-associated fibroblasts - bystanders turning into key players. Curr Op Gen Dev 19, 67-73 (2009).

Baranowska-Kortylewicz, J, Abe, M., Kurizaki, T., Pietras, K., Kortylewicz, Z.P., Nearman, J., Mosley, R.L., Enke, C.L., and Östman, A. Impact of PDGFr-b inhibition with STI571 on radioimmunotherapy. Cancer Research 65, 7824-7831 (2005).

Paulsson, J., Sjöblom, T., Micke, P., Pontén, F., Landberg, G., Heldin, C.H., Bergh, Brennan, D.K., Jirström, K. and Östman, A. Prognostic significance of stromal PDGF β-receptor expression in human breast cancer. Am. J. Path. 175, 334-341 (2009).

Augsten, M., Hägglöf, C., Olsson, E., Stolz, C., Panagiotis, T., Levchenko, T., Frederick, M.J., Borg, Å., Micke, P., Egevad, L. and Östman A. CXCL14 is an autocrine growth factor for fibroblasts and acts as a multi-modal stimulator of prostate tumor growth. Proc. Natl. Acad. Sci. USA 106, 3414-3419 (2009).

We present a novel method for deriving transcriptional network models from molecular profiles of human cancer tumors. The network models aim to (i) detect novel oncogenes and tumor suppressor genes, and (ii) to explain how such genes exert their function by regulating downstream targets.

Mathematically, we represent a tumor by a vector of transcript levels, a vector of gene dosage alterations and a matrix that describes the regulatory interaction between transcripts. When the system evolves in time, the rates of transcript synthesis and decay are controlled both by the tumor's alterations in gene dosage and by the regulatory interactions between genes. For a particular set of patients, we derive network structure and model parameters using an efficient algorithm based on convex optimization, that aims to achieve an optimal trade-off between data fit and biologically defined constraints.

To evaluate the predictive potential of the method, we analyzed DNA copy number alterations and mRNA expression profiles from 186 glioblastoma tumors from the Cancer Genome Atlas study. The derived network model rediscovered known oncogenes and contained interesting predictions, including two candidate tumor suppressors. We conducted experiments in glioblastoma cell lines to test the model's prediction, validating one of the candidate tumor suppressor and confirming its effect on a set of downstream targets.

Possible applications of our method include the discovery of regulatory interactions in several forms human cancer, and the selection of candidate therapeutic targets. Our method can be a helpful tool to meet the analytical challenges of the ongoing human cancer genome programs.

Organ specific growth control remains one of the major questions in developmental biology that has not been resolved; it is not understood what determines organ size and shape (reviewed in Lecuit and Le Goff, Nature 450:189, 2007). It is also not clear why tumors arising in different tissues harbor different oncogenic mutations (Taipale and Beachy, Nature 411:349, 2001). Much of what we know about physiological mechanisms controlling cellular growth in mammals has been revealed by human cancer genetics. These studies have revealed that a large number of genes can contribute to aberrant cell growth; there are more than 300 genes that have been linked to cancer, and mutations found in cancer are often cell type specific. For example, PTCH mutations are common in medulloblastoma, APC in colon cancer, and TMPRSS2-ERG in prostate cancer, suggesting that different pathways in different cell lineages are coupled to the cell cycle machinery. Our hypothesis is that the problems of organ-specific growth control and specificity of oncogenes to particular tumors represent two sides of the same coin; that is, mutations in tumors are tissue specific, because tumors arise by the most economical mutagenic route, aberrantly activating the organ-specific growth mechanisms.

To test this hypothesis, we have developed computational and experimental methods to identify direct target genes of oncogenic transcription factors, and used high-throughput RNAi screening to identify genes required for cell cycle progression. Combining these two sets of data allows identification of specific regulatory elements which drive growth in particular tissues and tumor types. Preliminary evidence suggests that Hedgehog (Hh) and Wnt signals appear to be directly coupled to expression of N-myc and c-Myc genes, but only in tissues and cell-types that display a proliferative response to these factors.

As the genomic era has been ceaselessly advancing, more genomes have been completely sequenced and more diseases have been identified in the Online Mendelian Inheritance in Man database (OMIM). To further the understanding of human disease and to develop new, effective medicines and therapies, there is a need for tools to extract biologically meaningful information if we are to keep pace with emerging diseases. Many genes in other organisms share function with human genes, particularly if they are orthologs, that is, if they are descended from the same genes in the last common ancestral species. Orthology is an indispensable key for transferring biological knowledge between species, from protein annotations to sophisticated disease models, especially for human diseases. However, orthology assignment is not trivial. The quality of ortholog clusters is hence essential for the appropriate selection of model organisms when designing drugs and for predicting human responses to treatments. We therefore present an updated version of OrthoDisease, a disease gene orthologs database. OrthoDisease offers a more complete evolutionary picture of disease genetics, allowing researchers to analyse orthologs to new disease genes, and to select suitable model systems for particular diseases. OrthoDisease provides disease ortholog clusters between Homo sapiens and 99 other organisms and offers four search features: 1. Disease search, 2. Gene search, 3. OMIM search, and 4. Text search. In addition, each search feature allows users to restrict the search to a species of interest. The database is accessible online at http://orthodisease.sbc.su.se.

Mass spectrometry based proteomics is currently the most accurate and efficient platform for analyzing protein content in biological mixtures. The techniques are in general computationally intensive, as the amount of data such assays generate is massive. Currently there is no consensus in how to perform and evaluate the outcome of the data processing. Here I will give a tutorial on the different techniques to identify peptides and proteins from mass spectrometry data and also give some insight into how the performance of such techniques may be improved by the usage of machine learning.

The mechanism by which multi-spanning helix-bundle membrane proteins are inserted into their target membrane remains a major unsolved problem in molecular cell biology. In both prokaryotic and eukaryotic cells, membrane proteins are inserted co-translationally into the lipid bilayer. Positively charged residues flanking the transmembrane helices are important topological determinants, but it is unclear if they act locally, affecting only the nearest transmembrane helices, or can act globally, affecting the topology of the entire protein. Here we show that the topology of an Escherichia coli inner membrane protein with 4 or 5 transmembrane helices can be controlled by a single positively charged residue placed in different locations throughout the protein, including the very C-terminus. This observation points to an unanticipated plasticity in the insertion mechanism.