In a very recent lecture (see full video from NIH VideoCasting) given for the NIH Systems Biology Special Interest Group, Trey Ideker presents a great overview of the various strategies his group has been developing in the recent years in order to integrate multiple types of large scale datasets. While one of the most pervasive 'meme' about high-throughput measurement is that they are "notoriously unreliable" (see Hakes et al, 2008, for a recent example), Trey beautifully illustrates how predictive computational models and novel biological insights can be generated by sophisticated data integration strategies. Three types of applications are presented in his talk:

1. mapping of transcriptional response pathways
2. functional mapping of protein complexes
3. disease diagnosis and stratification

In the last section, Trey presents the study recently published in Molecular Systems Biology (Chuang et al, 2007, video: 00hr:39min:15sec) where the information provided by microarray expression profiling is superposed to a protein-protein physical interaction network to identify 'subnetwork' biomarkers that classify metastatic vs non-metastatic breast tumors.

Research highlight by Charles Auffray, Functional Genomics and Systems Biology for Health, UMR7091, CNRS and Pierre & Marie Curie University—Paris VI, Villejuif, France

The work presented by Helikar et al. (2008) in a paper recently published in the PNAS represents a promising new step in the development of computational cellular physiology in eukaryotes. From curated cellular and biochemical data available in the literature, the authors have assembled a discrete Boolean model of signal transduction comprising 130 nodes, and examined in a systematic and controlled manner how varying combinations of external inputs translate into a range of cellular responses. The qualitative model is not only able to reproduce known input-output relationships representative of major transduction pathways, but it also provides evidence in support of the emergence of information-processing functions from the complex cellular network of molecular interactions. This is strikingly demonstrated by the fact that a large sample of randomly selected input combinations result in a very limited fraction of the possible outputs, which correspond to well-characterized global biological responses, a result which is obtained irrespective of the level of noise introduced in the inputs of the model. Moreover, similar input combinations are neatly clustered by the model into equivalence classes of global outputs, reflecting the ability of the cell to integrate complex environmental signals and translate them into robust specific responses and behaviours through common intracellular pathways. While discrete Boolean modelling makes it possible to highlight emergent properties of transduction networks, overcoming the hurdle of parameter estimation, very much as in classical physiology, it provides only high-order views in the form of black boxes with limited predictive and explanatory power. Integration with continuous models will be essential to unravel and engineer the underlying mechanisms.

Helikar T, Konvalina J, Heidel J, Rogers JA (2008). Emergent decision-making in biological signal transduction networks. PNAS 105, 1913-1918

(via Gene Expression)

A few days ago, an exciting review by Michael Lynch was published in Nature Reviews Genetics (The evolution of genetic networks by non-adaptive processes, Lynch 2007a ), a close follow-up of another review, published in PNAS a few months ago (The frailty of adaptive hypotheses for the origins of organismal complexity, Lynch 2007b). Michael Lynch has also written a book on the topic: The Origins of Genome Architecture (read a review)

The architecture of biological networks are often hypothesized as being "shaped" by adaptive evolution to confer global properties such as redundancy, robustness, modularity, complexity and evolvability. Lynch has some robust comments (others have some too, see Jonathan Eisen's "adaptationomics awards") on the “vast majority of biologists engaged in evolutionary studies [who] interpret virtually every aspect of biodiversity in adaptive terms” (Lynch 2007b). In contrast to what he perceives as a widespread belief, Lynch states clearly:

It is an open question as to whether pathway complexity is a necessary prerequisite for the evolution of complex phenotypes, or whether the genome architectures of multicellular species are simply more conducive to the passive emergence of network connections.(Lynch 2007a)

Beyond its somewhat controversial tone, Lynch's central lesson is the need to adopt a population genetics viewpoint (“nothing in evolution makes sense except in light of population genetics”) and he reminds us that, beside natural selection, three additional non-adaptive processes drive the evolution of living organisms: genetic drift, mutation and recombination. By analyzing the interplay between relative rates of loss and gain of regulatory sites (which depend both on mutation rate and mutational target size such as non-coding DNA), population size and recombination frequency, he demonstrates that purely non-adaptive forces can, in principle, determine the level of connectivity of regulatory networks--for example, determine the predominance of highly connected network motifs over linear pathways--without invoking any inherent advantages of the respective architectures on biological functions related for example to development or metabolism. It appears thus that, depending on the population genetics parameters, network structure can be profoundly "shaped" by the mere physical processes of mutation and recombination. At the very least, Lynch proposes that such models should be considered as "null hypothesis" when claiming that selection is engaged in a given aspect of organisms complexity.

In his review of Lynch's book, Massimo Pigliucci draws our attention to the fact that "the genome is only part of the story, arguably the simplest part to figure out", and that one of the greatest current challenges is to explain how phenotypes evolve. Lynch also recognizes that his models are simplified and do not, for example, consider kinetic or dynamical properties of biological networks. But here is a naive question: would it be possible to design an experimental strategy to test directly, in the lab, the evolution of simple (synthetic?) genetic circuits and observe the trends in connectivity under non-selective conditions or are the timescales involved too unrealistic?

How do proofreading and repair activities emerge and what is their impact on system-level properties such as robustness or evolvability?

Biological networks and the Internet may share common architectural principles common to "robust yet fragile" systems (Doyle and Csete, 2007). But the Internet does not (yet) repair itself while cells or tissues do. Repair may represent an example of "downward causation" (The Music of Life, Noble 2006 or a book review) in which a high-level functional property (eg repair activity) emerging from a multi-component system (eg the DNA repair machinery) acts on a component at a lower level of organization (eg one nucleotide). In other words, could repair be considered as an example of a "cross-scale" feedback motif? Can these types of motifs be generalized and how would they evolve? The topic could in fact be extended to the field of synthetic biology: how to assemble synthetic systems with self-repair capability?

Many questions... Who has answers?

see also: Fidelity and infidelity (2001) Radman, 2001

The relationship between genetic mutations and human diseases is often complex and ambiguous: a given disease can be associated with mutations in distinct genes and, conversely, mutations in a given gene can be associated with several diseases. Can this many-to-many relationship be exploited to construct a human disease network and extract information on the human disease landscape?

In their work just published in PNAS, Albert-László Barabasi, Marc Vidal and colleagues reconstruct such a "diseasome" network in which disorders are linked to the respective associated disease genes (Goh et al, 2007 PNAS). Two projections of the network are presented: a) the Human Disease Network (HDN), in which diseases are connected to each other if they share a common disease gene; b) the Disease Gene Network (DGN), in which genes are connected if they are associated with a common disease. The HDN has a giant component comprising almost half of the diseases, in which some classes of disorders cluster naturally (eg cancers or neurological disorders, but not metabolic disorders). The DGN, when integrated with functional annotations, expression and protein-protein interaction data, provides a first step towards a "network-based explanation of the emergence of complex polygenic disorders" in the sense that it reveals, perhaps not too surprisingly, how functionally related genes can lead to similar disorders.

The authors also look at the centrality of human disease genes in the protein-protein interaction network. An interesting twist comes when human disease genes are separated into essential and non-essential classes, according to the lethal or non lethal mouse phenotype resulting from the knockout of the respective orthologous genes. While essential genes tend to be associated with hubs in the interactome, disease genes that are non-essential (representing 78% of all disease genes) do not display a higher connectivity than non-disease genes. A somewhat complementary conclusion was recently reached by Lu and colleagues when looking at changes in gene expression in a mouse model of asthma: genes whose expression is the most affected by the disease have low connectivity while genes coding for hub proteins tend to display stable expression levels (Lu et al, 2007 Mol Syst Biol 3:98).

Reading this work, two main questions come to my mind:

First, if a majority of disease genes are not more central than non-disease genes, what will be the "network-based explanation" for the mere fact that they are implicated in a human disease? What kind of model will be needed to achieve this fundamental prediction?

Second and on a more general note, it looks to me that system-level approaches will be needed to integrate the environmental causes to human disease. While there is no question about the power of genetics and genomics to provide a global view on human diseases, I find it useful to remember that, as Jeremy Nicholson emphasizes,

the majority of people in the world die from what are, in the broadest sense, environmental causes. (Nicholson 2007, Mol Syst Biol 2:52)

One can only agree with Euan Adie, that "the way we present genomic and proteomic data on the web sucks" (read post on Nascent). And this holds for biological networks: depiction of protein-protein interactions as colorful hairballs results in impressive figures but is not obligatorily very useful. While the network representation is a powerful abstract representation of biological processes, it is trivial to say that a graph (with its jungle of nodes and edges) is far from resembling even remotely to an actual living cell as you see it under the microscope... In the crude visualization of biological process as simple graphs, space, time, multi-scale structure and biological context are missing.

Charles DeLisi makes an attempt to tackle the problem of visualization of complex mutli-scale biological networks by introducing the use of metagraphs (Hu et al, 2007, Nature Biotech 25:547). Metagraphs have so-called metanodes in addition to simple nodes. A metanode contains a subgraph composed of child (meta)nodes, which are revealed only when the metanode is in its "expanded" state. Edges link simple nodes while metaedges link "contracted" metanodes and are inferred from the links carried by nodes of the underlying subgraph. A key distinctive feature of metagraphs is that several instances (carrying different "labels") of a node can be shared between distinct metanodes (eg when a protein belongs to different complexes).

Metanodes can represent directly the multi-scale modular hierarchy of a network, incorporate biological context (eg sets of proteins sharing the same GO annotation) or even represent groups of orthologous genes. With this representation, implemented in the software VisANT (http://visant.bu.edu/), "semantic zooming" into the network is made possible. This would be similar to zooming into a Google Map, when not only the scale of the map changes but also the resolution of the labels and various abstract annotations, as is best seen using the "hybrid" mode superposing annotations with the satellite picture.

This analogy with Google Map illustrates also the limits of the current network representation as "maps" of cellular processes. There is still a long way until the graphs representing biological networks can really be mapped onto cellular structures to result into better visualization tools but also into more realistic computational models of the whole cell. In a sense, a "Google Cell" should also have a "hybrid" mode, where the abstract representation can be superposed onto the "satellite image" version of the biological object visualized. As if little tiny networks would be folded inside each voxel of a 3D full reconstruction of a cell, such as the one recently published by Antony and colleagues (Höög et al, 2007, see post). Something like integrating interaction networks, "ORFeome"-like datasets and electron tomography...

How does the connectivity of biochemical networks influence the response to drug combinations?
To address this question, Joseph Lehár and colleagues analyzed the response of yeast and human tumour cell lines to thousands of drug pairs, and compared these experimental data to a computational model (Lehár et al, 2007) . By systematically measuring the biological effects of drug pairs over a range of concentrations, 'response surfaces' could be constructed. This strategy enabled the authors to establish a clear relationship between the shape of the biological response surface and the type of connection linking the targeted biochemical pathways. This dependancy can be used to learn how biological systems are connected as well as to elucidate the mechanism of multi-drug treatments, which are increasingly preferred for many diseases.

It will be interesting to see how far this kind of analysis takes us in the real world: how much will 'off-target' effects confound the analysis? To what extent is the mapping between network topologies and response surfaces a one-to-one function? Further, it would be important to extend this idea to three or more drug components. (Yeh and Kishony, 2007)

Diseases such as cancer are often related to collaborative effects involving interactions of multiple genes within complex pathways, or to combinations of multiple SNPs. To understand the structure of such mechanisms, it is helpful to analyze genes in terms of the purely cooperative, as opposed to independent, nature of their contributions towards a phenotype (Anastassiou, 2007).

• Using an information-theoretic definition of synergy, Dimitris Anastassiou exposes a computational approach to identify ab initio sets of interacting genes linked to a given disease state or phenotype (Anastassiou, 2007). This definition of synergy, derived form a generalization of the concept of mutual information, can connect two levels of organization (for example: genes and disease phenotype) and reveal the structure of the cooperative effects underlying a phenotypic state.
• Jim Collins and colleagues apply network inference techniques to identify key pathways involved in prostate cancer progression (Ergün et al, 2007). A compendium of 1144 expression profiles spanning multiple cancer types is used to train the "mode-of-action by network identification" (MNI) algorithm. When applied on the test set of prostate cancer profiles, the androgen receptor and several of its cognate target genes are identified as top genetic mediators. This signaling pathway would not have been detected by expression change alone or by pathway analysis using Gene Set Enrichment Analysis (GSEA).