The practical problems of post-genomic biology

describes the emergence of new postgenomic disciplines like proteomics, epigenomics, and phenomics. He proposes that integrated “omic” databases in the future will allow the in silico re-enacting of the entire organism once the totality of the information at the various levels, from genome to proteome and functions, is available. This is an attractive idea: just think of virtual organisms that could serve for virtual clinical trials. However, the “omic” strategy is based on the same linear, monocausal, deterministic thinking that has reigned in molecular biology over the past decades2; it just extrapolates it to the entire genome. This is brute-force, genocentric reductionism in the guise of entireness, rather than a novel integrative approach devoted to wholeness, or as Nature Biotechnology put it in an editorial: “Complicated is not complex.”3 In fact, the present efforts in the drug development industry to simulate organismal function still ignore the very elementary parts, and instead employ the traditional “top-down” modeling approach of physiologists and systems engineers, in which organs and cells are black boxes. Such approaches vastly dominated a recent workshop on computer modeling in biology, despite the promising conference title: “From gene to organ.”* With their dream of simulating organisms “from bottom-up,” genomics scientists touch upon the old riddle of genome-to-phenome mapping. They will have to overcome the mentality of data collecting, clustering, and classification, and develop a qualitatively new conceptual understanding of the complexity of living organisms. Help in tackling this daunting task is already on its way: not only are physical scientists increasingly interested in studying complex systems in general, but one of the first institutions devoted to such studies, the New England Complex Systems Institute (NECSI) in Boston, MA, has launched collaborations with molecular biologists to harness the data flood triggered by the success of genomic technology. To counter, but not stifle Evans’ enthusiasm, let me give an overview of some of the major obstacles and challenges that we will have to confront in a post-genomic endeavor that will ultimately lead to the design of cyber-guinea pigs based on our genomic databases: Ontological questions concerning database structure and representation of “gene function”. A major goal of making genome databases more user-friendly is the systematic annotation of genes with functional properties, such as the biochemical function, interaction partners, and physiological role. This has worked pretty well for simple organisms, as demonstrated in the Saccharomyces Genome Database4. However, in more complex organisms, the same task represents an ontological challenge, as most proteins are embedded in a regulatory network and don’t just code for metabolic enzymes. For instance, genes such as ras, myc, rho, and NF-κB either stimulate growth and survival, or they induce apoptosis. The “function” of a gene product appears to depend on its “cellular context.” This raises fundamental questions. For example, what is a “gene function?” What is its “context?” And what is the modular entity to which the concept of context applies: a protein domain, a single protein, a stable complex, or a whole pathway? Recently, the principle of “modularity” in cellular processes has been proposed to facilitate the task of dealing with biological functions5. For example, a specific signal transduction pathway could be viewed as a functional module. However, it remains to be seen whether modules are “natural kinds” or just logical constructs of our mind and how their boundaries have to be delineated given the extensive “crosstalk” between the historically defined pathways. These questions must be addressed first before useful higher level “omic” databases can be constructed. Limitations of computational power. In proposing the in silico re-enacting of organismal function, Evans seems to ignore the immensity of computational capacity required. Take the human genome with 100,000 genes and let every gene be simply either “on” (expressed) or “off ” (silent). This minimal, idealized, and discrete setting alone would lead to the astronomical number of 1030,000 possible gene expression profiles! The computing and testing of all these patterns with the existing serial computers would take more time than the age of