The Human Genome Project is providing a genetics parts list of the human and model organism genomes [2,3]. We can identify most of the genes, and, accordingly, their corresponding proteins. This provides the scientific framework for global analyses – where the behavior of all (or most) of the elements can, in principle, be studied. The Human Genome Project is also the first example of discovery science in biology, where all of the elements in a biological system can be defined and placed in a database, e.g., the sequence of the 3 billion nucleotides in the 24 chromosomal strings of the human genome. Discovery science now encompasses a quantitative characterization of all the RNAs in particular cell types, their transcriptome; the quantitative characterizations of all the proteins in particular cell types, their proteomes; the quantitative characterization of the small molecule metabolites in a cell, its metabolome; the characterizations of the protein/protein interactions, the interactome; etc. Discovery science provides systems biology with systems parts lists that are essential to its hypothesis-driven iterative and integrative cycles (see below).

Centers and institutes have been established with cross-disciplinary environments. Biologists are slowly starting to realize the power of cross-disciplinary scientists working in close apposition with biologists. In particular, biologists recognize the critical role applied mathematics, statistics, and computer science is playing in generating the tools for computing, storing, analyzing, graphically displaying, modeling, and ultimately distributing biological information, as well as the role of engineers, chemists, and physicists to produce tools for the global capture and analyses of biological information.

The Internet and the Worldwide Web have provided biologists the capacity to transmit large datasets to colleagues throughout the world. Systems biology will thrive to the extent to which global datasets are openly available to all biologists, and will therefore greatly benefit from the ongoing developments of grid computing, both for enabling real-time large-scale data exchange and collaboration, and for performing high performance distributed computing.

The genome structure is based on a digital code encompassing the core information necessary to initiate development and physiological responses [4]. Since digital codes are ultimately completely knowable, biology is founded with a core of knowable information. The digital genome encodes two fundamental types of information: the genes encoding proteins – the molecular machines of life, and the gene regulatory networks, which specify the behavior of the genes. The gene regulatory networks include the control regions of genes with their DNA-binding sites and their cognate transcription factors [5]. The DNA-binding sites carry out two important roles: (1) they assemble the transcription factors and co-transcription actors for each gene and these collectively operate as a molecular machine to specify the behavior of the gene across developmental or physiologic time – the emergent properties of the gene include temporal and spatial control, as well as the amplitude of expression –; and (2) the DNA-binding sites determine the architecture of the gene regulatory network; that is, which genes are linked in a network by virtue of shared transcription factors. Thus, the control regions act in a manner analogous to integrating computer chips, always sensing changes in the concentrations of transcription factors across developmental or physiologic time.

There are two major types of biological information: the digital information typical of the genome, and the generally analog information of environmental signals. The environmental information falls into two categories: (1) deterministic, where a given signal usually specifies a particular outcome, and (2) stochastic or random, where signals may be quite noisy. Some biological systems, such as the immune system, may have the capacity to convert the stochastic events into information; for example, the stochastic diversification of somatic hypermutation and the assembly of the junctional regions of their B or T cell receptors can be converted by antigen-driven selection into information. Clearly, a major challenge in biology will be to separate signal from noise in large datasets.

Biological information operates across three distinct time spans: evolution – tens to billions of years –; development – hours to a significant fraction of the life span of the organism –; physiologic – milliseconds to weeks. Gene regulatory networks lie at the heart of understanding each of these processes, because the toolbox of protein machines is highly similar across the evolutionary tree of contemporary metazoans. These gene regulatory networks enable the distinct phenotypes of different organisms to emerge as systems properties [5].

Finally, as one moves from the digital genome to the environment, there is a vast hierarchy of types of biological information: DNA → RNA → protein → protein → interactions → biomodules (sets of interacting proteins executing particular phenotypic functions) → networks of biomodules within individual cells → networks of cells and organs → individuals → populations of species → ecologies. The important point is that information is added to the operation of biological systems at each one of these levels; hence, to do systems biology properly, one must acquire global sets of information from as many different levels as possible and integrate them into a coherent picture of the biological system.

Mapping and sequencing of the human genome, and of the genome of model bacterial, plant and animal genomes has been the major goal of the Human Genome Program. At present, the human genome has been announced as ‘completed’, the mouse genome is nearing completion, and those model organisms such as baker's yeast, nematode, fruit fly and the plant Arabidopsis have been essentially completed, as well as those of about a hundred bacteria, and many more are in the pipeline.

These advances have relied heavily on the progress made since the initial description of DNA sequencing chemistries in 1977. Initial attempts at automation of the Maxam–Gilbert and Sanger method were unsuccessful. Attachment of fluorescent dyes to nucleotides was described in 1985, enabling implementation of a first version of the DNA sequencer in 1986 [6]. However, scaling up of sequencing to the massive scale required to generate an accurate human genome sequence was possible, more than twelve years after the introduction of the first automated sequencer, only after significant improvements in data quality generation were made available through advances such as the use of dye terminators and fluorescence energy transfer for enhancement of signals, and when assessment of data quality across sequencing platforms and laboratories became a standard practice with software packages such as Phred/Phrap. In the process, a 3000-fold throughput increase has been achieved, and another 3000-fold increase is anticipated during the next decade, with the prospect of nanoscale instruments operating at the single molecule level.

The content and accessibility of public electronic repositories have increased dramatically during the same period, benefiting largely of advances in computer technology and the development of the World Wide Web after 1992. As a matter of fact, the size of the DNA sequence databases is increasing so rapidly that its growth exceeds Moore's law predicting a doubling of computer chip performance every 18 months to two years [4]. It is therefore anticipated that smarter ways of dealing with the avalanche of DNA sequencing data will have to be invented and implemented, while maintaining a high level of accuracy, since we cannot rely merely on the projected increase in computing power. Some of the most powerful computers, which were commonly used so far in fields such as forecasting or telecommunications, are now used in the field of genomics.

As simple as it might look, completion of the sequencing of large genomes is not an easy task, even if one considers only its euchromatin part, leaving the heterochromatin aside. In the initial ‘complete’ descriptions, the human genome sequence was split in over 100 000 fragments of various sizes, with significant uncertainties as to the order and orientation of many of them, as illustrated by a comparison of the genome assemblies of the publicly and privately generated versions (Tajashi Gojobori et al., personal communication). This might be due to the existence of regions of the genome that are intrinsically unstable, and therefore difficult to clone and sequence; they might contain important genes and regulatory DNA sequences, as suggested by the observation that disease-related genes tend to be found around the regions of discordance between the physical, genetic, radiation hybrid and cytogenetic maps [7].

As the sequencing of the human and mouse genomes was proceeding to completion, it could be anticipated that the number of genes would be narrowed down to a consensus number. The current guess is that there are 30 000–35 000 genes in these species. More may be identified if small regulatory RNAs are defined as genes and if many small protein-coding genes are identified.

Given that the number of potential genomes made of 3 billion base pairs of DNA is approximately equal to 10 to the power of 1 billion, a number so large that a computer made of all particles in the Universe (10⁷⁰) would have a hard time to browse through all of them, the debate as to whether a particular eukaryotic genome (ours) contains 10⁴ or 10⁵ genes seems to lack relevance. Emphasis should be rather placed on understanding the combinatorial rules that make use of a similar pool of genes by different species in different contexts, taking into account the other dimensions of the genome, including the various types of DNA polymorphisms, the effects of chromatin structure and methylation, etc.

RNA expression profiling to monitor the transcriptome has also a long history starting with RNA complexity measurements performed by kinetic reasssociation in the 1960s and early 1970s, the famous Rot curves. The first cDNA array experiments were performed soon after the initial description of cDNA cloning in 1975 [8]. The technology was based on bacterial colonies spotted by hand on nitrocellulose filters using toothpicks, then hybridized with radioactively labeled cDNA and revealed on X-ray films, providing semi-quantitative measurements. After the invention of PCR in 1985, cDNA inserts could be amplified in microtiter plates, spotted onto Nylon membranes; exposure to phosphor plates introduced around 1990 made it possible to obtain quantitative measures on a dynamic range spanning three orders of magnitude. Automation of the full process was made possible with the introduction of commercial picking and spotting robots after 1992, around the time when in situ photolithography oligonucleotide synthesis was first described [9]. The now popular two-color fluorescence cDNA arrays were introduced in 1995 [10], and deposition and in situ oligonucleotide synthesis by ink-jet technology in 1997 [11].

Although this technology has made important progress, particularly during the past decade, it has not yet reached the level of maturity and robustness of DNA sequencing. This is well illustrated by the fact that public repositories of microarray data have been introduced only very recently, and discussions on initial formats and standards are still ongoing. As a result of the availability of a myriad of instruments, reagents and tools for capturing and analyzing the data, quality assessment and standards have yet to come. This imposes severe restrictions on the possibility of comparing datasets generated using similar but somewhat different procedures, and on the depth of analysis possible when multiple samples are compared. In spite of early efforts to introduce statistical assessment of the data, it is only recently that mathematicians and statisticians have started to tackle the problems associated with microarray experiments, such as signal identification and measurement, data filtering and normalization, and data analysis and mining. Large amounts of data have been generated, most often with limited or no replication, and with little attention paid to the basics of experimental design. As a result, the complex experimental and biological variations associated with microarray data are rarely documented, preventing a thorough analysis and seriously limiting the power of integration with other types of data.

In the frame of systems biology, it is essential to perform transcriptome measurements based on a robust experimental design, including a precise description of the biological problem, and following standard operating procedures under quality assurance. The current technologies provide reasonable access to significant variations of relatively intense signals, which has been largely documented by a variety of other analytical techniques during the past decades. Reproducing established results, and generating some novel ones, such as identification of biomodules operating in galactose metabolism of yeast discussed below, is reassuring on the value of the technology in its present state.

However the challenge ahead is to access the small variations of the weak signals, corresponding to the vast majority of genes, since they potentially convey collectively more biologically relevant information than the limited number of those associated with strong signals and variations. Novel technologies under development, enabling massively parallel and controlled measurements down to the single molecule level, will help to overcome the current hurdles. Indeed, the single molecule approaches will move the analysis of transcriptomes from an analog mode (DNA and oligonucleotide arrays) to a digital mode where it will be possible to analyze transcriptomes down to the single mRNA per cell. This ability to visualize mRNAs expressed at very low levels is critical because much interesting biology operates at this low level of mRNAs (e.g., gene regulatory networks and some signal transduction pathways). The ability to measure digitally in cells down to the single RNA molecule level will also enable us to address other dimensions of the roles of RNA such as alternative splicing, editing, folding, stability and degradation, each of which adds to the complexity of the transcriptome.

An analysis of the proteome begins with a listing and quantitative enumeration of all the molecular species of proteins in individual cells. Proteomic analyses represent an ever-greater challenge than genomic or transcriptomic analyses for one simple reason: the dynamic range of protein expression in a cell vary from one to 10⁶ copies. There is no equivalent of PCR for proteins – hence the limits of detection for proteins are limited by the sensitivity of the analytic tools. A widely used tool for protein analysis – the mass spectrometer – may have at best down to attomole sensitivity (e.g., requires 10⁵ molecules). This challenge can be overcome, of course, by analyzing the proteins from large number of cells.

Proteomic analyses also pose two other challenges – shared in part by genomic analyses. First, cells change across physiologic and developmental time dimensions. One wants to capture the changing snapshots of the changing patterns of protein expression across these time dimensions – for they reflect the changing biology. Physiologic responses may occur in a fraction of a second. How can these real time snapshots be obtained with global analyses (e.g., all protein elements)?

Finally, proteins exhibit a hierarchy of different informational states and locations that inform the biology they execute. We need to be able to identify and quantitate all of the molecules species of proteins in a cell. We need to characterize all their chemical modifications and correlate them, where possible, with changes in biological function. We need to measure accurately the protein/protein and protein/DNA interactions for these data constitute the foundations for building protein and gene regulatory networks. We need to measure protein half-lives and activations. Where proteins are localized in cells is key. We also need ab initio and better experimental methods for determining dynamically changing three-dimensional structures and correlating these with protein function. A wide range of platforms will be necessary for all these measurements.

Two-dimensional gel electrophoresis (2D-GE), introduced in 1975, was probably the first analytical technique available to analyze the proteome, but it could not provide directly access to the protein sequence information. The introduction of the automated protein sequencer proved invaluable to collect such information on specific proteins, including some isolated by 2D-GE [12]. The automated peptide synthesizer made it possible to produce proteins or protein segments [13], and to generate specific reagents such as polyclonal and monoclonal antibodies to probe their structure and function as individual elements, or as part of molecular complexes. Although these methods could not manage the diversity of proteins, they helped revealing the multidomain structure of proteins, and their assembly in multimeric and supramolecular molecular machines performing simple and complex biochemical functions. A global approach to the diversity of protein–protein interactions (the interactome) was made possible by the yeast two-hybrid system in its various forms, although it remains to be seen how many of the interactions revealed by such techniques are relevant to the real context of the original biological system. In fact, comparison of independently derived datasets has indicated limited overlaps, indicating that many of the interactions identified may be yeast-context specific, or simply occur by chance. Recently, a number of novel technologies have started to provide the means by which to interrogate proteomes on a more global scale, combined with the ability to provide non-ambiguous identification of their elements. These include specific labeling of subsets of proteins, based on their content of specific residues or protein modifications (ICAT), and the use of liquid chromatography coupled to tandem mass spectrometry [14]. A variety of array-based technologies are currently under development to probe proteomes with antibodies, or to reveal protein-DNA interaction.

There are many more aspects of protein structure and function that require similar technological advances before global analyses become practical. Mass production of recombinant proteins will help decipher their structures by X-ray crystallography and Nuclear Magnetic Resonance. The generation of large collections of antibodies to all human proteins is under way to provide the reagents to probe tissue and cell arrays and identify the spatial and sub-cellular localization of these proteins.