Having in hand a table of gene entries, with summarized annotations is not sufficient to handle genomic information. Data can be organized based on biological principles so as to reflect current knowledge at best, and to be viewed at a glance in global X-omic experiments, or for comparative purposes. The question of knowledge representation in reductionist terms (i.e. based on the fact that some levels of knowledge can be reconstructed by the integration of parts of knowledge of lower levels) is a difficult epistemological question and can hardly be debated here. Current data integration strategies are dependent on available consensus methods, with their compromises and imperfections that have been slowly adopted by the scientific communities. Biological (and chemical) knowledge is primarily produced in the form of research publications, books, patents and other un-structured texts, and since two decades in structured databases (biologists are used to fill Genbank forms prior to paper submissions). Data and information can be organized as semantic networks, following combinations of ontologic hierarchies and praxeologic schemes 80. In brief, an ontological hierarchy is designed to organize entities (here biological or chemical entities) following inclusion/subsumption principles ("A" is part of, is a component of, etc. "B"). Best known examples are the taxonomic trees ("species" is part of "genus", etc.) or the Gene Ontology, or GO 31, although this later was built from a lose definition of the ontology (see below). A praxeologic scheme allows a representation of the activity or function (praxis) (for instance enzymatic activities caused by proteins, assembly of molecular structures, biological effects caused by drugs, etc.) based on the transformation or alteration of an entity into another, through time ("C" is converted into "D"). By contrast with ontologies, which should be stable and non-conflicting hierarchies, praxeologies can describe cyclic processes and can vary over time. Best know examples are metabolic graphs and fluxes, and their variations in different physiological conditions. Interestingly, the enzymatic activities can be organized following hierarchical principles (e.g. the EC numbers proposed by the Enzyme Commission): a praxeologic scheme such as a metabolic pathway is therefore linked to one or more ontologies for the metabolites (hierarchical categorization of molecules) and for enzymes (hierarchy following the GO, the EC, etc.). In practice, although they are of very distinct nature, ontologic hierarchies and praxeologic schemes are handled by bioinformaticians as "semantic networks" and relational graphs senso lato.

Following recommendations of the Gene Ontology (GO) consortium 31, malaria gene function was defined using a semantic network organized in three hierarchical axes, i.e. "molecular function", "biological process" and "cellular component", according to a controlled hierarchical vocabulary. The GO-based annotation circumvents the problems raised by heterogeneous key word annotations, allowing subsequently a cross-species comparison with genomes annotated similarly, and complies with in silico management requirements. Genes thus annotated can further be embraced in higher order representations of biological knowledge.

Concerning enzymes, entries can be linked to graphical representations of the metabolic reactions they catalyze. To that end, in the same way genes were defined following the GO procedure, enzyme substrates and products are defined following a chemical ontology (CO), and the reaction itself has to be defined following an enzymatic reaction ontology. A reaction can be viewed as a small graph in which the enzyme is associated with the line that connects the nodes corresponding to the substrates and products. Metabolic pathways are connected by shared nodes, and can be viewed at different scales. Current representations of malaria metabolism have been made available using generic methods, i.e. KEGG 4881 or MetaCyc 498283, or specifically designed for Plasmodium, i.e. the Malaria Parasite Metabolic Pathways (MPMP) coordinated by Hagai Ginsburg at the Hebrew University of Jerusalem 5051.

The Kyoto Encyclopedia of Genes and Genomes (KEGG) resource provides a set reference of metabolic schemes, manually designed so as to represent all possible primary metabolic reactions, and formatted with the KEGG Markup Language (KGML) 48. These reference schemes can be explored on the KEGG web portal, with a very clear view of the global metabolic map, connecting all pathways. Organism-specific schemes are generated, based on sequence similarities (using the KEGG Orthology, or KO orthologue identifier) with references to the KEGG Gene catalogs. Thus 79 metabolic schemes have been generated for P. falciparum. Each scheme provides links to metabolite (substrates, products, co-substrates) information, and enzyme descriptions following the Enzyme Commission (EC) classification. EC numbers give access to multiple sequence alignments, protein motifs, genomic mapping, links to Genbank, UniProt, PDB, etc. Tilling KEGG schemes from different organisms highlights metabolic similarities and differences, and could be of help for anti-malarial purposes, highlighting for instance metabolic reactions occurring in Plasmodium and not in humans. However, the maps are designed based on the enzymatic reactions and, for instance, the fatty acid synthesis due to the type 1 fatty acid synthase (FASI, a multiprotein complex) from the human cell cytosol, strictly overlaps with the type 2 fatty acid synthase (FASII dissociated enzymes) from the Plasmodium apicoplast. Information on the protein structure and cell compartmentalization of the enzymes would have been sufficient to distinguish between FASI and FASII. Furthermore, metabolites that are generated in one compartment (i.e. diacylglycerol generated in one of the numerous cell membranes) may not be available for an apparent downstream reaction occurring in another compartment. Thus, used without caution, the KEGG schemes may seem to be fully valid for the entire living world, with a misleadingly clear and fully cross-connected global overview of metabolism, and can lead to unrealistic representations. The missing enzymes, which have been experimentally assayed, are not shown, and it is unclear whether gaps within pathways are due to absent enzymes or to incomplete data. Additionally, the KEGG representation is not intended for the design of schemes other than those pre-defined. KEGG outputs need to be reexamined for accuracy of interpretation (see below).

As an alternative, the design of MetaCyc schemes for Plasmodium, called PlasmoCyc, has been initiated 4982. As for KEGG, a reference of the complete metabolic pathways has been designed manually and loaded in a MetaCyc database. In contrast with the fully connected overview of the KEGG metabolic map that can make the user overconfident, the MetaCyc global view of metabolism is fragmented, reflecting knowledge gaps, incomplete design of some pathways (such as the tricky pathways for lipid syntheses), and the versatility of the MetaCyc tool for implementation of new schemes. Using P. falciparum gene annotation and information of the MetaCyc reference database, the PathoLogic module of the Pathway Tool Software 84 allows the generation of a Pathway/Genome Database (PGDB). A total of 113 metabolic pathways (complete or fragmented) have been generated for P. falciparum. As in KEGG, each graph gives access to metabolite, reaction and enzyme information. The Gene-Reaction Schematic (GRS) representation allows a visualization of the relation between the genes, the enzymes, the catalyzed reactions, even in the case of complex or multienzymatic proteins. This model is useful to distinguish for instance FASI and FASII proteins. It is further useful for Plasmodium proteins that are often multienzymatic. For each reference pathway, the occurrence or absence of enzyme homologues in the Plasmodium genome is documented. As in the case of KEGG, compartmentalization information is missing from the pathway model, in spite of some effort to highlight some specific pathways (for instance the apicoplast fatty acid synthesis).

A synthetic representation of Plasmodium biological processes can further be viewed at the Malaria Parasite Metabolic Pathways web portal 50. Among 120 schemes representing numerous cellular processes, half represent metabolic pathways and were fully designed for malaria researchers. All schemes were built using KEGG pathways, cleaned of irrelevant information and curated by international experts. The quality of each representation is very high and specialized, and benefits from a sustained effort in biological and molecular investigation and validation. Missing data are concisely documented. Most importantly, enzyme subcellular localization is shown. Graphs are not automatically generated, but drawn by experts. While MetaCyc graphs are self generated taking into account all information of the underlying database, the MPMP scheme are not self generated: in particular, the subcellular localization representation is not generated automatically from a subcellular attribute filled in the underlying database. Thus, although the MPMP representations are of higher quality, their update is not dynamic and cannot be used for in silico graph-based treatments.

From this short overview, it is clear that metabolic representations should be carefully used. On one hand, generic approaches (KEGG, MetaCyc) do not include cell compartmentalization data and the chemical ontology for lipids is not finished. On the other hand, the curated MPMP representations are based on high-quality data but they are static, with no graph management tools, giving click-access to remote information. Future challenge will be to design underlying models for building graphs from genomic data (following the GO), metabolite data (following a stable Chemical Ontology), reaction connections and compartmentalization information for both gene products and metabolites. Construction of PlasmoCyc benefited from information of the MPMP 5082, and MPMP was designed after cleaning KEGG schemes. Models for other knowledge bases and graphs should benefit from the important effort of the MPMP in defining metabolic and other biological processes.

The application of functional genomics strategies to assign functionality to each gene product of an organism has recently attained increased attention in the field of post-genomics research of Plasmodia, includes the understanding of the transcriptome, proteome and interactome of the parasite to elucidate mode-of-action of inhibitory compounds, allow optimization of such inhibitor activities, explain resistance mechanisms to known drugs, chemically validate potential drug targets and ultimately identify and/or functionally describe new drug targets (Figure 1). Some malaria X-omic studies did not meet the initial expectations, such as transcriptome analyses of P. falciparum in which the modification of the transcriptomic pattern to different treatments was particularly low. Some others were technically highly biased, such as interactome analyses in which results from protein fragments are recorded together with results from intact proteins, and raise imperfect, but interesting sets of data. In this part, an overview of malaria X-omic analyses is given, with some insights on the current strategies for in silico integration of their functional outputs with genomic knowledge.

The P. falciparum transcriptome has been extensively investigated resulting in comprehensive profiles of transcript expression throughout the complete life cycle of the parasite 8586. The overall conclusions demonstrated that the majority (~87%) of the predicted genes are actively transcribed during the lifecycle but that 20% are specific of the intraerythrocytic developmental cycle and are produced in a periodic nature in a 'just-in-time' fashion. These early reports have been followed by investigations designed to answer numerous biological questions, including transcription and post-transcription specific aspects of the regulation of protein expression, transcriptional machinery and inheritance, interstrain conservancy, gametocytogenesis and antigenic variation control mechanisms 8788899091. A high degree of correlation exists between the in vitro and in vivo transcriptomes of P. falciparum with an overexpression seen for genes encoding a sexual stage antigen as well as gene families that encode surface proteins, providing interesting new vaccine candidates 9293.

Proteomics studies are essential to conclusively prove mechanistic changes and explain global protein expression profiles, differential protein expression, posttranscriptional control, posttranslational regulation and modifications, alternative splicing and processing, subcellular localization and host-pathogen interactions (Figure 1). Reassuringly, there is a good correlation between the abundance of transcripts and the proteins encoded by these during the P. falciparum lifecycle 8788 with a majority of discrepancies attributed to a delay between transcript production and protein accumulation. Analysis of the P. falciparum proteome 94 and a comprehensive and integrated analyses of the genome, transcriptome and proteome of P. berghei and P. chabaudi chabaudi, which represents the state-of-the-art of functional genomics applied to the lifecycles of Plasmodia 9596, indicated that over half of the proteins in these parasites were detected solely in one stage of the lifecycle. This implies a considerable degree of specialization at the molecular level to support the demanding developmental programme and suggested a highly coordinated expression of Plasmodium genes involved in common biological processes.

In-depth understanding of the protein-protein interaction network (defined as interactome) can provide insights into the function of proteins, regulatory mechanisms and functional relationships of these (Figure 1). An extensive protein-interaction study of P. falciparum (nearing global-type analysis) combining interaction information with co-expression data and GO annotations indicated unique interactions, identified groups of interacting proteins implicated in various biological processes, and predicted novel functions to previously uncharacterized proteins 97. Interestingly, comparison of the Plasmodium interactome with that from other organisms indicated a marked divergence with very little conservation with other protein complexes 98.

The abovementioned X-ome datasets are therefore now available for in silico data mining approaches. As such, data mining of the transcriptome using an extensive sequence similarity search identified 92 putative proteases in the P. falciparum genome, 88 of which are actively transcribed 99. This strategy has also been applied to the kinase family 100. The transcriptome datasets are nowadays extensively utilized in PlasmoDB, MPMP and other specialized sites 101. PlasmoDB additionally has links to proteome data and also allows access to the interactome data. Proteome experiments based on 2D-gel electrophoresis can furthermore be aided by Plasmo2D software to allow the identification of proteins in such platforms 102.

An important question in mining different global X-omic datasets is how they can be compared. Draghici et al. 103 made clear that the inconsistencies between the various microarray platforms (in situ synthesized short oligos, longer oligos, spotted oligos or cDNAs) are so high that it is almost impossible for the moment to compare results from different platforms. How can a transcriptomic profile be compared to a proteomic profile, given the errors, the linearity of signals and the magnitude of variations of each method, and the biological stability and turnover of RNA transcripts and of proteins? Is the enzyme profile correlated with the metabolite profile 104? A pragmatic solution is to avoid the multiplicity of methods and by agreeing on some standards 103105106107. Although "global" invariant references might not exist, it is nevertheless worth trying to find in large gene expression matrices the most invariant (or less variant) genes 106. This quest is of general concern and is currently one of the challenges proposed by the European Conferences on Machine Learning and the European Conferences on Principles and Practice of Knowledge Discovery in Databases 106109. In the absence of absolute references and standards, outputs from malaria X-omic experiments should be analysed and compared with caution, particularly when obtained with platforms that are distinct from those used to feed the public data repositories. Consequently, in addition to referential public repositories for functional genomics, versatile software for local analyses are strongly needed.

Various tools aim to provide biological interpretation of gene clusters but these mostly specialize in only one or two types of analyses. FatiGO 110, GeneLynx 111 and Gostat 112 are powerful tools for Gene Ontology mining; GoMiner 113, MAPPFinder 114 and DAVID 115 use GO and metabolic pathway interpretation whereas GeneXPress 116 and MiCoViTo 117 use metabolic pathways and incorporate transcription regulation visualization. Improvements on these include a web interface called MADIBA (MicroArray Data Interface for Biological Annotation, 118) that has been initially designed for malaria transcriptomics (Dr. C. Claudel-Renard, personal communication). This interface links a relational database of various data sets to a series of analysis tools designed to facilitate investigations of possible reasons for co-expression of clusters of genes (e.g. from gene expression data) and to deduce possible underlying biological mechanisms. Clusters of co-expressed genes are automatically subjected to five different analytical modules including 1) search for over-represented GO terms in clusters, 2) visualization of related metabolic graphs with KEGG representations, 3) chromosomal localization, 4) search for motifs in the upstream sequences of the genes and finally 5) Plasmodium-specific genes without human homologues. MADIBA analysis of the transcriptome dataset from Le Roch et al. 85 resulted in an improved annotation of the Plasmodium genes (41% vs. 37%) and characterization of 6 additional clusters with GO annotations, of which one exclusively contained glycolysis in its entirety (except for fructose-bisphosphate aldolase) and another identifying gene as potential as drug targets due to their Plasmodium-specific characteristics. Therefore, MADIBA allows versatile analyses of a vast variety of transcriptomic profiles. These analyses can highlight potential drug targets by providing functionality to co-clustered expressed genes in a guilt-by-association manner, including those of un-annotated proteins, by predicting co-regulated expression via chromosomal localization as well as the identification of motifs for cis-regulatory elements and lastly by identifying unique Plasmodium-specific genes involved in specific biological mechanisms.

With the advent of integrative investigations of datasets from the transcriptome, proteome, interactome etc. new analysis tools are being developed including a Partial Least Squares (Projection to Latent Structures-PLS) method which has been used to integrate yeast transcriptome and metabolome data 119. Linear modelling was used to investigate the changes in the transcriptome due to environmental perturbations and, assuming that the metabolome is a function of the transcriptome, the metabolic variables were modeled with PLS. A genome-wide investigation of protein function was recently performed by computationally modelling the P. falciparum interactome 120 to elucidate local and global functional relationships between gene products. This novel approach entailed an integration of in silico and experimental functional genomics data within a Bayesian framework to create the network of pairwise functional linkages. This resulted in predicting functionality based on associations between characterized and uncharacterized proteins for 95% of the currently annotated hypothetical proteins in the P. falciparum proteome. Only 107 hypothetical proteins show interaction with other hypothetical proteins potentially representing new pathways or previously uncharacterized components of known pathways.

This overview shows that the integration and mining of global functional genomic experiments are in the frontline in drug discovery processes 121122. Malaria X-omics data provide comprehensive information to, amongst others, understand the mode-of-action of inhibitory compounds, allow optimization of drug action, validate drug targets (chemical validation strategies), identify families of genes/gene products that are more amenable as drug targets ('druggable genes'), annotate the function of hypothetical proteins by 'guilt-by-association' and point out specialized gene expression regulation systems (Figure 1) 123.

In the case of drug treatments, the sought effects on the metabolism of targeted tissues or organisms include up- or downregulation of the protein target(s), the upregulation of detoxification pathways (cytotoxic responses) and the upregulation of alternative or compensatory pathways of the affected organism that can be reflected in changes in the transcriptome/proteome of the organism. The characterization of the transcriptional response induced by drug challenge has been applied with success in the antibacterial field, creating reference compendia of expression profiles after drug challenge that provide insight into a drugs' MOA 123124125. Transcriptional profiling of drug challenged malaria parasites has been limited to only a few studies to date including Serial analysis of gene expression (SAGE) of chloroquine treated parasites 126, a high-density short-oligonucleotide array study on parasites treated with phosphatidylcholine biosynthesis inhibitors 88 and custom arrays originating from suppression, subtractive hybridisation (SSH) libraries on parasites treated with polyamine biosynthesis inhibitors 127. Drug-specific transcriptional responses were seen in the chloroquine and polyamine inhibition studies indicating the presence of a feedback signaling mechanism. Le Roch et al. 88 compared transcriptional responses with proteome analyses and showed that more pronounced changes were induced at the protein level after drug challenge. This is also true for antifolate inhibited parasites where a marked increase were seen in folate biosynthesis protein levels upon treatment with inhibitors against DHRF-TS 128. Global-level proteome response analysis of the combination of artemether and lumefantrine also revealed drug-specific changes in the proteome 129. Subproteomic investigations have additionally become particularly important to determine the molecular binding partner/target protein of an inhibitory compound and/or to describe the mode-of-action of such compounds. This has been applied to ferriprotoporphyrin IX were identified 130, kinase inhibitors 131 and the quinoline family of compounds 132.

Any of the above-mentioned strategies are potential starting points to the discovery of unsuspected drug targets in P. falciparum, whether it is a new/additional functionality that is ascribed to a known protein or the characterization of novel function of a previously 'unknown' protein. As the hypothetical proteins represent more than half (~60%) of the malarial proteome (see above), these are some of the most attractive areas to the drug target discovery effort. The basic rationale behind using expression profiles to assign functionality to genes is based on the principle of guilt-by-association 133134 in which genes coding for proteins with similar functionality often exhibit the same expression profiles and protein-interacting partners. Coincidentally, the expression profiles of genes specific to a given organelle also display similar expression patterns. This principle has been applied with unsupervised robust k-means clustering 85. Improvements on this clustering approach were proposed using a semi-supervised clustering method called ontology-based pattern identification (OPI) 135. OPI uses previous gene annotation data to generate clusters with greater specificity and confidence whose members then additionally share the same expression profiles. However, Llinás and del Portillo 136 warns against using only classical guilt-by-association methods, showing that many genes that are functionally unrelated show similar expression profiles during the asexual development of P. falciparum.

Vital malaria proteins may have no counterpart in humans or sequence dissimilarity with their human homologues that may be sufficient to become a therapeutical target without disturbing essential function in the human host. The literature on potential protein targets for anti-malaria treatments is already crowded, and will hopefully be enriched and better documented in the future. Once a target has been identified, an important decision is whether or not to enter into costly drug- or vaccine-development programmes. Determination or prediction of the target three-dimensional structure and in silico experiments can be achieved to connect these biological targets with the pharmacological space of small compounds 137, and assist this decision and provide initial clues on possible therapeutical strategies.

In this paper, five aspects of the in silico storage, organization and mining of data from malaria genomics and post-genomics, were examined in the context of the prediction and characterization of targets and drugs: 1) the comparison of protein sequences including compositionally atypical malaria sequences, 2) the high throughput reconstruction of molecular phylogeny, 3) the representation of biological processes particularly metabolic pathways, 4) the versatile methods to integrate genomic data, biological representations and functional profiling obtained from X-omic experiments after drug treatments, 5) the determination and prediction of protein structures and their virtual docking with drug candidate structures. Data management that should be at the genomic scale, including multiple species, and should therefore be as reliable as possible. A "biological space" should be formatted so as to represent scientific knowledge and to connect genomic data and post-genomic functional profiles in the most versatile way, allowing the mining of information with diverse methods (Figure 4a–e). This "biological space" should be linked to a "chemical space" that contains all small molecules stored in chemolibraries (millions of compounds) including known drugs (Figure 4f–h). Thus, progresses toward a chemogenomic knowledge space will benefit on the referential public repositories and particularly UniProt and PlasmoDB, on recent theoretical advances in genomics and post-genomics data management and mining and on the power of computer grids (Figure 4).

Genomic data (Figure 4a), i.e. protein sequences, can be organized based on sequence similarity (Figure 4b). This projection of protein sequences should allow the high throughput reconstruction of molecular phylogenies both at the intraspecific (connecting paralogs and alleles) and interspecific (connecting homologues among which orthologs) levels, following statistically accurate methods (Figure 4b). This task is difficult because of the compositional bias and high insert content of malaria sequences. Statistically valid protein sequence comparisons are now available to allow genome scale alignment and high throughput phylogeny reconstruction (named TULIP), including malaria atypical sequences, and providing quality scores on which one can rely after automatic genome scale treatments. Benefits from the TULIP method include an easy upgrade and update of the obtained protein spatial projection. Another substantial side of a biological space contains representations of the knowledge of biological processes, using stable ontologies, and dynamic graph representation (Figure 4c). Here, efforts are still needed to improve existing representations of the metabolism and numerous projects are under progress, most importantly PlasmoCyc and PMPM. The PlasmoCyc existing metabolic graphs have the advantage of being more easily updated and usable for in silico mining methods, as long as the output is examined by biological experts. In a brief overview of malaria X-omic profiling, some tools allowing a linkage between genomic data, biological process representations and global functional profiles have been introduced (Figure 4d). There is still a large diversity of data treatment and mining strategies, reflecting the diversity of technologies and mining methods. This step is one of the doors to connect the biological space with the chemical space, following the "reverse chemical genetic" way, i.e. "from known drugs to biological response" (toxicity, mode of action). Basic analytical tools like MADIBA and sophisticated mining approaches will be needed to understand and compare the biological responses to anti-malarial drugs. The other door to connect the biological space and the chemical space follows the "direct chemical genetic" way, i.e. "from known biological target to drug candidate". In addition to crystal structures of malaria proteins, the automated structural annotation of the malaria proteome should be initiated (Figure 4e). Based on protein structure information, virtual docking campaigns such as the WISDOM challenges can be achieved using the power of computer grids.

This paper did not review the "chemical" side for chemogenomics space. By numerous aspects the in silico organization of the small molecules stored in chemolibraries (Figure 4f) was not achieved the way biological information was. Clustering of small molecular structures based on properties is highly debated (Figure 4h). Collections of small molecules are generally designed to obey the pragmatic Lipinski's "rule of five" 181 making them likely candidates for drug discovery. Numerous studies are under progress in order to identify which small molecule descriptors can comply with chemogenomic approaches (e.g. the Accamba project for the analysis of chemolibraries and the building of bioactivity models; 182). A chemical ontology (CO) has been recently introduced 183, but it has not yet been used and validated by the scientific community the way the GO was. A database for Chemical Entities of Biological Interest (ChEBI) has also been launched 184. The modelling of the three-dimensional structures of small molecules (Figure 4g) can be predicted by numerous public or commercial methods which should be examined with attention (see the WISDOM challenges). PubChem, a repository for molecules acting on biological targets was recently launched 183185 and the UniProt protein knowledge base was recently upgraded to report toxic doses of small molecules on proteins 186187, however these initiatives are just starting points. Access to an ocean of small molecular structures and to a deluge of biological sequences raises an enthusiastic challenge: "The goal for the coming decades will be to explore the overlap between chemistry space and protein space" 188. Is this prediction exuberant? From the methodological survey summarized in this paper, this next milestone for malaria research is not out of reach.

Beyond virtual screening, the grid technology provides the collaborative IT environment to enable the coupling between molecular biology research and goal-oriented field work 189. It proposes a new paradigm for the collection and analysis of distributed information where data do no longer need to be centralized in one single repository. On a grid, data can be stored anywhere and still be transparently accessed by any authorized user. The computing resources of a grid are also shared and can be mobilized on demand so as to enable very large-scale genomics comparative analysis and virtual screening. A longer term perspective is, therefore, to enhance the ability to share diverse, complex and distributed information on a given disease for collaborative exploration and mutual benefit. The concept of a knowledge space is to organize the information so that it can be reached in a few clicks. This concept is already successfully used internally by pharmaceutical laboratories to store knowledge 175. The grid permits the building of a distributed knowledge space so that each participant is able to keep the information he owns on his/her local computer. A set of grid services would particularly take advantage of the developments in the area of semantic text analysis for extraction of information in biology and genome research. (Figure 4, lower part). The literature on malaria biology, physiopathology and medicinal chemistry is, at least in significant parts, stored as unstructured texts that make an invaluable source of knowledge which access depends on advances in terminology analysis and term extraction methods. A first attempt at using terminology analysis for the "harvesting" of relevant concepts in a defined disease area has been undertaken in another biomedical grid project, the recently started EU Integrated Project @neuRIST 190. In this project, terminology analysis lead to the refinement of a disease-specific text corpus and provided a shortlist of relevant terms. Moreover, not only the genes associated with the defined disease area could be identified by automated methods, but also single nucleotide polymorphisms (SNPs) published for these disease-associated genes, could be automatically identified (Dr. Laura Furlong, IMIM, Barcelona, personal communication). An important focus of future activities in this project is the construction and validation of fine-granular disease-specific ontologies. This concept can easily be adopted for a knowledge base on any disease, including malaria. Existing databases can be complemented by this automatically generated semantic layer, which subsequently would also be helpful for data mediation. Moreover, a structured knowledge space would produce grid services for indexing of distributed data resources and thus improve navigation through knowledge and retrieval of relevant information.

Finally, present and future in silico information must be supported and validated by data gathered in vitro and in vivo. As other X-omic strategies approach, the necessity of in silico mining is unquestionable; it is also susceptible to generate a huge amount of theoretical data that will need years to be confronted to the experimentation. The challenges in this domain remain severe. Difficulties of cloning and expressing parasite proteins in heterologous systems, the validation of druggable targets using siRNA etc, and the experimental assignment of functionality to the many hypothetical proteins, will occupy scientists in parallel for some time to come. Data generated from in silico analysis leads to a need for further laboratory work, such as the in vitro testing of ligands identified as potential drug leads through the WISDOM project. A strong linkage between scientists undertaking in vitro research and in silico researchers is therefore essential to support an iterative approach to knowledge generation and analysis, in the context of malaria chemogenomics.