Past, Present, and Future of Chemogenomics.

Abstract

Chemogenomics as a strategy started more than a decade ago with the idea to use the freshly published human genome to facilitate the identification of novel drugs for the newly discovered genomic targets. Since then, the field has advanced tremendously, enabled by the largescale availability of small molecule-bioactivity data, more biological omics data, and the development of various computational tools to mine these data and generate novel ligand-target hypotheses. What started as a protein family strategy for efficient ligand design, has now advanced to a broad toolbox including e.g. repurposing of known drugs, prediction of targets and pathways for hits from phenotypic screening, analyzing and designing polypharmacology, the prediction of adverse drug effects, or screening biologically relevant compound selections. From the early beginnings on, Chemogenomics was a discipline with a strong interaction between experimental and computational methods. For example it is not feasible to screen very large numbers of compounds against very large numbers of targets. This gap creates an opportunity for compound-target prediction algorithms for prioritizing compounds or targets for experimental validation. Another key opportunity for fruitful interactions between computational and experimental methods can be found in the area of analyzing the very large and complex data matrices of many compound activities against many targets and target families. This is not a trivial task, requiring computational methods to make these data available in a form that is accessible to humans, so it can be further used to draw conclusions and generate new hypotheses. Finally, this wealth of data is also a genuine resource for interactions with other scientific fields like systems biology, structural bioinformatics, and computer science, among others. In this special issue in Molecular Informatics, we want to focus on the computational aspects of Chemogenomics, review the current stage of the field, highlight promising directions, and discuss the future of the field. In the selection of topics we have focused on five major themes: the underlying Chemogenomics data, advanced compoundtarget interaction prediction algorithms, visualization of Chemogenomics data, other new developments, and expert opinions on the future of Chemogenomics. Since Chemogenomics is an active field of research we have included a number of reviews from scientists that outline their own recent work concerning Chemogenomics. The wordle in Fig. 1 was generated from the full texts of this special issue. The figure nicely reflects the main topics in Chemogenomics: “analysis” and “prediction” of “interactions” (“activity”, “interactions”, “binding”, “similarity”) between ligands (“compound”, “drug”, “chemical”) and targets (“target”, “protein”). Chemogenomics strongly depends on large sets of “data”, e.g. ligand-target data from “ChEMBL” or “DrugBank”. The first section of this special issue contains two contributions on Chemogenomic data sources. While the impact of having by now large and openly accessible Chemogenomics data sources cannot be emphasized enough, it should also be acknowledged that we have a growing and experienced user community that is willing to share their experiences. In this first section, Southan et al. compare the content of various Chemogenomics data sources and Kallikoski et al. discuss possible sources of errors in compound target activity databases. For the second section on the prediction of ligandtarget interactions, we choose to invite three author teams that work on integrating additional -omics information for the improvement of ligand-target interaction prediction. This reflects the major importance of this topic to Chemogenomics, and highlights the diversity of approaches and experiences from multiple authors. The review of Brown et al. describes the integration of ligand and target similarity to the kernel based prediction of compound-target interactions, and how such approaches could be used also for cross-family predictions and extracting features responsible for the interactions. The group of Eric Martin has generated one of the most advanced toolkits for the prediction of kinase inhibitors that makes extensive use of 2D and 3D target-family information and structure activity relationships of ligands, that enables the prediction of activities for kinases without known ligands. Figure 1. Wordle of the full text of all contributions of this special issue.