Manual Annotation of Protein Interactions

Abstract

Protein interactions are the basic building blocks for assembly of pathways and networks. Almost any biologically meaningful functionality (for instance, linear signaling pathways, chains of metabolic reactions, transcription factor dimmers, protein complexes of transcriptosome, gene-disease associations) can be represented as a combination of binary relationships between "network objects" (genes, proteins, RNA species, bioactive compounds). Naturally, the assembled pathways and networks are only as good as their "weakest" link (i.e., a wrongly assigned interaction), and the errors multiply in multi-step pathways. Therefore, the utility of "systems biology" is fundamentally dependent on quality and relevance of protein interactions. The second important parameter is the sheer number of interactions assembled in the database. One needs a "critical mass" of species-specific interactions in order to build cohesive networks for a gene list, not a constellation of non-connected proteins and protein pairs. The third issue is semantic consistency between interactions of different types. Transient physical signal transduction interactions, reactions of endogenous metabolism, transcription factor-promoter binding, and kinetic drug-target interactions are all very different in nature. Yet, they have to fit well into one database format and be consistent in order to be useful in reconstruction of cellular processes.High-quality protein interactions are available in peer-reviewed "small experiment" literature and, to a much smaller extent, patents. However, it is very challenging to find the interactions, annotate with searchable (and computable) parameters, catalogue in the database format in computer readable form, and assemble into a database. There are hundreds of thousands of mammalian interactions scattered in tens of thousands of papers in a few thousands of scientific journals. There are no widely used standards for reporting the interactions in scientific texts and, therefore, text-mining tools have only limited applicability. In order to generate a meaningful database of protein interactions, one needs a well-developed technology of manual curation, equipped with computational solutions, managerial procedures, quality control, and users' feedback. Here we describe our ever-evolving annotation approach, the important annotation issues and our solutions, and the mammalian protein interactions database MetaBase which we have been working on for over 8 years.

Systems-Based Analyses of Brain Regions Functionally Impacted in Parkinson's Disease Reveals Underlying Causal Mechanisms

“…A total of seven proprietary and publicly available ontologies were extracted from the MetaBase resource [69]: 1) Pathway maps represent a portfolio of over 1,000 canonical, disease-specific and toxicity-specific pathways. The goal of each pathway map is to show the cellular response to particular stimulus.…”

“…Expression regulators correspond to those proteins that are responsible for the observed gene expression changes. All network-based analyses are based on the MetaBase resource, a manually curated interaction database [69].…”

The PluriNetWork: An Electronic Representation of the Network Underlying Pluripotency in Mouse, and Its Applications

“…Nevertheless, we see a lot of value in hand-curated network data, which is complementary to networks based on machine learning or text mining. As pointed out by Bureevas et al [23], an advantage of manual curation is accuracy. Since there exist no standards for reporting an interaction or a regulation link in a paper, a human curator is best suited to understand the precise semantics of the textual descriptions provided by authors.…”

“…The advantages and disadvantages of this approach have been discussed e.g. by Bureeva et al [23] and we will discuss them towards the end of the paper.…”

The Role of Historical Bioactivity Data in the Deconvolution of Phenotypic Screens

“…The Ingenuity IPA database (Redwood City, CA) 21 was first queried to identify genes whose complete or partial deletion in mouse models leads to a modification of TNF-α levels. GeneGO MetaCore (Philadelphia, PA) 22 was then mined to identify proteins upstream of TNF-α. The linear canonical pathways (noodles) containing TNF-α were extracted from MetaBase.…”