reStructured

============= Interactomics =============

Abstract

How to integrate the vast amount of interaction information out there in order to reveal biological processes.

A meaningfull integration seems to be the critical part in the game as well as the evaluation of the quality and reliability of the information from different data sources. It might required to consult biophysics and modeling of precise kinetic data.

Introduction

Conventional network-approaches use qualitative (binding or no binding) rather than quantitative data. Even an oversimplified model of interaction in binary terms (0 or 1) is already very predictive. For instance, with a straightforward guilt-by-association method we achived a predicitve power of 8/9 true positives. Of, course, there is a lot of room for further improvements as always. Also, the interactome of, for instance, mammals is not so deeply covered yet and definitely requires innovative ways of tackling this problem.

In intruiging question is whether 3D-structures of individual molecules could add any value to it. There are huge amount of solved structures already available and its a fact that only a limited amout of unique folds are actually invented by nature, as nature is very economically. It should also. It should also be possible to get the 3D structure of a protein from its primary sequence (theoretical), though this is migh still be quite tricky yet. There was very little attempt so far to integrate structural data with interaction information, which is actually a logical next step! In principle its quite simple: if the surfaces of two proteins fit together they should tend to interact more likely.

Also subcellular location should be taken into account and expression as well. If two genes are not expressed at the same time in the same cell and compartment (i.e. space), there should be no interaction.

Consequently, instead of just black and white, we could have different colours between two entities describing the strength and likelihood of an interactions. 0 would be again no binding (i.e. no expression together, or different compartments), 1 would be molecules which are always together (maybe due to covalent bonds). Then we have a spectrum of different binding values in-between. We could simply add different types of evidence to it.

With respect to 3D-structural information, there are methods in which charge-patterns of protein-surfaces are calculated and used to predict protein-protein interactions.' Machine learning algorthim for example can be used to interrogate this [22701576 ].

Structural data is probaly the best source for getting some form of binding-strength into interactions (in a scalable manner). For enzyme-metabolits there are also Michaelis menten parameters, but it is questionable if they can b used on a large scale.

The Reactome data model of interaction is by describing them as reversible reactions: A + B => C AND C => A + B

Combining with high-performance supercmoputing sources. dynamic 3D modelling. made it possible to virtually screen more than 8 million potential drug compounds. ZINC is a free database of more than 21 million commercially available compounds which can be used for virtual screening [http://www.biocompare.com/Life-Science-News/115579-Moving-3D-Computer-Model-Of-Key-Human-Protein-Is-Powerful-New-Tool-In-Fight-Against-Cancer/] 22647192].

There are two major approaches to predict 3D protein structures: 1) sequence homology based prediction and 2) ab initio (or de novo) method. Sequence homology approach uses sequence alignment to identify best matching 3D structure fo different components and threads them to predict the overall 3D structure. In contrast, ab initio method is based on energy minimization principle and predicts the whole structure soley from the sequence. Improved ab initio methods integrate the biochemical and biophysical properties [http://www.biology-online.org/articles/bioinformatics_microbial_biotechnology/3d_structure_modeling_docking.html].