Is It Time to Trade “Wet-Work” for Network?

Abstract

Systems biology aims to create a predictive mathematical model of the living organism. Such a goal may seem to be an exciting challenge to those with a quantitative background but it often sounds like a preposterous idea to others with less affinity to computers. Therefore, a talk on this topic should start with presenting several concrete examples on how systems biology may aid solving both practical and theoretical problems in biomedical research. These will include the difficulties associated with finding combinatorial therapeutic targets and a related theoretical problem of understanding the robustness of genetic networks in order to facilitate efficient drug design. The limitations of massively parallel biological data acquisition will be also discussed in order to provide an assessment of the data quality quantitative approaches will rely on [1,2].

The basic steps of modeling, such as data collection, reverse engineering, forward simulations, parameter optimization, are well known but each of these steps is associated with specific issues due to the fact that the object of the study is a highly complex, heterogeneous, adaptive system. There is no doubt that recent interest in systems biology has been ignited by the development of massively parallel measurement techniques in molecular biology, such as microarray chips. It is all the more important to point out that current modeling efforts can make little use of microarray data due to their well documented lack of accuracy [2]. Large-scale measurements, unless they significantly improve, can be used only in probabilistic models, where independent validation will be required before inclusion into reliable models. The noise level of measurements can be used to estimate the information content of these data sets, which in turn will impose limitations on modeling itself. The dichotomy of high throughput low quality measurements versus low throughput, high quality measurements will be discussed in terms of actual applications to modeling efforts.

Reverse engineering spans from whole genome sequencing projects to parameter optimization of dynamic models. A major difficulty in reverse engineering human genetic networks is the fact that at present functional annotations exist for less than half of all human genes, and most of the existing annotations are incomplete. It seems, however, that in microorganisms unknown functional annotations can be correctly guessed with a reasonable probability exploiting phylogenetic profiles, sequence homology and protein interactions databases [3]. This is based on the fact that proteins that share function tend to be present in the same microorganisms, tend to interact and tend to have sequence similarity. We are in the process of extending this method for human proteins. We have created a microarray based large scale (~10,000 genes) gene expression database on a wide variety of primary human cell lines. This will provide a “differentiation profile” of human proteins, analogous to the phylogenetic profile of genes in microorganisms. This information will be combined with other “knowledge data bases” in order to produce probable functional annotations.

Reverse engineering from time-series data for a genetic network with a given complexity has a well-defined information requirement. These estimates, however, should be adjusted according to the specific characteristics of data derived from genetic networks [1]. Methods developed by the AI community are widely used in systems biology. Evolutionary algorithms, for example, are exploited both in parameter optimization of large data driven models and gene expression based classification of various disease states. As an example, our recent work on the applications EA algorithms to breast cancer classification will be discussed.

Finally, understanding robustness is one of the deeper intellectual challenges of systems biology with several practical implications. Fundamental concepts associated with robustness are well known to the AI community. I will, therefore, present some surprising results derived from breast cancer associated gene expression measurements that can be best explained in terms of abstract concepts such as “attractors”.

Taken together, I will argue as a practicing biologist that, although with due caution, it is time to add computational “genetic network” approaches to traditional, “wet-lab” biology.