Plectix Blog

Getting Over the Hump

By Isha Antani

For many experimental biologists, the field of computational biology is a vast and unexplored terrain. Sure, there is curiosity about this exotic land, especially now with the increasing overlap between computational biology and the rediscovered systems biology. But, when asked to take the opportunity to embark on an expedition to explore computational biology, most are not interested.

It is not that they do not appreciate what the field has to offer. It is more that the community has had success through its traditional laboratory techniques and corresponding tried-and-tested analysis tools. Trained from their university days to utilize tools such as Microsoft Excel to digest data sets generated from Western blots and such, most experimental scientists cannot find strong reasons to get into a whole new field.

Furthermore, there is a distinct disconnect brought on by computational tools. Computational biology, which is a convergence of three distinct concentrations — biology, computer science and mathematics — often utilizes tools that require programming experience and is thus favored by biological scientists with solid foundations in some aspect of programming. Experimental biologists not trained to approach science from a programming perspective often find computational tools tasking, requiring precious time to learn and master. Additionally, translating laboratory data into the unfamiliar and unfriendly languages used by computational tools (e.g., ODEs, scripts, etc) is a tricky hill to surmount.

For computational biology to be properly integrated into the mainstream scientific process, there need to be better ways to merge the existing synergy between current biology and the new computational arm. Learning new techniques and transforming schools of thought requires time, which many scientists do not have. Perhaps then, the best way to better integrate computational biology with traditional laboratory approaches is to incorporate some computational curricula at an earlier stage in a scientist’s education. This can be done, for example, during the days of undergraduate studies, when no particular methodology has been engraved into the routine and the fertile mind is in the stage of gaining an understanding of the many scientific fields.

Systems Biology in a Nutshell

By Joshua Havumaki

As proof of concept applications such as Merrimack’s MM-121 (an ErbB3 inhibitor) slowly make their way into the mainstream, the systems biology community is continuing to grow and people are starting to appreciate the necessity of computational methods to model and explain complex biological systems. 

Within the community, computer scientists and mathematicians are rapidly developing methods for almost every step of the modeling process. Experimentalists are starting to dabble in small representative models consisting mostly of a few reactions or differential equations. In addition, certain community standards like SBGN (Systems Biology Graphical Notation) and SBML (Systems Biology Markup Language) are evolving and improving. However, in my mind, there are still a few problems that elude the explosion of the community and prevent it from unifying.

Computer scientists and mathematicians are developing methods to improve and/or redefine the modeling workflow. A quick browse through the CMSB (Computational Methods in Systems Biology) 2009 program provides a cross-section of methods ranging from the very theoretical (e.g. Computing Reachable States for Nonlinear Biological Models) to the directly applied (e.g. Improved Parameter Estimation for Completely Observed Ordinary Differential Equations with Application to Biological Systems). One major issue seems to be that most methods presented use a limited pool of applications and examples (e.g. the Raf/MEK/ERK cascade) and therefore rarely come to life for the biological community.

Additionally, it is a great undertaking for the average experimentalist to build an accurate, comprehensive model of a biological system. Commonly, experiments are not designed with specific mechanistic questions in mind. An experimentalist may ask whether a protein “activates” another protein, while a modeler must know if the interaction is direct and how the activation occurs. There is also a lack of specific parameter and protein abundance information in the literature, which is necessary for building biologically accurate models. Since experimentalists mostly think about biological systems from a higher level of granularity (e.g. not detailed mechanistically) the learning curve to adopt these methods for them may be too great to bear.

Standards like SBML and SBGN have yet to achieve widespread adaption by the community. One reason is the actual necessity and design of these standards don’t address the major pitfalls encountered by those trying to model complex systems. For example, SBGN doesn’t scale well with size and complexity.  This often results in trading comprehensively and accuracy for usability and understanding.

Lastly, biology, mathematics and computer science all have biases concerning scientific method and experiments. In math and computer science, even the most theoretical and abstract concepts can be explored while theoretical biological questions are impeded by money, time and technology. These problems, in conjunction with a lack of consensus on the definition of systems biology, result in a field that is fragmented and is yet to come out of its shell.

From Databases to Models?

By Ty Thomson

There are currently hundreds of online resources for biology, including countless databases containing information on proteins, protein-protein interactions, signaling pathways, and other biological processes.  In a recent perspective in Molecular Systems Biology, Bauer-Mehren et al. asked how easily the information from these databases can be harnessed and combined to produce a more complete picture of the cell, and perhaps even to produce computational models.  They found that although the information in many pathway databases is accurate and relatively comprehensive, there are still many challenges to merging information from multiple resources to construct computational models.

One challenge that they identified is that the scope of models greatly depends on the purpose of the model, therefore the modeler must carefully decide on the model boundaries in order to know what information is pertinent.  However, in order to decide on the relevant model boundaries, the modeler must know the relevant information.  This catch-22 means that information retrieval must be done in concert with evaluating the information in order to decide on appropriate model boundaries and to construct a useful model.  For example, while Bauer-Mehren et al. were able to automatically extract a number of signaling components and reactions in the EGFR pathway from the Reactome database, many components and reactions that the authors expected were not retrieved.  They had to make additional database queries to complete the model that they were constructing.  In this case, Bauer-Mehren et al. had carefully defined the boundaries of the model ahead of time by attempting to recapitulate the contents of a previous model.  If they had been constructing a model from scratch, they would not necessarily have known what components or reactions were missing without attempting fishing expeditions of additional queries.

Another challenge faced by the authors is that information in these online resources is often at differing levels of detail, even within the same resource.  For example, in the Reactome database, some reactions consider ERK1 and ERK2 independently, whereas others mention only a combined ERK1/ERK2 entity.  Unfortunately, the link between these species is not explicitly defined in the database.  Better semantic tagging of species and reactions in these resources could help alleviate these problems. 

Interestingly, these challenges are not unique to building a model from information in databases and online resources – modelers themselves face these same issues when building models from the primary literature.  To me, this demonstrates that although these resources may facilitate model construction, they are currently not in a position to revolutionize the careful manual construction of models.  Certainly, the fact that about half of the reactions and species in the EGFR pathway that the authors expected to find were not retrieved at all, even after additional database queries, means that literature review must remain a vital part of model building.  So despite the large amount of content in online pathway databases, the incompleteness and differing resolution of the data means that model building will likely continue as a manual process, at least for the foreseeable future.