The Digital Biologist

Literature Update: Synthetic biology applied to the engineering of tissue homeostasis

In synthetic biology, the challenges to engineering individual cells to do useful things are already significant given the monstrous complexity of even the simplest of cells. Now imagine trying to do enginering at the individual cell level, in order to modulate the bulk properties of a population of such cells working cooperatively as tissue in vivo. That is exactly the problem that a bioengineering research group at MIT set out to tackle, as described in their PLOS One paper that appeared this month.

The applications of this problem are clinically and economically significant - a fact amply illustrated by this team's decision to focus upon the challenges around creating synthetic tissue for transplant into patients with diabetes. The central problem described in this paper, is the creation of engineered stem cells capable of maintaining a stable population of insulin-producing β-cells in a diabtetes patient. Tissue homestasis in this case, is defined by the authors as "the property of balancing growth, death, and differentiation of multiple cell-types within a multicellular community". 

 The authors found that the design of such a system defies any simple, intuitive thinking, requiring a multifactorial approach that takes into acount a number of parameters governing the equilibrium between the population of self-renewing stem cells and the steady state population of adult, insulin-producing β-cells. Far from optimizing a system of synchromized, homogeneous cells, the best solutions proposed by the mathematical model were characterized by the optimization of the heterogeneity of the cell population against the required steady state behavior - a key component of this approach being a phenotypic sensitivity analysis to determine how the behaviors of the various synthetic functional module contribute to the performance of the system as a whole.

The full text of the paper is available online at the NCBI.

  © The Digital Biologist | All Rights Reserved

Literature Update: Modeling potential toxicity for kinase inhibitors

Here is a 2010 paper that somebody shared with me recently - a really nice example of the application of computer modeling to real world problems in drug disocvery and development. A research group at the University of Chicago have developed a computer model of the realtionship between adverse events that occur during kinase inhibition therapy, and the actual kinase targets. In addition to predicting sets of clinically established adverse events (such as diarrhea and rash in patients treated with ERGFR inhibitors), the model was also able to predict sets of adverse events that had been documented, but that had not been so widely reported in publicly available clinical data.

Here is the abstract for the paper:

"Kinase inhibition is an increasingly popular strategy for pharmacotherapy of human diseases. Although many of these agents have been described as “targeted therapy”, they will typically inhibit multiple kinases with varying potency. Pre-clinical model testing has not predicted the numerous significant toxicities identified during clinical development. The purpose of this study was to develop a bioinformatics-based method to predict specific adverse events (AEs) in humans associated with the inhibition of particular kinase targets (KTs)."

The full text of the paper is also available online at the NIH.

 © The Digital Biologist | All Rights Reserved

Biologists Flirt With Models

This is an updated version of an article that I published in 2009. Alas in the 3 years or so that have passed since I wrote it, little seems to have changed beyond the fact that the crisis in the pharmaceutical industry has deepened to the point that even the biggest companies in the sector are starting to question whether their current business model is sustainable.

Introduction

The enormous challenge posed by the complexity of biological systems represents a potential intellectual impasse to researchers and threatens to stall future progress in basic biology and healthcare. In recent years, increasing reliance on correlative approaches to biology has failed to resolve this situation. The burgeoning volumes of laboratory data gathered in support of these approaches pose more questions than they answer until such time as they can be assimilated as real knowledge. Modeling can provide the kind of intellectual frameworks needed to transform data into knowledge, yet very few modeling methodologies currently exist that are applicable to the large, complex systems of interest to biologists. Early developments in biological modeling, driven largely by the convergence in systems biology of complementary approaches from other disciplines, hold the potential to forge a knowledge revolution in biology and to bring modeling into the mainstream of biological research in the way that it is in other fields. This will require the development of modeling platforms that allow biological systems to be described using idioms familiar to biologists, that encourage experimentation and that leverage the connectivity of the internet for scientific collaboration and communication.

The challenge of biological complexity

Researchers in biology and the life sciences have yet to embrace modeling in the way that their peers in the fields of physics, chemistry and engineering have done. There are some very good reasons for this, not the least of which is that the systems of interest to biologists tend to be far more refractory to modeling than the systems that are studied in these other fields. It is much more straightforward, for example, to capture in lines of computer code the orbit of a satellite or the load on a suspension bridge than it is to build a computational model of even the simplest and most well-studied of biological organisms. At the molecular level the fundamental processes that occur in living systems can also be described in terms of physics and chemistry. However, at the more macroscopic scales at which these systems can be studied as “biological” entities, the researcher is confronted with an enormous number of moving parts, a web of interactions of astronomical complexity, significant heterogeneity between the many “copies” of the system and a degree of stochasticity that challenges any intuitive notion of how living systems function, let alone survive and thrive in hostile environments. Biology is, in a word, messy.

Complexity seems to be the mot du jour in biology right now and the field arguably stands at a crossroads from which significant future progress will largely depend upon the ability of its researchers to at least tame the monster of complexity, if not to master it. Driven by the rapid development of the technology available to scientists in the laboratory, the quantity, scope and resolution of biological data continues to grow at an accelerating pace. Invaluable though this burgeoning new wealth of data may be, it often poses more questions than it really answers until such time as it can be assimilated as real knowledge. The expectations for the Human Genome Project for example were as huge as the tsunami of data that it generated, spawning a whole new genomics industry within the life sciences sector and even moving some leaders in the field to make predictions about cures for cancer within a few years. Looking back over almost a decade since the first working draft of the human genome was completed - while there have certainly been some medical benefits from this work, I think it is fair to say that the impact has not been on anything like the scale that was initially anticipated, largely as a result of having underestimated how difficult it is to translate such a large and complex body of data into real knowledge. Even today, our understanding of the human genome remains far from complete and the research to fill the gaps in our knowledge continues apace. The as yet unfulfilled promise of genomics is also reflected in the fact that many of the biotechnology companies that were founded with the aim of commercializing its medical applications, have disappeared almost as rapidly as they arose. This is not to say that the Human Genome Project was in any way a failure - quite the opposite. In its stated goals to map out the human genome it succeeded admirably, even ahead of schedule, and the data that it generated will only become more valuable with time as we become better able to understand what it is telling us. The crucial lesson for biology however, is that as our capacity to make scientific observations and measurements grows, the need to deal with the complexity of the studied systems becomes more not less of an issue, requiring the concomitant development of the means by which to synthesize knowledge from the data. Real knowledge is much more than just data - it does not come solely from our ability to make measurements, but rather from the intellectual frameworks that we create to organize the data and to reason with them.

Models: Frameworks to turn data into knowledge

So what are these intellectual frameworks? Modeling is one of the most powerful and widely used forms of intellectual framework in science and engineering . There is a tendency to think of models only in terms of explicit modeling - the creation of tangible representations of real world objects such as a clay model of a car in a wind tunnel or a simulation of a chemical reaction on a computer for example. In actuality however, scientists of all persuasions are (and always have been) modelers, whether or not they recognize this fact or would actually apply the label to themselves. All scientific concepts are essentially implicit models since they are a description of things and not the things themselves. The advancement of science has been largely founded upon the relentless testing and improvement of these models, and their rejection in the case where they fail as consistent descriptions of the world that we observe through experimentation. As in other fields, implicit models are in fact already prevalent in biology and are applied in the daily research of even the most empirical of biologists. Every experimental biologist who constructs a plasmid or designs a DNA primer for a PCR reaction, is working from an implicit model of the gene as a dimer of self-describing polymers defined by a 4-letter alphabet of paired, complementary monomers. Interestingly, the double-helical structure of DNA published by Watson and Crick in 1953, remained essentially just a hypothetical model until it could actually be observed by x-ray crystallography in the 1970s, yet it earned them a Nobel prize a full decade earlier as a result of the profound and wide-ranging synthesis and integration of knowledge that it brought to the field. 

It is no accident that the majority of the successes that explicit modeling approaches have enjoyed in biology tend to be confined to a rather limited set of circumstances in which already established modeling methodologies are applicable. One example is at the molecular level where the quantitative methods of physics and chemistry can be successfully applied to objects of relatively low complexity (by comparison with even a single living cell). Modeling approaches can also be successfully applied to biological systems that exhibit behavior that can be captured by the language of classical mathematics, as for example in the cyclic population dynamics of the interaction of a predatory species with its prey, expressed in the Lotka-Volterra equations. Unfortunately for the modern biologist, many if not most of the big questions in biology today deal with large, complex systems that do not lend themselves readily to the kind of modeling approaches just described. How do cells make decisions based upon the information processed in cell signaling networks? How does phenotype arise? How do co-expressing networks of genes affect one another? These are the kinds of questions for which the considerable expenditure of time, effort and resources to collect the relevant data typically stands in stark contrast to the relative paucity of models with which to organize and understand these data.

Correlation versus causation

One response to the challenges posed by biological complexity in both academia and the healthcare industry has been the pursuit of more correlative approaches to biology in which an attempt is made to sidestep the complexity issue altogether by focusing (at least initially) on the data alone. If the data are measured carefully enough and can be weighted and scaled meaningfully with respect to one another, parametric divergences that can be detected between similar biological systems under differing conditions may reveal important clues about the underlying biology as well as identify the critical components in the system. This is very much a discovery process that aims in effect to pinpoint the needle in the haystack from the outside, without all the mess and fuss of having to get in and pull the haystack apart. These phenomenological approaches have been in widespread use for some time in both academia and industry in the form of genomic and proteomic profiling, biomarker discovery and drug target identification, and have also given rise to a plethora of high-throughput screening methodologies. 

While these correlative approaches have enjoyed some successes, they do tend to compound the central problem alluded to earlier of generating data without knowledge. Moreover, their limitations are now starting to become apparent. In recent years for example, steadily intensifying investment and activity in these areas by drug companies has seen the approval rate of new drugs continue to fall as levels of R&D investment soar. The field of biomarkers has also seen a similar stagnation, despite years of significant investment in correlative approaches. Cancer biomarkers are a prime example of this stagnation. Since we don’t yet have a good handle on the subtle chains of cause and effect that divert a cell down the path towards neoplasia, we are forced to wait until there are obvious alarm bells ringing, signaling that something has already gone horribly wrong. The ovarian cancer marker CA125 for example, only achieves any significant prognostic accuracy once the cancer has already progressed beyond the point at which therapeutic intervention would have had a good chance of being effective. To use an analogy from the behavioral sciences, broken glass and blood on the streets are the "markers" of a riot already in progress but what you really need for successful intervention are the early signs of unrest in the crowd before any real damage is done. The lack of new approaches has also created a situation in which many of the biomarkers in current use are years or even decades old and most of them have not been substantially improved upon since their discovery. Much more useful than the current PSA test for prostate cancer for example, would be a test with which a physician could confidently triage prostate cancer patients into groups of those who would probably require medical intervention and those whose disease is relatively quiescent and who would likely die of old age before their prostate cancer ever became a real health problem.

Given the general lack of useful mechanistic models or suitable intellectual frameworks for managing biological complexity, the tendency to fall back on phenomenology is easy to understand. Technology in the laboratory continues to advance, and the temptation to simply measure more data to try to get to where you need to be, grows ever stronger as the barriers to doing so get lower and lower. We should also not underestimate the bias in the private sector towards activities that generate data, since they are far more quantifiable in terms of milestones and deliverables than the kind of research needed to translate that data into knowledge; this also makes these data-generating activities much more amenable to the kinds of R&D workflow management models that are common in many large biotechnology and pharmaceutical organizations. In effect what we have witnessed in biology over the last decade or so is a secular movement away from approaches that deal with underlying causation, in favor of approaches that emphasize correlation. However, true to the famous universal law that there’s no such thing as a free lunch, the price to be paid for avoiding biological complexity in this way is a significant sacrifice with respect to knowledge about mechanism of action in the system being studied. Any disquieting feeling in the healthcare sector that it is probably a waste of time and money to simply invest more heavily in current approaches is perhaps the result of an uneasy acknowledgement that much of the low-hanging fruit has already been picked and that any significant future progress will depend upon a return to more mechanistic approaches to disease and medicine. 

Modelling biological systems

Having arrived at such an intellectual impasse, there has arguably never been a more opportune time for biologists to incorporate modeling into their research. Biological modeling has to date tended to be almost exclusively the realm of theoretical biology, but as platforms for generating and testing hypotheses, models can also be an invaluable adjunct to experimental work. This is already well recognized in other fields where modeling is more established in the mainstream of research. Astronomers for example, use computer-based gravitational models to accurately position telescopes for observations of celestial objects. and civil engineers routinely test the load-bearing capacity of new structures using computers, before and during their construction, comparing the simulation results with experimental data to ensure that the limits of the structure’s design are well understood. 

One misconception that is common amongst scientists who are relatively new to modeling is that models need to be complete to be useful. Many (arguably all) of the models that are currently accepted by the scientific community are incomplete to some degree or other, but even an incomplete model will often have great value as the best description that we have to date of the phenomena that it describes. Scientists have also learned to accept the incomplete and transient nature of such models since it is recognized that they provide a foundation upon which more accurate or even radically new (and hopefully better) models can be arrived at through the diligent application of the scientific method. Physicists understand this paradigm well. For example, the discrepancies that astronomers observe trying to accurately map the positions and motions of certain binary star systems to the standard Newtonian gravitational model, can actually provide the means to discover new planets orbiting distant stars. The divergence of the astronomical observations from the gravitational model can be used to predict not only the amount of mass that is unaccounted for (in this case, a missing planet), but also its position in the sky. Models therefore, can clearly have predictive value, even when they diverge from experimental observations and appear to be “wrong”. 

For models to be successful in fulfilling their role as intellectual frameworks for the synthesis of new knowledge, it is essential that the chosen modeling system be transparent and flexible. If these requirements are satisfied, a partial model can still be of great value since the missing pieces of the model are a fertile breeding ground for new hypotheses and the model itself can even provide a framework for testing them. What is meant by transparent and flexible in this case? Transparency here refers to the ease with which the model can be read and understood by the modeler (or a collaborator). Flexibility is a measure of how easily the model can be modified, for example when new knowledge becomes available or existing knowledge turns out to be wrong. A model that is hard to read and understand is also difficult to modify and, very importantly in this age of interconnectedness, difficult to share with others. The importance of this last point cannot be overstated since one of the most often ignored and underestimated benefits of models is their utility as vehicles for collaboration and communication. A model that must be completely rewritten in order to add or remove a single component is not flexible and once built, there exists a significant barrier to modifying it that discourages experimentation. Flexibility also implies that the models should be capable of being built in a simple, stepwise fashion, based only upon what is known - free of any significant constraints on scope and resolution or any a priori hypotheses about how the assembled system will behave. The co-existence of these last two properties is essential for flexibility in a modeling system, since decisions about which components of a model to omit in order to keep it within workable limits (e.g. for the purposes of memory, storage or execution), must of necessity be predicated upon some pre-existing notions of how the assembled system will behave and which components will be critical for that behavior. It is interesting to note that biological models based upon classical mathematical approaches generally fall far short of these ideals with respect to both transparency and flexibility.

A cornerstone of any modeling approach should also be the recognition that biological systems are dynamic entities - in fact “biology” is essentially the term that we apply to the complex, dynamic behavior that results from the combinatorial expression of their myriad components. For this reason, models that can truly capture the “biology” of these complex systems are also going to need to be dynamic representations. This entails a big step beyond the computer-generated pathway maps that biologists working in cell signaling are already starting to use. These pathway maps are essentially static models in which the elements of causality and time are absent. They are a useful first step, but just like a street map, a pathway map does not provide any information about the actual flow of traffic. An ideal modeling platform would offer a “Play” button on such maps, allowing the biologist to set the system in motion and explore the its dynamic properties. Notwithstanding these requirements for transparency, flexibility and dynamics, the biology community in large part are unlikely to adopt modeling approaches that require them to become either mathematicians or computer scientists, irrespective of how much math and/or computer science there might be “under the hood”. How fewer cars would there be on our roads if all drivers were required to be mechanics? Ideally biologists will be able to couch their questions to these new frameworks in an idiom that approximates the descriptions of biological systems that they are used to working with - and in a similar fashion, receive results from them whose biological interpretation is clear. Finally, let us not forget that thanks to the internet, we live in an era of connectivity that offers hitherto unimaginable possibilities for communication and collaboration. The monster of biological complexity is in all likelihood, too huge to ever be tamed by any single research group or laboratory working in isolation and it is for this reason that collaboration will be key. With knowledge and data distributed widely throughout the global scientific community, a constellation of tiny pieces of a colossal puzzle resides in the hands of many individual researchers who now have the possibility to connect and to work together as never before, and to assemble a richer and more complete picture of the machinery of life than we have ever seen. Potentially huge opportunities might therefore be squandered if future modeling platforms are not purposefully designed around the connectivity and semantic web potential that the internet offers. It is interesting to reflect on a potential future in which the monster of biological complexity, the emergent property of a multitude of tiny interactions within living cells, is eventually tamed by an analogous emergent property of the semantic web - the concerted scientific efforts of an interconnected, global research community.

© The Digital Biologist | All Rights Reserved

Update: The truly staggering cost of inventing new drugs

One of the areas of application for digital biology that most interests me, is in the discovery and development of new drugs. On The Digital Biologist last year, I published an article entitled A new prescription for the pharmaceutical industry? in which I quoted some oft-used estimates of the time and cost of bringing a new drug to market. These estimates at that time, stood at something like a decade or more for research and development and a cost of around $1.2 billion. The numbers I used were based upon a study done by the Federal Trade Commissions's Bureau of Economics in 2008, that essentially reproduced the results of a study done a few years earlier by DiMasi et al.

Oh what a difference 4 years can make ...

After an intrepid reporter or two at Forbes revisited these numbers very recently and adjusted them using the latest data on drug failure rates at some of the largest pharma companies, an even gloomier picture has emerged of the true cost of drug development - and the new numbers are (as the title of the report states) truly staggering.

"The average drug developed by a major pharmaceutical company costs at least $4 billion, and it can be as much as $11 billion."*

In the article, the authors demonstrate that given recent drug failure rates, some companies are spending as much as $11 billion dollars on R&D for each drug that they successfully bring to market, with the "most successful" companies in the study still spending almost $4 billion per drug.

If these new numbers are an accurate reflection of the situation in the pharma industry right now - what has already looked for some years like an increasingly unsustainable business model for these large companies, would now seem to be a problem of potentially disastrous proportions and an urgent call for a new, more efficient and cost-effective way of doing pharamaceutical R&D.

You can see the full article and the table of R&D costs for each company featured in the report at Forbes online.

© The Digital Biologist | All Rights Reserved

Happy New Year for 2012

I hope that 2012 will be one of happiness, prosperity and fulfillment for digital biologists everywhere and a banner year for the field of digital biology in general. My very best wishes to all of of you.

An "Uncertainty Principle" for traditional mathematical approaches to biological modeling

If you're a biological modeler, chances are there are two words that keep you up at night and that on occasion, might even have given you serious pause to question the wisdom of your choice of profession.

Those two words are combinatorial complexity.

For anyone not entirely familiar with this concept, imagine one of the simplest possible biomolecular systems, with a kinase K that can bind to and phosphorylate a substrate S at either of two positions a and b, as shown in this first diagram. Even this very simple system can produce 13 possible molecular species: K unbound, S unbound in one of its 4 phosphorylation states, K bound at a with S in one of its 4 states, K bound at b with S in one of its 4 states.

Taking the traditional biological modeling approach of using ordinary differential equations (ODEs), you would therefore have to write 13 rate equations to describe this system. So far so good.

But now let's add the phosphatase P that dephosphorylates the sites a and b on S. If we do a similar analysis of the possible molecular species for our new system, taking into account now, the possible bound and unbound states of P and K on S, we discover that the addition of this single agent P yields 21 new molecular species in addition to the 13 that we already had! Furthermore, since we are working with a model of interdependent ODEs, we will also need to rewrite our original set of rate equations.

If it takes all this work to describe what is almost the simplest imaginable kind of system, how many equations would we need for a real biological system? How many rate equations would we need to describe the canonical epidermal growth factor receptor (EGFR) signaling pathway for example?

Hold on to your hats ... drum roll ... somewhere north of 10³⁰ equations.

Yikes!

All this said however, biological modelers have built models of complex cellular systems like the EGFR pathway, so how on earth have they done it? The answer is by simplifying the system, typically by either ignoring features that are presumed to minimally impact the system's behavior, or by aggregating features to create a less granular description of the system, again under the presumption that this will not significantly affect the model's behavior.

The danger inherent in such approaches is that they require a set of a priori hypotheses about what are and what are not, the important features of the system i.e. they require a decision about what aspects of the model will least affect its behavior before the model has ever been run.

The famous Uncertainty Principle that we all learned in high school physics states that it is impossible to simultaneously determine with any accuracy, both the position and the momentum of an electron. A recasting of this principle for traditional biological modeling might be

"Scope or resolution, but not both at the same time"

One could argue that whereas the original physical principle is absolute, in the case of biological modeling the limitation is one of technology - "If we had a big enough, fast enough computer ..." etc. Perhaps, but when you compare the storage and processing time required to solve a system of 10³⁰ equations with the scale of our universe, the biological modeling version of the principle seems pretty darn absolute to me.

Did I hear someone say "quantum computer"?

Just let me know when they've built one that could address this problem and I will gladly publish an update :-)

The author Gordon Webster, has spent his career working at the intersection of biology and computation and specializes in computational approaches to life science research and development.

© The Digital Biologist | All Rights Reserved

A new report by "The Economist" laments the lack of innovation in the biopharma sector

On "The Digital Biologist" a couple of months back, I published an article about a troubling trend in the pharmaceutical industry in which R&D investment costs are soaring even as the approval rate for new drugs actually declines. One of the reasons I posited for the apparent stagnation in this sector, was the inability of corporate leaders to abandon old practices and embrace innovation and new methodologies. "Do what you've always done and you'll get what you've always gotten".

Now a new report from "The Economist" (sponsored by Quintiles) provides some real numbers to back up this idea - and while it may not be a strictly scientific or comprehensive survey of the sector, it does provide real evidence of a kind of corporate inertia at the highest levels, with regard to embracing new ideas and changing course from the rocky road that the industry currently finds itself on.

I would encourage you to read this interesting and well written report in full, but here are some of its major findings in summary ...

Less than half of industry executives even have faith that their own R&D programs currently in place, are adequate to meet the needs of their companies.

As much of 70% of the global R&D budget for the sector may actually be wasted in fruitless programs.

Barely half of all companies who responded to the survey say that they are prioritizing change and innovation and this includes the companies who admit that the programs they do have in place are largely inadequate.

An inflexible corporate culture that fears change is cited as the leading impediment to improved innovation.

In a sector that is seen as being driven more by fear than by ambition, the companies that are doing well are those in which life science innovators are able to create a culture that recognizes and rewards effort rather than penalizing failure.

You can download a full copy of this report here

The author Gordon Webster, has spent his career working at the intersection of biology and computation and specializes in computational approaches to life science research and development.

© The Digital Biologist | All Rights Reserved

Papers, Scissors, Rock: A bacterial strategy game

As the old saying goes: "There's nothing new under the Sun" and indeed it turns out that Nature also invented "Paper, Scissors, Rock" many millennia before the first kids ever settled a schoolyard dispute using this game. In a recent PNAS article, researchers from the University of Washington, Seattle describe the survival dynamics of a community of three strains of bacteria that compete for resources.

In addition to a healthy, wild type strain, there is a second strain that produces the bacterial toxin colicin as well as a bunch of other proteins that allow it to survive its own toxic strategy (without which it would be shooting itself in the proverbial foot). A third strain has evolved resistance to the colicin protein produced by the second strain, by acquiring mutations that prevent colicin from binding to its own proteins.

However, in keeping with the maxim that there's no such thing as a free lunch, these latter two strains of bacteria pay a price for their competitive strategies. There is a significant cost in energy and resources for the bacteria that produces colicin and all of the accompanying proteins that enable it to survive its own offensive strategy. The colicin-resistant strain also pays a price insofar as the resistance mutations it needs to survive its competitor's colicin strategy, compromise to some degree, the functions of certain vital proteins that it also needs for its general well-being.

The scenario we have then is that the colicin producing strain can outperform the wild type bacteria in the absence of the colicin resistant strain. The colicin resistant strain can outperform the colicin producing strain when the wild type strain is not around. The wild type strain can outperform the colicin resistant strain when the colicin producing strain is not around. ... the bacterial equivalent of the "Paper, Scissors, Rock" game.

So what happens when all three strains are present at once? A very interesting dynamic of "survival of the weakest" emerges.

If one of the strains really starts to dominate the strain it grows better than, this actually favors its other competitor. If the toxin producing strain starts to dominate the wild type strain, this favors the growth of the toxin resistant strain. If the toxin resistant strain starts to dominate the toxin producing strain, this favors the growth of the wild type strain. If the wild type strain starts to dominate the toxin resistant strain, this favors the growth of the toxin producing strain.

"Paper, Scissors, Rock"

When the authors of the study cultured these three strains of bacteria in their experiments, this scenario is indeed what they observed in the laboratory. When all three strains were present in the same culture, the researchers consistently observed over many experiments, that no single strain emerges on top - the result of an interesting competitive scenario that actually encourages restraint amongst the players. The authors point out that this kind of competitive dynamic has also driven the evolution of behavioral traits in the natural world, such as mating.

While not exactly the same situation, this scenario kind of reminds me of that famous scene in the movie "A Beautiful Mind" in which John Nash, the great luminary in the field of game theory, has an epiphany in which he imagines himself and his colleagues, all competing for the attentions of the same pretty blonde. Sometimes it might be very much in everybody's best interests if disputes could be settled over a game of "Paper, Scissors, Rock" instead of a barroom brawl, a showdown with law enforcement or even a war.

The author Gordon Webster, has spent his career working at the intersection of biology and computation and specializes in computational approaches to life science research and development.

© The Digital Biologist | All Rights Reserved

A new prescription for the pharmaceutical industry?

The past two decades have been witness to a bleak and implacable trend in the pharmaceutical industry. Investment in pharmaceutical R&D has been rising while the number of new drugs getting approval has actually declined over the same period. In other words, the industry is spending ever more money only to see the returns on its investments get ever smaller. By recent estimates, the current average cost of bringing a new drug to market is about $1.2 billion. In 2007 at an AAAS science and technology forum,  William Haseltine, the founder of Human Genome Sciences said "What I see is an industry in very deep trouble.", supporting his statement by reference to the industry statistics compiled that year, showing that the annual number of new drug approvals was down substantially, despite a 10-fold increase in R&D spending over the same period.

To put this in perspective, we have reached the point at which the average pharmaceutical company needs its new drug to be a billion dollar molecule just to recover the costs of its development! So how on earth did it come to this?

As somebody much smarter (and wittier) than me once noted, the truth is never pure and rarely simple, and there are probably no easy answers to this question. During this period of declining returns on investment, there have however been some other trends in the industry that I believe to be worthy of consideration.

Interestingly, this period has seen an effective doubling of the amount of scientific data gathered each year. Thanks to incredible advances in laboratory instrumentation and computing, both the scope and the resolution of the data that can be captured in the laboratory, have increased to the point at which scientists working on data-intensive projects must routinely deal with terabytes of data. And while having more and better data is always valuable, a huge problem facing scientists right now, is that the technology for gathering and storing data has outstripped our ability to assimilate it and to turn it into real knowledge. In some sense, we "know" more and more about the biological systems we study, yet far from helping us to develop new and better therapies, the divergence between our investment in collecting all of this data and the productivity that results from it, continues to grows larger with each passing year.

All great truths begin as blasphemies
George Bernard Shaw

In tandem with this technology-driven surge in our ability to collect more and better data, there has been an accompanying movement in the life sciences towards the kind of phenomenological and data-driven approaches that these new technologies have enabled. It seems that every area of biology now has its own 'omics, in spite of the fact that the much-vaunted potential of the original 'omic field of genomics, still has yet to live up to its promise. Hot on the heels of The Human Genome Project, some luminaries in the field were making predictions such as a cure for cancer within a couple of years, and yet here we are a decade or more later, still waiting for the "magic bullet" therapies that our "knowledge" of the human genome was supposed to unleash.

Now don't get me wrong, the data gathered in the course of The Human Genome Project is invaluable, but it is still data, not knowledge. In retrospect, we can now see that what was perhaps not fully understood at the time, was how incomplete our real knowledge of the human genome was (and still is). Turning data into real knowledge requires an intellectual framework to organize and make sense of it. To use an extreme analogy, even if we could map the connections between every neuron in a human brain, this vast data set would not be sufficient by itself to explain anything much about human consciousness. Similarly, the human genome project has yielded vast volumes of sequence data, yet we still do not have the means to translate most of this information into the kind of useful insight that would allow us to make better drugs. In spite of the hyperbole surrounding the project, even at the time there were those who cautioned against unrealistic expectations.

Do what you've always done and
you'll get what you've always gotten

It is my conviction that one of the major reasons for the stagnation in real progress in the pharmaceutical industry over the last couple of decades, has been an overly heavy reliance on the kind of data-driven approaches characterized by the 'omics fields and discovery strategies such as high throughput screening, to the detriment of more knowledge-driven approaches. This is not to say that these data-driven approaches have not had their successes, nor that the choice to pursue these approaches is entirely unjustified. They are a response to the daunting challenge of tackling the vast complexity of biological systems and the product of considerable optimism about the capabilities of the new technologies that make these approaches possible. Rather than trying to really understand the system, these approaches aim to sidestep biological complexity by the outwardly more pragmatic approach of, well ... trying stuff to see what works.

The declining success rates for new drugs offers little comfort to those who feel that the industry is still on a good course. Only 15% of biologics obtain approval compared to a mere 7% of traditional small molecule drugs. These figures also fail to take into account the incredible costs associated with "successful" drugs that actually reach the market and start generating revenue, only to turn into money pits themselves as a result of safety issues that time and patient numbers start to make evident. The list of drugs like Vioxx, Avandia and Accutane that have become household names for all of the wrong reasons, continues to grow, with lawsuit damages and settlements compounding the already burgeoning costs of drug development.

Daunting though it might be to contemplate tackling the complexity of biological systems head-on, it is my belief that a greater emphasis on knowledge-driven approaches would be of enormous benefit to the pharmaceutical industry. A far more rigorous and thorough characterization of a new drug is possible when it is accompanied by a mechanistic understanding of the underlying biology, and of the impact of the drug on the activity of the biological networks from which the "biology" of the studied system is an emergent property.

Computational modeling and simulation approaches that are much more in the mainstream of research in other fields such as physics and engineering, offer the potential to provide the kind of intellectual frameworks mentioned earlier, with which biological data can be organized and explained. Until quite recently, biological modeling has been severely limited in scope and resolution by virtue of the difficulties of describing complex biological systems in the language of traditional mathematics. New approaches to biological modeling however, enable life scientists to create models of the kind of scope and resolution that make it possible to perform meaningful simulations of relatively macroscopic biological processes.

There are impediments to the adoption of these new approaches. The great majority of senior executives in the pharmaceutical industry have had little or no exposure to the kind of systems-based thinking that underpins these approaches and are often in departments and divisions that have invested heavily in data-driven approaches such as high throughput screening. Nor is systems biology currently part of the main work flow for drug development in most pharmaceutical companies. There have been numerous attempts (and re-attempts) to integrate systems biology approaches into the drug development process, but any time the economics of the industry has required a little belt-tightening, most companies have tended to view systems biology departments from a kind of "last in, first out" perspective when looking for expenditure items that can be circled with the red pen.

Change is difficult and the heavy levels of investment in the current approaches to drug discovery and development make it even harder for the industry to change course. The life science sector's greatest resource however, is its people, and their incredible level of talent and education are reason enough to be optimistic that the industry will in time, find both the will to change its current course, and the right people to spearhead its advance in a new direction.

 The author Gordon Webster, has spent his career working at the intersection of biology and computation and specializes in computational approaches to life science research and development.

  © The Digital Biologist | All Rights Reserved

Code Tutorial: Getting started with Python in the lab

My previous article was inspired by my impression that many life science researchers are missing out on some of the wonderful computational tools that are out there, and that could be invaluable in helping them to manage and analyze their laboratory data. This may be due to a lack of awareness that these tools exist and that many of them are inexpensive or free, or it may be out of fear that learning to use them will take too much time out of an already hectic schedule. In starting this series of code tutorials for life science computing, I hope to point life scientists with relatively little exposure to programming languages, in the right direction to be able to start using these tools to write useful code that can solve real problems.

In this tutorial we will create a simple program that can predict the pattern of single-strand DNA fragments that you would expect on your electrophoresis gel after doing a restriction digest. So without further ado, let's introduce one of our favorite programming languages Python. Python  is an incredibly versatile programming environment, being well suited for creating solutions from small scripts of a few lines in length, to large scale applications that are used throughout global organizations. The Python development environment is freely available for download from the official Python web site. For users of  Windows and Mac OS X platforms, Python comes in nice self-installing packages that pretty much do everything for you. For Linux users doing a manual install, you will need to unzip the install package and run a shell script or (more conveniently), you could just do a one-click install with one of the excellent software package managers that come with most Linux OS distributions. The good news for Linux users is that chances are, your Linux OS already comes with Python included and if you're running your own Linux system, you probably don't need our help installing stuff anyway!

We'll be writing our code using the nice clean IDLE development environment that comes with Python. IDLE is really a kind of specialized text editor that can actually run the Python code you type into it. On Mac OS X or Linux, it can be launched from a terminal window (usually by just typing "idle" if everything is installed and configured correctly). On Windows you will probably need to launch it from the command line ("run") window.

Is IDLE running now? Congratulations, you just ran your first Python application because IDLE itself is actually written in Python!

Just Do It

So let's start by creating a (very) simple DNA sequence in Python. In IDLE, type the following line and hit the return key

mysequence = "atcg"

In programming jargon, 'mysequence' is a string of 4 characters. IDLE will probably color code the "atcg" part to show this. Now type the following and hit return

len(mysequence)

If everything went well, IDLE should now display the number 4 which is the length of the string in characters returned by the 'len' function. Now try typing the following lines. Make sure you indent the 'print ch' line by one tab space just as I have done and you will have to hit the return key twice after the 'print ch' line to actually run the code.

for ch in mysequence:
    print ch

You should now see your sequence printed out in order, one character per line like this

a
t
c
g

The code you just typed translates into english as "for each item in 'mysequence', give the item the temporary name 'ch' then print 'ch'". This kind of loop construction is known as an iterator in Python. Python knows that strings are made up of smaller items, in this case characters, so when you iterate over a string using the 'for' command, the iterator returns each character in the string, in order of occurrence. I used the name 'ch' because the items in this case are characters, but I could just have easily used any other name like 'item' or 'c' or 'roger_the_shrubber'.

The indent in the 'print ch' line is significant. Indents are the way to create blocks of code in Python that are to be executed all together. Every line of code that is indented by one level underneath the 'for' statement, gets executed on each round trip through the loop, so if we had for example, also wanted to print the position of the nucleotide in the sequence, we could have added one or two more indented lines under the 'for' statement like this

i = 0
for ch in mysequence:
    print ch
    print i
    i += 1

In this modified version, a variable i is declared before the loop and then also printed and incremented by 1 on each pass through the loop. The 'i += 1' syntax is Python shorthand for 'i = i + 1'. The important thing to note here is that all 3 of the indented lines are executed for each pass through the loop.The output from the loop now looks like this

a
0
t
1
c
2
g
3

If we were to unindent  the 'i += 1' line, this would remove it from the loop and it would become the first line to be executed after all the passes through the loop had been completed. Can you figure out how this would change the output printed by the loop?

Let's bring in the string section

Another feature of strings (and all list-based objects in Python) is that their characters/items can also be accessed according to their numerical position in the object, starting from 0 (all lists, sequences, arrays and so on in Python are numbered from 0), up to 1 minus the length of the object. So mysequence[0] = 'a', mysequence[1] = 't' and so on. You can also slice strings up using [from:to] syntax like this

name = "Arthur King"
print len(name)
  11
print name[0:6]
 Arthur
print name[:6]
 Arthur
print name[7:11]
 King
print name[7:]
 King
print name[-4:]
 King

Note the following features of Python indices:  the last number in the 'from:to' range is exclusive; if the 'from' or 'to' index is excluded, the beginning or end of the string/list is assumed; a negative index for a string/list denotes the position -n from the end of the string.

Dictionaries are your friend

So now we know how to create DNA sequences as strings and how to iterate over them, let's introduce the Python dictionary. Dictionaries resemble in some aspects the kind of list object that we already encountered in strings, but instead of the items being labeled according to their numerical positions, they can be labeled pretty much any old way we choose. For example ...

mydictionary = {"galahad":"pure","lancelot":"brave"}

Now if you type

print mydictionary["lancelot"]

you get

'brave'

and so on. Dictionaries are composed of 'key:value' pairs and the stored value for each key can be accessed using the mydictionary[key] syntax. Dictionaries can also be added to like this

mydictionary["bedevere"] = "wise"

So now let's create a dictionary to store the molecular weights of the 4 standard DNA nucleotides

nucleotides = {'a':131.2,'t':304.2,'c':289.2,'g':329.2}

Note that we used single quotes here. Python allows both single and double quotes to be used interchangeably but the closing quote must be the same as the opening quote. What this dictionary does in effect, is to translate the symbol for the nucleotide into a molecular weight, for example

print nucleotides['t']

prints the value

304.2

So now we are in a position to actually do a useful calculation. By combing the 'for' iterator for our sequences and the nucleotide dictionary, we can calculate the molecular weight of any DNA sequence. Here's how ...

molecularweight = 0.0
for ch in mysequence:
    molecularweight += nucleotides[ch]
print molecularweight

When you run this code, you get the molecular weight for our little 4 nucleotide sequence

1053.8

Now we could also do all kinds of fancy stuff like adding the molecular weight of a 5' phosphate if one is present, or accounting for DNA adducts or chemically modified bases, but you get the general idea.

Organize your code with functions

Since we're going to want to use the same molecular weight calculation on different sequences, let's create a Python function that takes a DNA sequence as input and returns its molecular weight.

def calculateMolecularWeight(sequence):  
    molecularWeight = 0.0
    for ch in sequence:
        molecularWeight += nucleotides[ch]
    return molecularWeight

In Python the 'def' statement defines a function by its name 'calculateMolecularWeight' and by the parameters that it takes as input - in this case a single parameter 'sequence'. Note that Python is a dynamically typed language that does not require you to define what kind of data 'sequence' is before you run it, so the calculateMolecularWeight function would happily accept an integer like 2 as the input parameter, but it would raise an error when it tried to process it since an integer has no items in it for the 'for' loop to iterate over. Notice also how all of the code that forms the body of the function is indented (with the code in the 'for' loop being indented one more level still). The 'return' statement in a function does two things - it exits the function and if it is accompanied by a parameter as in this case, it sets the return value of the function to that parameter.

Now that we have our function, we can calculate the molecular weight for any DNA sequence that we give it as an input parameter, like this

mwt = calculateMolecularWeight(mysequence)

or we can even supply the sequence parameter explicitly like this

mwt = calculateMolecularWeight("gatgctgtggataa")

Lists: The missing ingredient

So now let's look at adding the final feature - the ability to detect restriction sites in the sequence and calculate the molecular weights of all the DNA fragments produced by a restriction digest.

We can define restriction sites as sequences in a similar way as we did for DNA sequences. For example

bamH1 = "ggatcc"
sma1 = "cccggg"

but in addition to the restriction enzyme's recognition motif, we also need to include information about where the enzyme cuts the DNA strand within the motif, so let's exapnd our deinition of a restriction site by using the very versatile list format that Python offers.

bamH1 = ["ggatcc",0]
sma1 = ["cccggg",2]

Lists can be formed by grouping items in square brackets as shown above. Python lists can contain other Python objects of any kind - even other Python lists! The second item (an integer) in our restriction site definition is the position in the motif after which the enzyme cuts the DNA (again numbering from 0 as the first position). The items in lists can be accessed by their positions, just like the characters in a string, so typing

print sma1[1]

returns the value

2

Remember that Python lists are numbered from 0, so 'sma1[1]' is actually the 2nd item in the list. Now that we have our restriction sites, we need a way to search for them in a DNA sequence. Fortunately there is a very handy 'find' method for Python strings that does exactly what we want to do. The 'find' method is a property of strings in Python and such a property can be accessed using the dot notation like this

position = mysequence.find("tag")

This looks for the next occurrence of the amber stop codon in 'mysequence' and returns its position in the sequence. If the search motif is not found, 'find' returns the value -1. Notice the phrase 'next occurrence' above. The 'find' method returns the position of the first occurrence of the search motif that it finds.  If you want to look for multiple occurrences, an additional parameter can be supplied to 'find' to tell it which position in the sequence to start searching from, like this

position = mysequence.find("tag",172)

Then for example, if you find the first stop codon at position 172, you can search again from that position to look for the next occurrence and so on. If no position parameter is supplied to 'find' (as in the first example), the search is always done from the beginning of the string.

The .find syntax works for any Python string  because in object-oriented programming (OOP) terms, 'mysequence' belongs to the general string class and 'find' is a method defined for that class. We will hardly delve at all into OOP in this introductory tutorial, except to say that Python is an OOP language and with a little OOP knowledge there are ways that we could write the code in this tutorial much more cleanly and elegantly in an OOP style. For instance, we could define our own DNA sequence and restriction site classes, with their own properties and methods. Fortunately for us however, unlike some other OOP languages, Python does not force the programmer to work in an OOP style all the time, so we can learn just the general nuts and bolts of Python for now and save OOP for another day.

We already have our molecular weight function, so let's add another function for doing restriction digests of our sequences. This function will take a sequence and a restriction site definition as input and return a list of the digested DNA fragments. Here it is ...

def digestDNA(sequence,rs):
    frags = []
    last = 0
    while 1:
        i = sequence.find(rs[0],last)
        if i == -1:
            frags.append(sequence[last:])
            break
        else:
            frags.append(sequence[last:i+rs[1]])
            last = i+rs[1]
    return frags

The line 'frags = []' creates an empty list that we will use to store our DNA fragments. The variable 'last' will be used to record the last place that we found the restriction site defined in 'rs'. Initially, 'last' must obviously be set to 0 so that the search starts at the beginning of the sequence.

The Python 'while' command is another kind of iterative loop that works differently from the 'for' loop that we have already encountered. The 'while' loop will continue to cycle for as long as the condition after 'while' is true. It is easier to understand how the 'while' loop works by studying this example

x=0
while x < 10:
    print x
    x += 1

This loop will print the numbers from 0 to 9 and then stop once x reaches 10. In our function however, we don't seem to have any condition at all after 'while', but just a rather mysterious integer, the number 1. The reason for this is that integers in Python can be used as shorthand for true or false where an integer greater or less than 0 is 'true', with 0 being 'false'.

In our function then, the loop would seem to run forever since 1 is always true and never gets changed. If you look in the indented body of the loop however, you will see the 'break' command which exits the loop. The loop in our function is essentially configured to run forever, until we encounter the situataion where there are no more occurrences of the restriction site motif (in which case sequence.find() returns -1) and then we simply add the remainder of the sequence to our list of fragments and exit the loop using the 'break' command.

The 'if' construct tests whether the value returned from 'find' equals  -1. Notice that in the conditional statement we use 'i == 1' which means "return 'true' if i equals 1 or 'false' if not", as opposed to "i=1" which means "set the value of i to 1". If the condition following 'if' is true, everything in the indented 'if' block is done. The 'else' block is a way to tell Python what to do in the event that the 'if' condition is false. In our case, if 'find' does find another occurrence of the restriction site, we use the restriction site cut position to define the new fragment we're generating, we add the new fragment to our fragment list using the 'append' method that is a property of  Python lists and we update the 'last' variable so that next search for the restriction site will continue from the position of its last occurrence. Pretty simple right?

All together now ...

Now we can finally put it all together to create a Python script that will calculate the molelcular weights of the DNA fragments produced by a restriction digest. You will notice that the 'digestDNA' function returns a list of DNA sequences, so after we have run the function, we can use our iterative 'for' loop to go through the returned list of fragments one at a time and calculate the molecular weight of each one, like this

for fragment in digestDNA(s,sma1):
    print item
    print calculateMolecularWeight(item)

Notice that instead of assigning the return value of the 'digestDNA' function to a variable and then iterating over the variable like this

fragmentList = digestDNA(s,sma1)
for fragment in fragmentList:
    ...
    ...

we just used the function itself directly where the variable would go. This is perfectly OK to do in Python and in fact, it is often good form to do this to make the code more concise and readable.

So let's create a fake sequence with a couple of SmaI restriction sites in it and test our code.

s = "atcgatcgatcgcccgggatcgatcgcccgggatcgatcgatcg"
for fragment in digestDNA(s,sma1):
    print fragment
    print calculateMolecularWeight(fragment)

If everything went well, running this code should yield the following output

atcgatcgatcgcc
3739.8
cgggatcgatcgcc
3962.8
cgggatcgatcgatcg
4438.2

If we wanted to do a multiple digest, for each additional restriction site, we could apply the same approach to the fragments generated by the previous digest(s). In this case, we could write our code in such a way that it could be applied recursively to our original sequence and the subsequent fragments. Recursive methods are easy to write in Python, but we'll save recursion for a future tutorial.

 One of the great things about Python is that it is a very popular language with a large and flourishing community. This means that for most problem domains, including the life sciences, somebody somewhere has probably already developed a library of Python functions for your particular problem. As you might have guessed, there already exist Python libraries for the kind of bioinformatics problems that we have tackled here. Since the goal here was to learn enough basic Python to be able to write some useful code, we did not rely on any of them for our simple problem. Here are some links however, to give you an idea of the kinds of life science libraries that are available in Python.

Biopython - a set of freely available tools for biological computation
Biskit - an object-oriented library for structural bioinformatics research
SciPy - a library for mathematics, science, and engineering, with many useful algorithms for life science computing

This list is by no means exhaustive, but it gives some idea of the universe of useful Python tools that are available to the life scientist.

Don't stop there!

So now that you have hopefully seen enough to appreciate how easy it can be to write useful code, we encourage you not to stop here. We will be publishing future tutorials, but in the meantime, why not head on over and browse the many fantastic references, tutorials and other resources for new Python users that can be found at the documentation area of the Python web site.

  © The Digital Biologist | All Rights Reserved