Reproducibility on a technical level vs a conceptual level

I am and have been a huge proponent of reproducibility in scientific computation for most of my training. I think that reproducibility is important both on a technical level and on a conceptual level. Reproducibility at a technical level—what many call replicability—is the ability successfully generate a similar/identical output given the source code and data. Reproducibility at the conceptual level—what we often refer to as reproducibility in a more general sense—is being able to get a consistent result using different methods (e.g. re-implement the logic in your own software), different data (e.g. do your own data collection on the subject of interest), or both. The latter—reproducibility at the conceptual level—is the ultimate test, and gives much stronger support of a theory/model/result than replicability.

In some ways, reproducing a computational result in biology is much easier than reproducing an experimental result. Computers are just machines* that execute sets of instructions (*gross oversimplification noted), and assuming I provide the same instructions (source code) and input (data) to another scientist, it should be trivial for them to exactly and precisely reproduce a result I previously computed.

However, I speak of an ideal world. In the real world, there are myriad technical issues: different operating systems; different versions of programming languages and software libraries; different versions of scientific software packages; poor workflow documentation; limited computing literacy; and a variety of other issues that make it difficult to reproduce a computation. In this way, computational biology is a lot more like experimental biology: getting the “right” result depends highly on your environment or setup, your experience, and in some cases having some secret sauce (running commands/programs in the right order, using certain settings or parameter values, etc). In this way, computational biology is a lot more like experimental biology than we let on sometimes.

It doesn’t have to be this way. But it is.

There is a great discussion going on in the social media right now about what should be expected of academic scientists producing software in support of their research. I hope to contribute a dedicated blog post to this discussion soon, but it is still very nascent and I would like to let my thoughts marinate for a bit before I dive in. Several points are clear, however, that are directly relevant to the discussion of replicability and reproducibility.

  • Replicability in computation supports reproducibility. I’m not sure of anyone that disagrees with this. My impression is that most disagreements are focused on what can reasonably be expected of scientists given the current incentive structure.
  • Being unable to replicate a computational study isn’t the end of the world for academic science: important models and theories shouldn’t rely on a single result anyway. But the lack of computational replicability does make the social enterprise of academic science, already an expensive and inefficient ordeal under constant critical scrutiny, even more expensive and inefficient. Facilitating replicability in computation would substantially lower the activation energy required to achieve the kind of real reproducibility that matters in the long run.
  • There are many academics in the life sciences at all levels (undergrad to grad student to postdoc to PI) that are dealing with more data and computation right now than their training has ever prepared them to deal with.
  • A little bit of training can go a long way to facilitating more computational replicability in academic science. Check out what Software Carpentry is doing. Training like this may not benefit cases in which scientists deliberately obfuscate their methods to cover their tracks or to maintain competitive advantage, but it will definitely benefit cases in which a result is difficult to replicate simply because a well-intentioned scientist had little exposure to computing best practices. (Full disclosure: I am a certified instructor for Software Carpentry, although I receive no compensation for my affiliation or participation. I’m just a huge proponent of what they do!)

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s