Validating assumptions

Oftentimes, our interpretations of our experimental data are dependent on the assumptions that we bring to the table. I have previously discussed the problem of holding unexamined assumptions (see here and here), and how they can pervade our thinking. Here I want to ask the question of how we would validate our assumptions? In particular, if we have a model that is built on assumptions, what kind of data count as evidence in favor of the model and its assumptions?

In the early Drosophila embryo, Dorsal (Dl) is present at high levels on the ventral side (black nuclei on right). This gradient is established by the action of Toll signaling on the ventral and lateral parts of the embryo (not shown) causing Cactus (Cact), the inhibitor for Dorsal, to split up from Dorsal and be degraded (left). Previously, it had not been proposed that Cactus or Dorsal/Cactus complex could enter the nuclei (left). For more on Dorsal, Cactus, and Toll, see here.

This is a pertinent question for me because, in my lab, we do both modeling and experiment, and sometimes our modeling work is not trusted in the same way that experimental results would be. We had a paper a couple of years ago (O'Connell and Reeves, 2015) that made a few new assumptions about the Dorsal/Cactus system (see here for a short intro to that system), which resulted in a much better explanation of past data, but it also resulted in a slightly novel way of looking at the system (see Figure). Specifically, we suggested that not just free Dorsal, but also Dorsal/Cactus complex and free Cactus could enter the nuclei.

This idea had not been proposed before (at least, not at this stage in Drosophila development), and reviewers were suspicious. We even had a reviewer of a subsequent manuscript try to reject that manuscript because of the assumptions we made in the 2015 paper!

Now, given that we have a wet lab, one way to test our assumptions would be to simply perform well designed experiments. But sometimes, that's easier said than done. In this particular case, because Cactus is difficult to image (both in fixed embryos, but also in live imaging), all of the proposed experiments would be indirect validations of our results, at best.

So the question is: do we actually need to validate these assumptions? That's not as radical of a question as it might sound at first. I am basically asking: weren't the assumptions already validated by the fact that the model provided a better explanation of previous data? This the question that we'll begin to look at next time.