Post

We have a paper in Nature today! www.nature.com/articles/s41... Several years in the making, Delphy (www.delphy.bio) is one of the tools for outbreak response that we've been building over the last several years with Pardis Sabeti over at the Broad Institute.

The project is the brainchild of the inestimable Patrick Varilly, and the tool is 100-1000x faster(!!) than others. For our part, we worked with Patrick on how to make Delphy usable & accessible beyond the handful of PhDs who know the necessary command line incantations to run this kind of analysis.

While it's always nice to run faster, what's *actually* exciting about that is the way that it can change the way you work. It means trying things early instead of waiting for perfection. You can immediately identify overall trends, bad data, outliers, etc and fix these problems before the next run.

Instead of spending all your time cleaning data first, there's essentially no cost to taking a quick look and proceeding with your research. That's really powerful, because *all data is bad* and the pursuit of what passes for “good” is both time consuming and has its own kinds of badness.

(That last part is a post for another time. This is already too long.)

Another thing we enjoyed about this project is the ability to visualize the uncertainty, ranges, and error behind its calculations. The output of this analysis is a phylogenetic tree, but like all such trees, it's inherently a lie. It's the least worst lie we can come up with, but it's still a lie.

Like the map of a hurricane path, with its error bars that everyone agrees is a range yet mentally perceives it as an actual path, a tree can be a distraction. So we've been trying to lean into that, and showing the haziness everywhere we can.

Part of understanding a dataset is the idea of being able to interact with the data to learn about it. This is something I take from my days during undergrad working in child development. The basic idea in constructivism is that you can't just shovel information at kids and have them memorize it.

How can they poke at it, look at it from different sides, play with it? It's not about single answers, it's about building a mental model of what the data *means*, whether for a research paper about how an outbreak has spread, or for a policymaker to make decisions about outbreak response.