Tweet

Adam Ward

Apr 1 • 30 tweets • 26 min read

🚨⬇️ New pub! ⬇️🚨

TLDR: data science to generate interdisciplinary hypotheses

This is the paper I'm MOST PROUD OF in my career to date. A (long) thread to tell you why. 1/n

#hyporheic #river #stream #water #hydrology

doi.org/10.1002/hyp.14…

@geosociety

Scene: @geosociety meeting in Charlotte, NC, 2012.Brainstorming with @stefankrauseh2o about hyporheic science.

Problem: If we only measure what is in our narrow areas of expertise, we can plausibly explain our data as a function of unmeasured.

Solution: Lets do this!

@LeverhulmeTrust

Next step: Acquire funding.
Assembled a team, wrote a proposal to support many-disciplinary studies of hyporheic processes.

With support from @LeverhulmeTrust and @bIGIdeas_UoB we embarked on a series of studies to learn how to work together in truly integrated science

From 2012-2016 we undertook a series of experiments at various sites and scales. This helped us learn to work together, norm in the way we were thinking about hyporheic zones, and build a collaborative environment

@Hydro_DoctorJ

2016 is when it all came together, in a field campaign that I led along with @Hydro_DoctorJ at @HJA_Live. A team of about 20 scientists descended on a 5th order basin, armed with our sampling gear, ideas, and enthusiasm.

It took YEARS to process all of the data. Solute tracers, eDNA, water isotopes, metabolomics, spatial analysis, macroinvertebrates, nutrients, anions & cations

Data documented here:
essd.copernicus.org/articles/11/15…

The first analysis of these data was a look at stream solute tracers across the basis, testing some hypotheses published by my science BFF Steve Wondzell
hess.copernicus.org/articles/23/51…

@CiaranJHarman

That is also the pub where - after at least 8 years of extreme patience by @CiaranJHarman - I could understand StorAge Selection functions

@bIGIdeas_UoB

As the data came together (summer 2018), I was a fellow at @bIGIdeas_UoB & @LES_UniBham. On that visit, Aaron Packman - long-time mentor - pushed me to articulate a vision for the field, & for where we were going with this data set
doi.org/10.1002/wat2.1…

@biogeogenia

Team meetings continued at conferences and occasionally via zoom, and I plodded along and wasn't making much progress. Enter @biogeogenia.

@biogeogenia

@biogeogenia organized a "Slow Science Workshop" for our team in Spring 2020. We rented a large house where ~15 of us lived and worked, synthesizing data from the series of experiments we embarked on starting in 2012.

These ideas had been simmering for years, but this retreat gave us the time and space to formalize our vision for river corridors:
doi.org/10.1029/2021WR…

@KirchnerLabETH

At that retreat, I remembered hearing @KirchnerLabETH talking about work on the Plynlimon data set - essentially plotting 'everything against everything' to try and identify relationships worth exploring.

How could we do that, esp. expecting multivariate interactions?

@FengMao_

Now its @FengMao_ to the rescue. He took the data set and started building random forests, classifying, just playing around with things. That inspired me to start playing with SVMRs, which Tyler Balson (student on my team) had just been teaching me.

💡lightbulb moment. Now we can develop a web of relationships between variables spanning scales, disciplines, etc. We can handle loads of information and sift through it. A machine learning enabled approach to sift through the data and identify relationships.

Why did we need that approach? Each of our 62 sites was distilled to 157 variables describing physical, chemical, & biological features. There are nearly 25,000 pairwise relationships that 'could' exist. Too many to look at. Our approach distilled this into 672 worth considering

And here it is. 672 'ribbons' linking different variables to one another with directional relationships ( i.e., A=f(B) does not require B=f(A) ). The "NOT" and "S" are because we also considered if variables were spatially structured or not

Remember these are mathematically meaningful relationships, but are they mechanistic? Did we find correlation or causation?

@HJA_Live

We did recover expected relationships consistent with several past studies including (to name a few):

River continuum concept (where @HJA_Live was a key field site)

Our prior HESS pub on solute tracers in the basin

& Work by @nwisnoski
doi.org/10.1007/s00442…

So how much was rediscovery vs. novel relationships? 84% of the relationships we identified HAVE NOT been previously considered in the literature.

Remember - The goal here was to get outside of our disciplinary lenses. The machine learning is a 'hypothesis generator', not a substitute for detailed, mechanistic studies.

We didn't attempt to spin mechanistic hypotheses for every relationship we found, but we did attempt to explain a few novel findings:

Why are metabolomics data most informed by geological variation?

What controls N-acquiring enzymatic activity in a nitrogen-limited ecosystem?

Conclusions:
Building connections between existing studies requires explicitly planning for synthesis in future efforts. Here, we demonstrated the value of collecting data sets that enabled synthesis within and between locations, disciplines, and scales.

This does not diminish the value of traditional, disciplinary hypothesis testing and deductive approaches to science. Instead, common metadata and even a small number of out-of-group observations may enable synthesis efforts based on inductive approaches

Ultimately, inductive approaches are a useful way to generate hypotheses from existing observational datasets and advance our scientific understanding.

@jddrummo

The team. This has been 10 years with people coming and going, contributing, etc. I'm going to attempt to tag and acknowledge as many of our team and funders as I can. Here we go: Aaron Packman, Susana Bernal, Nicolai Brekenfeld, @jddrummo, @emilybonnell_g, @freshwaterflows...