NetCoupler: Inferring causal pathways between high-dimensional metabolic data and external factors

# NetCoupler: Inferring causal pathways between high-dimensional metabolic data and external factors
### <p>Luke W. Johnston</p>
<p>Steno Diabetes Center Aarhus, Denmark</p>

---

<div class="my-footer">
<span>
<img src="../../common/dda-logo.png" alt="DDA", width="75">
<img src="../../common/sdca-logo.png" alt="SDCA", width="55">
<a href="https://slides.lwjohnst.com/iarc/2020-12-16/">slides.lwjohnst.com/iarc/2020-12-16</a>
</span>
</div>

---

# Outline:

- Background on analytical problem

- NetCoupler history and implementation

- R package work and usage

- Examples using NetCoupler

- Current challenges

???

Overall though, I'm very much looking forward to feedback, thoughts, or comments
on NetCoupler to help make it better.

---

# Background: Analytic problem

---

## Modern studies generate a mass of metabolic data

- Metabolic biomarkers
]

- Complex networks
]

---

## "Traditional" analysis may use: Dimensionality reduction

<div class="figure" style="text-align: center">
<img src="data:image/png;base64,#../../au-ph/2019-08-15/images/pca.png" alt="Reducing number of variables with PCA." width="280" />
<p class="caption">Reducing number of variables with PCA.</p>
</div>

???

This has the advantage of making things simpler while trying to maximize
variance in the data. Afterward you can do modelling on each principal component.
The disadvantage of this approach is that it loses a lot of information since
the interdependence and connections between variables it not maintained.

---

## "Traditional" analysis may use: Many regression-type models

.center[
```
O1 = M1 + covariates
O1 = M2 + covariates
...
O1 = M7 + covariates
O1 = M8 + covariates
```
]

???

Some ways you might go about analyzing this data is by running many regression
models, one for each metabolic variable for instance.

This of course has problems since you're simply running a bunch of models and
not taking account of the inherent interdependencies between variables.

---

## "Traditional" analysis may use: Network analysis

???

This approach is nice in that you can extract information about the connection
between metabolic variables. But there is no way to incorporate the disease
outcome with this approach and in order to construct the network properly most
methods require you provide a prespecified base network, which you might not know.

---

## But, what if we...

- want info about network structure?
--

- don't know the network structure?
--

- have an exposure, metabolites, and outcome?
--

- are interested in causal links?

---

## ... if what we want to know is something like this?

---

# (Potential) solution: NetCoupler

## History and implementation

---

## Initial development

.pull-left[
- Developed by Clemens Wittenbecher for his 
[PhD thesis](https://publishup.uni-potsdam.de/opus4-ubp/frontdoor/deliver/index/docId/40459/file/wittenbecher_diss.pdf)

- Algorithm that:
    - Finds most likely network structure
    - Allows inclusion of exposure and outcome
    - Identifies causal links between and within network
]

.pull-right[
![Clemens Wittenbecher](data:image/png;base64,#https://avatars3.githubusercontent.com/u/33724052?size=200)
]

---

## Four basic phases of the algorithm

---

## Infer (potentially) causal pathways with graphical model output

---

# Developing R package and usage

???

Met him at EDEG, asked if it was an R package. Started working together after
that.

---

## Current state of NetCoupler as <svg style="height:0.8em;top:.04em;position:relative;fill:#214c78;" viewBox="0 0 581 512"><path d="M581 226.6C581 119.1 450.9 32 290.5 32S0 119.1 0 226.6C0 322.4 103.3 402 239.4 418.1V480h99.1v-61.5c24.3-2.7 47.6-7.4 69.4-13.9L448 480h112l-67.4-113.7c54.5-35.4 88.4-84.9 88.4-139.7zm-466.8 14.5c0-73.5 98.9-133 220.8-133s211.9 40.7 211.9 133c0 50.1-26.5 85-70.3 106.4-2.4-1.6-4.7-2.9-6.4-3.7-10.2-5.2-27.8-10.5-27.8-10.5s86.6-6.4 86.6-92.7-90.6-87.9-90.6-87.9h-199V361c-74.1-21.5-125.2-67.1-125.2-119.9zm225.1 38.3v-55.6c57.8 0 87.8-6.8 87.8 27.3 0 36.5-38.2 28.3-87.8 28.3zm-.9 72.5H365c10.8 0 18.9 11.7 24 19.2-16.1 1.9-33 2.8-50.6 2.9v-22.1z"/></svg> 📦

.center[
<img src="data:image/png;base64,#../../iarc/2020-12-16/images/netcoupler-github.png" width="70%" style="display: block; margin: auto;" />
]

.footnote[
- [github.com/NetCoupler](https://github.com/NetCoupler/NetCoupler)
- **Goal**: Submit to CRAN by mid-2021.
]

---

## General framework and features

- Pipe ([magrittr](https://magrittr.tidyverse.org/) `%>%` or base R
`|>`) friendly

- Uses [tidyselect](https://tidyselect.r-lib.org/) helpers (e.g.
`starts_with()`, `contains()`)

- Auto-complete friendly (e.g. start function names with `nc_`)

- Inputs/outputs generally
[tibbles](https://tibble.tidyverse.org/)/dataframes or 
[tidygraph tibbles](https://tidygraph.data-imaginist.com/)

- Flexible with type of model

- Website with beginner-focused documentation

---

## Visual demo of input and output

---

## Example code

```r
std_data <- dataset %>% 
    nc_standardize(starts_with("metabolite"))
network <- std_data %>% 
    nc_estimate_network(starts_with("metabolite")) %>% 
    as_edge_tbl()
```

---

## Example code

```r
outcome_estimates <- std_data %>%
    nc_estimate_outcome_links(
        edge_tbl = network,
        outcome = "HbA1c",
        model_function = lm
    )
```

---

## Example code

```r
outcome_estimates <- std_data %>%
    nc_estimate_outcome_links(
        edge_tbl = network,
        outcome = "HbA1c",
        model_function = lm,
*       adjustment_vars = c("age", "sex")
    )
```

---

## Example code

```r
outcome_estimates <- std_data %>%
    nc_estimate_outcome_links(
        edge_tbl = network,
*       outcome = "incident_diabetes",
*       model_function = glm,
        adjustment_vars = c("age", "sex"),
*       model_arg_list = list(
*           family = binomial("logit")
*       ),
*       exponentiate = TRUE
    )
```

---

## Example plot of output

---

# Example projects using NetCoupler

---

## Clemens' PhD research

- **Aim**: Identify potential metabolic links between exposure to dietary risk
factors and later type 2 diabetes incidence based on metabolomics networks.

---

## Red meat, acylcarnitines, and incident diabetes in EPIC-Potsdam

---

## UK Biobank: Metabolic pathways between components of stature and HbA1c

---

## UK Biobank characteristics

.pull-left[
- Basics:
    - ~480,000 participants
    - Cross-sectional
    - Stature measures
    - Various demographics
]

.pull-right[
- Metabolic variables:
    - Cholesterol
    - Albumin
    - Alanine Aminotransferase
    - Apolipoprotein A and B
    - Aspartate Aminotransferase
    - C-reactive Protein
    - Gamma Glutamyltransferase
    - HDL and LDL Cholesterol
    - Triglycerides
]

---

## Link between leg length, liver markers, HbA1c

---

# Current challenges

---

## Several limitations or things to improve

- **Mainly**, performance can be slow
    - E.g. larger data or networks

- Untested on larger networks

- Untested on non-cross-sectional/time-to-event data

- Visualizing can be tricky

- Interpreting estimates can be tricky

???

Before addressing many of these I want the API to be stable first and the general
interface to be well-established before moving on to these things.

---

# Thanks! Comments, thoughts, feedback?