Skip to content

Commit

Permalink
changed README language etc
Browse files Browse the repository at this point in the history
  • Loading branch information
@jr-leary7
jr-leary7 committed Nov 7, 2023
1 parent ca28d29 commit ca542b8
Show file tree
Hide file tree
Showing 2 changed files with 20 additions and 18 deletions.
4 changes: 2 additions & 2 deletions .github/workflows/render-README.yaml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,13 +22,13 @@ jobs:
- uses: r-lib/actions/setup-pandoc@v2

- name: install CRAN packages
run: Rscript -e 'install.packages(c("rmarkdown","ggplot2", "dplyr", "purrr", "remotes", "BiocManager"))'
run: Rscript -e 'install.packages(c("rmarkdown","ggplot2", "dplyr", "purrr", "remotes", "devtools", "BiocManager"))'
- name: install BioConductor packages
run: Rscript -e 'BiocManager::install(c("SingleCellExperiment", "scater", "scran"))'
- name: install GitHub packages
run: Rscript -e 'remotes::install_github("jr-leary7/scLANE")'
- name: render README
run: Rscript -e 'rmarkdown::render("README.Rmd", output_format = "md_document")'
run: Rscript -e 'devtools::build_readme()'
- name: commit rendered README
run: |
git config --local user.name "jr-leary7"
Expand Down
34 changes: 18 additions & 16 deletions README.Rmd
View file
Original file line number Diff line number Diff line change
Expand Up @@ -54,7 +54,7 @@ Our method relies on a relatively simple test in order to define whether a given

First we'll also need to load a couple dependencies & resolve a function conflict.

```{r libraries, results='hide'}
```{r libraries, results='hide', message=FALSE}
library(dplyr)
library(scater)
library(scLANE)
Expand Down Expand Up @@ -87,7 +87,7 @@ plotPCA(sim_data, colour_by = "subject") +

## Trajectory DE testing

Since we have multi-subject data, we can use any of the three model backends to run our DE testing. We'll start with the simplest model, the GLM, then work our way through the other options in order of increasing complexity. We first prepare our inputs - a dataframe containing our cell ordering, a set of genes to build models for, and a vector of per-cell size factors to be used as offsets during estimation. In reality, it's usually unnecessary to fit a model for every single gene in a dataset, as trajectories are usually estimated using a subset of the entire set of genes (usually a few thousand most highly variable genes). For the purpose of demonstration, we'll select 50 genes each from the dynamic and non-dynamic populations. **Note**: in this case we're working with a single pseudotime lineage, though in real datasets several lineages often exist; in order to fit models for a subset of lineages simply remove the corresponding columns from the cell ordering dataframe passed as input to `testDynamic()`.
Since we have multi-subject data, we can use any of the three model modes to run our DE testing. We'll start with the simplest model, the GLM, then work our way through the other options in order of increasing complexity. We first prepare our inputs - a dataframe containing our cell ordering, a set of genes to build models for, and a vector of per-cell size factors to be used as offsets during estimation. In reality, it's usually unnecessary to fit a model for every single gene in a dataset, as trajectories are usually estimated using a subset of the entire set of genes (usually a few thousand most highly variable genes). For the purpose of demonstration, we'll select 50 genes each from the dynamic and non-dynamic populations. **Note**: in this case we're working with a single pseudotime lineage, though in real datasets several lineages often exist; in order to fit models for a subset of lineages simply remove the corresponding columns from the cell ordering dataframe passed as input to `testDynamic()`.

```{r prep-data}
set.seed(312)
Expand All @@ -97,7 +97,7 @@ order_df <- data.frame(X = sim_data$cell_time_normed)
cell_offset <- createCellOffset(sim_data)
```

### GLM framework
### GLM mode

Running `testDynamic()` provides us with a nested list containing model output & DE test results for each gene over each pseudotime / latent time lineage. In this case, since we have a true cell ordering we only have one lineage. Parallel processing is turned on by default, and we use 4 cores here to speed up runtime.

Expand All @@ -109,7 +109,7 @@ scLANE_models_glm <- testDynamic(sim_data,
n.cores = 4)
```

After the function finishes running, we use `getResultsDE()` to generate a sorted table of DE test results, with one row for each gene & lineage. The GLM backend uses a simple likelihood ratio test to compare the null & alternate models, with the test statistic assumed to be [asymptotically Chi-squared distributed](https://en.wikipedia.org/wiki/Likelihood-ratio_test).
After the function finishes running, we use `getResultsDE()` to generate a sorted table of DE test results, with one row for each gene & lineage. The GLM mode uses a simple likelihood ratio test to compare the null & alternate models, with the test statistic assumed to be [asymptotically Chi-squared distributed](https://en.wikipedia.org/wiki/Likelihood-ratio_test).

```{r glm-results}
scLANE_res_glm <- getResultsDE(scLANE_models_glm)
Expand All @@ -120,9 +120,9 @@ select(scLANE_res_glm, Gene, Lineage, Test_Stat, P_Val, P_Val_Adj, Gene_Dynamic_
col.names = c("Gene", "Lineage", "LRT stat.", "P-value", "Adj. p-value", "Predicted dynamic status"))
```

### GEE framework
### GEE mode

The function call is essentially the same when using the GLM backend, with the exception of needing to provide a sorted vector of subject IDs & a desired correlation structure, the default being [the AR1 structure](https://en.wikipedia.org/wiki/Autoregressive_model). We also need to flip the `is.gee` flag in order to indicate that we'd like to fit estimating equations models (instead of mixed models). Since fitting GEEs is more computationally complex than fitting GLMs, DE testing with the GEE backend takes a bit longer. Using more cores and / or running the tests on an HPC cluster speeds things up considerably.
The function call is essentially the same when using the GLM mode, with the exception of needing to provide a sorted vector of subject IDs & a desired correlation structure, the default being [the AR1 structure](https://en.wikipedia.org/wiki/Autoregressive_model). We also need to flip the `is.gee` flag in order to indicate that we'd like to fit estimating equations models (instead of mixed models). Since fitting GEEs is more computationally complex than fitting GLMs, DE testing with the GEE mode takes a bit longer. Using more cores and / or running the tests on an HPC cluster speeds things up considerably.

```{r gee-models}
scLANE_models_gee <- testDynamic(sim_data,
Expand All @@ -146,9 +146,9 @@ select(scLANE_res_gee, Gene, Lineage, Test_Stat, P_Val, P_Val_Adj, Gene_Dynamic_
col.names = c("Gene", "Lineage", "Wald stat.", "P-value", "Adj. p-value", "Predicted dynamic status"))
```

### GLMM framework
### GLMM mode

We re-run the DE tests a final time using the GLMM backend. This is the most complex model architecture we support, and is the trickiest to interpret. We recommend using it when you're most interested in how a trajectory differs between subjects e.g., if the subjects belong to groups like Treatment & Control, and you expect the Treatment group to experience a different progression through the biological process. Executing the function with the GLMM backend differs only in that we switch the `is.glmm` flag to `TRUE` and no longer need to specify a working correlation structure.
We re-run the DE tests a final time using the GLMM mode. This is the most complex model architecture we support, and is the trickiest to interpret. We recommend using it when you're most interested in how a trajectory differs between subjects e.g., if the subjects belong to groups like Treatment & Control, and you expect the Treatment group to experience a different progression through the biological process. Executing the function with the GLMM mode differs only in that we switch the `is.glmm` flag to `TRUE` and no longer need to specify a working correlation structure.

```{r glmm-models}
scLANE_models_glmm <- testDynamic(sim_data,
Expand All @@ -162,13 +162,13 @@ scLANE_models_glmm <- testDynamic(sim_data,
n.cores = 4)
```

{% note %}
:::{.callout-note}

**Note**: the GLMM backend is still under development, as we are working on further reducing runtime and increasing the odds of the underlying optimization process converging successfully. As such, updates will be frequent and functionality / results may shift slightly.
The GLMM mode is still under development, as we are working on further reducing runtime and increasing the odds of the underlying optimization process converging successfully. As such, updates will be frequent and functionality / results may shift slightly.

{% endnote %}
:::

Like the GLM backend, the GLMM backend uses a likelihood ratio test to compare the null & alternate models. We fit the two nested models using maximum likelihood estimation instead of [REML](https://en.wikipedia.org/wiki/Restricted_maximum_likelihood) in order to perform this test; the null model in this case is a negative binomial GLMM with a random intercept for each subject.
Like the GLM mode, the GLMM mode uses a likelihood ratio test to compare the null & alternate models. We fit the two nested models using maximum likelihood estimation instead of [REML](https://en.wikipedia.org/wiki/Restricted_maximum_likelihood) in order to perform this test; the null model in this case is a negative binomial GLMM with a random intercept for each subject.

```{r glmm-results}
scLANE_res_glmm <- getResultsDE(scLANE_models_glmm)
Expand All @@ -183,7 +183,7 @@ select(scLANE_res_glmm, Gene, Lineage, Test_Stat, P_Val, P_Val_Adj, Gene_Dynamic

### Model comparison

We can use the `plotModels()` to visually compare different types of modeling backends. It takes as input the results from `testDynamic()`, as well as a few specifications for which models & lineages should be plotted. While more complex visualizations can be created from our model output, this function gives us a good first glance at which models fit the underlying trend the best. Here we show the output generated using the GLM backend, split by model type. The intercept-only model shows the null hypothesis against which the scLANE model is compared using the likelihood ratio test and the GLM displays the inadequacy of monotonic modeling architectures for nonlinear dynamics. A GAM shows essentially the same trend as the `scLANE` model, though the fitted trend from `scLANE` is more interpretable & has a narrower confidence interval.
We can use the `plotModels()` to visually compare different types of models. It takes as input the results from `testDynamic()`, as well as a few specifications for which models & lineages should be plotted. While more complex visualizations can be created from our model output, this function gives us a good first glance at which models fit the underlying trend the best. Here we show the output generated using the GLM mode, split by model type. The intercept-only model shows the null hypothesis against which the scLANE model is compared using the likelihood ratio test and the GLM displays the inadequacy of monotonic modeling architectures for nonlinear dynamics. A GAM shows essentially the same trend as the `scLANE` model, though the fitted trend from `scLANE` is more interpretable & has a narrower confidence interval.

```{r plot-models-glm}
plotModels(scLANE_models_glm,
Expand All @@ -197,6 +197,8 @@ plotModels(scLANE_models_glm,
plot.scLANE = TRUE)
```

Model comparison using the GEE mode is similar, with the only change being that we now provide a vector of subject IDs.

```{r plot-models-gee}
plotModels(scLANE_models_gee,
gene = "DGUOK",
Expand All @@ -211,7 +213,7 @@ plotModels(scLANE_models_gee,
plot.scLANE = TRUE)
```

When plotting the models generated using the GLMM backend, we split by lineage & color the points by subject ID instead of by lineage. The gene in question highlights the utility of the scLANE model, since the gene dynamics differ significantly by subject.
When plotting the models generated using the GLMM mode, we split by lineage & color the points by subject ID instead of by lineage. The gene in question highlights the utility of the scLANE model, since the gene dynamics differ significantly by subject.

```{r plot-models-glmm, fig.width=9, fig.height=9}
plotModels(scLANE_models_glmm,
Expand Down Expand Up @@ -274,9 +276,9 @@ The smoothed dynamics can then be used to generate expression cascade heatmaps,

# Conclusions & best practices

In general, starting with the GLM backend is probably your best bet unless you have a strong prior belief that expression trends will differ significantly between subjects. If that is the case, you should use the GEE backend if you're interested in population-level estimates, but are worried about wrongly predicting differential expression when differences in expression are actually caused by inter-subject variation. If you're interested in generating subject-specific estimates then the GLMM backend should be used; take care when interpreting the fixed vs. random effects though, and consult a biostatistician if necessary.
In general, starting with the GLM mode is probably your best bet unless you have a strong prior belief that expression trends will differ significantly between subjects. If that is the case, you should use the GEE mode if you're interested in population-level estimates, but are worried about wrongly predicting differential expression when differences in expression are actually caused by inter-subject variation. If you're interested in generating subject-specific estimates then the GLMM mode should be used; take care when interpreting the fixed vs. random effects though, and consult a biostatistician if necessary.

If you have a large dataset (10,000+ cells), you should start with the GLM backend, since standard error estimates don't differ much between modeling methods given high enough *n*. In addition, running the tests on an HPC cluster with 4+ CPUs and 64+ GB of RAM will help your computations to complete swiftly. Datasets with smaller numbers of cells or fewer genes of interest may be easily analyzed in an R session on a local machine.
If you have a large dataset (10,000+ cells), you should start with the GLM mode, since standard error estimates don't differ much between modeling methods given high enough *n*. In addition, running the tests on an HPC cluster with 4+ CPUs and 64+ GB of RAM will help your computations to complete swiftly. Datasets with smaller numbers of cells or fewer genes of interest may be easily analyzed in an R session on a local machine.

# Contact information

Expand Down

0 comments on commit ca542b8

Please sign in to comment.