Smaller transcriptomic models do not exhibit superior generalization performance

Manuscript description

Existing guidelines in statistical modeling for genomics hold that simpler models have advantages over more complex ones. Potential advantages include cost, interpretability, and improved generalization across datasets or biological contexts. In cancer transcriptomics, this manifests as a preference for small “gene signatures”, or groups of genes whose expression is used to define cancer subtypes or suggest therapeutic interventions. To test the assumption that small gene signatures generalize better, we examined the generalization of mutation status prediction models across datasets (from cell lines to human tumors and vice-versa) and contexts (holding out entire cancer types from pan-cancer data). We compared two simple procedures for model selection, one that exclusively relies on cross-validation performance and one that combines cross-validation performance with regularization strength. We did not observe that more regularized signatures generalized better. This result held across multiple problems and both linear models (LASSO logistic regression) and non-linear ones (neural networks). When the goal of an analysis is to produce generalizable predictive models, we recommend choosing the ones that perform best on held-out data or in cross-validation, instead of those that are smaller or more regularized.

Manubot

Manubot is a system for writing scholarly manuscripts via GitHub. Manubot automates citations and references, versions manuscripts using git, and enables collaborative writing via GitHub. An overview manuscript presents the benefits of collaborative writing with Manubot and its unique features. The rootstock repository is a general purpose template for creating new Manubot instances, as detailed in SETUP.md. See USAGE.md for documentation how to write a manuscript.

Please open an issue for questions related to Manubot usage, bug reports, or general inquiries.

Repository directories & files

The directories are as follows:

• content contains the manuscript source, which includes markdown files as well as inputs for citations and references. See USAGE.md for more information.
• output contains the outputs (generated files) from Manubot including the resulting manuscripts. You should not edit these files manually, because they will get overwritten.
• webpage is a directory meant to be rendered as a static webpage for viewing the HTML manuscript.
• build contains commands and tools for building the manuscript.
• ci contains files necessary for deployment via continuous integration.

Local execution

The easiest way to run Manubot is to use continuous integration to rebuild the manuscript when the content changes. If you want to build a Manubot manuscript locally, install the conda environment as described in build. Then, you can build the manuscript on POSIX systems by running the following commands from this root directory.

# Activate the manubot conda environment (assumes conda version >= 4.4)
conda activate manubot

# Build the manuscript, saving outputs to the output directory
bash build/build.sh

# At this point, the HTML & PDF outputs will have been created. The remaining
# commands are for serving the webpage to view the HTML manuscript locally.
# This is required to view local images in the HTML output.

# Configure the webpage directory
manubot webpage

# You can now open the manuscript webpage/index.html in a web browser.
# Alternatively, open a local webserver at http://localhost:8000/ with the
# following commands.
cd webpage
python -m http.server

Sometimes it's helpful to monitor the content directory and automatically rebuild the manuscript when a change is detected. The following command, while running, will trigger both the build.sh script and manubot webpage command upon content changes:

bash build/autobuild.sh

Continuous Integration

Whenever a pull request is opened, CI (continuous integration) will test whether the changes break the build process to generate a formatted manuscript. The build process aims to detect common errors, such as invalid citations. If your pull request build fails, see the CI logs for the cause of failure and revise your pull request accordingly.

When a commit to the main branch occurs (for example, when a pull request is merged), CI builds the manuscript and writes the results to the gh-pages and output branches. The gh-pages branch uses GitHub Pages to host the following URLs:

• HTML manuscript at https://greenelab.github.io/generalization-manuscript/
• PDF manuscript at https://greenelab.github.io/generalization-manuscript/manuscript.pdf

For continuous integration configuration details, see .github/workflows/manubot.yaml.

License

Except when noted otherwise, the entirety of this repository is licensed under a CC BY 4.0 License (LICENSE.md), which allows reuse with attribution. Please attribute by linking to https://github.com/greenelab/generalization-manuscript.

Since CC BY is not ideal for code and data, certain repository components are also released under the CC0 1.0 public domain dedication (LICENSE-CC0.md). All files matched by the following glob patterns are dual licensed under CC BY 4.0 and CC0 1.0:

• *.sh
• *.py
• *.yml / *.yaml
• *.json
• *.bib
• *.tsv
• .gitignore

All other files are only available under CC BY 4.0, including:

• *.md
• *.html
• *.pdf
• *.docx

Please open an issue for any question related to licensing.