Compressed Transcriptomics

This repository open sources code and model weights for automated optimum compression of brain transcriptomic data using deep autoencoding, as detailed in our article.

Table of Contents

• What is this repository for?
• Why should I use this model?
  • Expressivity
  • Accuracy
  • Downstream inference
• Usage instructions
• Use queries
• Citation
• Funding

What is this repository for?

• The architecture of the brain is too complex to be intuitively surveyable without the use of compressed representations that project its variation into a compact, navigable space.
• The task is especially challenging with high-dimensional data, such as transcriptomic / gene expression data (for example provided by the Allen Brain Atlas) where the joint complexity of anatomical and transcriptional patterns demands maximum compression.
• The current established practice is to use standard principal component analysis (PCA), which comes at the cost of limited expressibility, which can impact quality of research findings.
• As an alternative, we here provide the code and model weights for a series of trained deep auto-encoders.

Why should I use this model?

Expressivity

Even when reducing 15,633 gene expression values into merely 2 components, note the intricate variational structure across auto-encoded (AE) representations, especially in comparison to methods such as PCA, kPCA, or NMF.

Accuracy

Across a number of possible latent dimensionalities, the auto-encoder achieves superior accuracy in reconstructing original source transcriptomic data in comparison to UMAP, PCA, NMF, and kPCA.

N.B. source reconstruction evaluation is presently not possible using t-SNE.

Downstream inference

Deep auto-encoders yield superior predictive utility across signalling, microstructural, and metabolic targets in applied machine prediction using transcriptomic data.

Usage instructions

In the uploaded Jupyter Notebook we provide a tutorial that:

• 1. Overviews running inference to compress transcriptomic data into an auto-encoded latent space of 2,4,8,16,32,64, or 128 dimensions.
• 2. Provides the code and architecture to retrain the a auto-encoder for use with other data.

The model weights are open-sourced here.

The NIFTI images depicting the two component latents of all models in both 4mm and 8mm space are open-sourced here.

Use queries

Via github issue log or email to j.ruffle@ucl.ac.uk.

Citation

If using these works, please cite the following paper:

Compressed representation of brain genetic transcription. James K Ruffle, Henry Watkins, Robert J Gray, Harpreet Hyare, Michel Thiebaut de Schotten, Parashkev Nachev. Human Brain Mapping, 2024. DOI https://doi.org/10.1002/hbm.26795. 

Funding

The Medical Research Council; The Wellcome Trust; UCLH NIHR Biomedical Research Centre; Guarantors of Brain; British Society of Neuroradiology.