Reproduce “Compressed representation of brain genetic transcription” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
James K Ruffle, Henry Watkins, Robert J Gray, Harpreet Hyare, Michel Thiebaut de Schotten, Parashkev Nachev
The authors compare auto-encoder compression of brain voxel gene-expression profiles with PCA and other methods. Its encoder maps each profile to a compact representation, and its decoder reconstructs it. On held-out voxels, the auto-encoder has the lowest reconstruction error. Across 24 prediction targets, it has the best mean scores, but does not significantly outperform t-SNE or UMAP. The voxel split does not test generalization to new people.
Key insights
- Deep auto-encoders consistently generate compressed representations that capture anatomically plausible spatial organization at multiple scales, outperforming traditional linear methods.
- Principal Component Analysis (PCA) is inadequate for representing the non-linear and complex nature of brain genetic transcription, yielding spatially reductive and biologically implausible maps.
- The ability to reconstruct original data from a compressed representation, a key feature of auto-encoders, is crucial for verifying representational fidelity and enabling generative modeling, distinguishing them from other effective non-linear methods like t-SNE and UMAP.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.