Parse data

[nbedanova]

ADMM cytokine deconvolution function added. It made sense to me to enforce non-negativity in the W (convolution matrix) but let me know your thoughts on that.
You can run the ParseADMM figure to try out the code and visualize results.

[Copilot]

Pull request overview

This PR adds an ADMM-based cytokine deconvolution method to decompose the cytokine factor matrix A into W @ H, where W represents cytokine interactions and H represents effect patterns. The ADMM algorithm includes optional non-negativity constraints on W (enforcing that cytokines only activate, not inhibit) and provides better convergence guarantees compared to gradient-based methods. A companion visualization figure is included to compare the ADMM results with the original factor matrix.

Key Changes:

• Added deconvolution_cytokine_admm() function implementing ADMM optimization with L1 regularization, adaptive penalty parameter, and configurable non-negativity constraints
• Created figureParseADMM.py to visualize the deconvolved matrices and cytokine interaction heatmap
• Reformatted gradient computation in existing deconvolution_cytokine() for better readability

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 12 comments.

File	Description
pf2rnaseq/factorization.py	Added ADMM-based cytokine deconvolution function with comprehensive parameter control, convergence tracking, and detailed logging; reformatted existing gradient computation for clarity
pf2rnaseq/figures/figureParseADMM.py	New figure file to visualize ADMM deconvolution results including deconvolved matrix H, original effects A, and cytokine signaling matrix W as a heatmap

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[aarmey]

@nbedanova let me know if this works better. It looks like your initial matrices were a local minimum, so ADMM would quit early. I see it fitting over multiple iterations now.

[nbedanova]

@nbedanova let me know if this works better. It looks like your initial matrices were a local minimum, so ADMM would quit early. I see it fitting over multiple iterations now.

I think this helped, starting to see some more patterns that make sense. Thanks!