Public code leaf
Matched neighbourhood controls and label diagnostics for spatial encoders
This page describes the public repository PeterPonyu/tessera-st. The analysis code, the protocol and its departures are on the analysis branch linked below; code, derived inputs and result tables are archived on Zenodo.
What the code does
It asks what a learned spatial encoder adds beyond averaging each observation with its neighbours, and how much the construction of a reference annotation decides that comparison before any method is run.
- Label diagnostics from a reference annotation and its coordinates: chance-corrected homophily across scales, label informativeness, a label correlation length, and how well each observation's own expression predicts its label
- Tuned neighbourhood-mean controls, selected on other patients with the same rule as each encoder's settings
- Parameter-free propagation on each encoder's own graph, matched to its message-passing depth
- Unmodified runs of GraphST and STAGATE through adapters, beside the Tessera-ST gated graph encoder
- Component removals for all three encoders, including spatial positives without message passing
Tessera-ST is one of the three encoders compared.
Code and archives
Analysis code, protocol and departures
Zenodo archive (code, derived inputs and result tables)
Analysis protocol, deposited before the prospective cohorts’ expression data were downloaded
git clone https://github.com/PeterPonyu/tessera-st cd tessera-st git checkout revision-r3
The repository README lists the commands for each step, from the saved result tables to complete refits. Complete refits need the public source data; the README explains how to obtain and check them.
Source: github.com/PeterPonyu/tessera-st.