DLPFC Case Study - Transductive

Transductive spatial transcriptomics case study on human DLPFC using Visium data. This repo benchmarks gene-only and histology-guided graph models (VQ-VAE/GAT with DINO/UNI features), with ablations and diagnostics for robust layer-aware clustering.

What is in this repo

• Core models in model/ (including DINO-enabled variants).
• Data loading and preprocessing in data/.
• Training/benchmark scripts in the repo root.
• Ablation and diagnostic scripts for model behavior checks.
• Notebook workflows for iterative experiments.

Quick start

1. Create environment:
   • conda env create -f environment.yml
   • conda activate geometric
     (or your environment name if different)
2. Verify data paths under data/Visium_DLPFC/.
3. Run a benchmark script, for example:
   • python benchmark_dino_gated.py

Common entry points

• benchmark_dino_gated.py — main transductive benchmark pipeline.
• ablation_study_no_dino.py — gene-only / architecture ablations.
• ablation_gate_learning.py — gating behavior ablations.
• diagnostic_dino_failure.py — targeted diagnostics for multimodal failure modes.

Notes

• Most scripts assume local Visium DLPFC files are available.
• Cached image features (if present) are reused for speed.
• Outputs are written as .json, .csv, and figures in root/result folders.