Plate IV
HCGT-PG
HCGT-PG: Hierarchical Counterfactual Graph Transformer, Perturbation-Grounded — causal biomarker discovery in Alzheimer's disease microglia
[SUPPLY]
Correlation is cheap here
Single-cell data makes correlation almost free. Thousands of genes track with Alzheimer’s pathology, and nearly none of that tells you what would happen if you actually changed one. The question worth asking is interventional: if you knocked this gene down in a microglial cell, what would happen to its disease state?
This is not GWAS data. Readers assume it is, and the assumption changes what they think the estimand is.
The estimand
HCGT-PG estimates the causal effect of a gene knockdown on a microglial disease-state score — and then checks that estimate against CRISPR screens where the knockdown was really performed. The validation is the point: a causal claim that cannot be checked against an actual intervention is a correlation with better vocabulary.
The data
- Morabito et al. 2021 — paired snRNA-seq and snATAC-seq multiome.
- ROSMAP — cohort-scale expression with neuropathology.
- Dräger et al. 2022 — CRISPRi/a CROP-seq screens in human microglia, the ground truth for the interventional check.
- Liu et al. 2025 — epigenomic atlas.
Status
The thesis proposal is accepted. The programme begins Fall 2026 at BRAC University under Dr. Badhan Das. There are no results yet; the design is the contribution.