Modeling Alzheimer's Cell States with ParticleGAN
DiffLab's open research explores how generative models of brain immune cells can help prioritize Alzheimer's experiments, inspired by the upcoming ALZ challenge.
Can AI help turn Alzheimer's data into a focused experiment? Our new ALZ research report explores that question using ParticleGAN to model microglia, the brain's immune cells.
The ALZ challenge
The hackathon-style Prima Mente Alzheimer's Translation Challenge is a global AI competition connecting computational models with laboratory testing. Starting in spring 2027, participants will build models on a planned atlas of 150 million cells. Strong performers advance to proposing therapeutic hypotheses; finalists' ideas will be tested in the lab.
We started exploring that model-to-experiment workflow now. This is an independent proof of concept using public SEA-AD human brain data, not the forthcoming challenge atlas or an official challenge submission.
Our approach
We trained five independent ParticleGAN models on 14,762 cells from 33 human donors, covering 2,000 genes across four Alzheimer's pathology stages. Rather than modeling an average cell, we generated cell populations to examine variation and compare gene-expression patterns across stages.
We checked those patterns against observed donors, sensitivity analyses, and published human-cohort evidence. One candidate stood out: CLEC5A was higher in High-pathology than Not AD cells, with all five models agreeing on the direction and support from a published ROSMAP association.
CLEC5A is an existing biological lead, not a target we discovered. Published mouse knockout work motivates the next question: could reducing CLEC5A help human microglia clear amyloid without increasing synapse loss? That requires an experiment in living human cells.
What is ParticleGAN?
A generative adversarial network, or GAN, trains a generator against a critic that compares generated samples with real data. ParticleGAN changes the generator's starting points: instead of drawing only from a fixed noise distribution, it learns a collection of latent vectors called particles alongside the generator. The aim is to better represent distinct patterns in complex data.
Here, those patterns are cell populations conditioned on pathology stage. The model does not simulate gene edits, predict drug responses, or establish causation. Only three training donors were Not AD, and some stage profiles remain imperfect.
Explore the report and gene explorer or inspect the open-source code and evidence. The output is a testable hypothesis, with its supporting evidence and limitations available for inspection.