Selected gene
CLEC5A
Amber: mean model; faint amber: individual fits. Neutral: observed donors. Cross-sectional stages, not time or treatment response.
Dataset / SEA-AD
An independent proof of concept using public human brain data. Prima Mente’s challenge atlas is planned for Spring 2027 and was not used here. Challenge ↗
Research question / Alzheimer’s
Clear amyloid.
Protect synapses.
Using SEA-AD data, we built a model of human brain immune cells. One signal worth following: CLEC5A. The next step is a test in living human cells.
Can reducing CLEC5A help microglia remove amyloid without increasing synapse loss?
The signalCLEC5A
The model misses the Low-stage peak.
01 / What we learned
A repeatable signal.
A reason to look closer.
Our ParticleGAN model generates populations of cells at four Alzheimer’s pathology stages. This lets us inspect variation and repeat the same comparison across five separately trained models.
CLEC5A is higher in High-pathology cells than in Not AD cells. That direction holds across our models, donor checks, and a published human cohort.
The model captures an association. Mouse knockout results supply the reason to test reducing it. Human benefit is still an open question.
- 5 / 5
- models agree on the direction
- +0.572
- observed contrast after covariate adjustment
- 0.00628
- published ROSMAP global FDR ↗
Within two donor-supported subtypes. Expression contrast: mean log1p(panel CP10k). The observed effect is +0.489; the model mean is +0.454. Agreement is not proof that a gene causes disease.
02 / How it performs
A partial match
to real cells.
- 65.9%
- gene effects share the observed direction
- 0.502
- correlation of model and observed gene effects
The model captures part of the pattern across 2,000 genes. It often underestimates the size of a change and misses some stage profiles.
These comparisons use training-donor associations. They describe what the model learned—not independent prediction accuracy.
The main limitation: only three Not AD training donors. Sex and APOE imbalance can produce convincing signals. Some disappear after adjustment; CLEC5A keeps its direction.
03 / What to test
Does better clearance
come at a cost?
Test CLEC5A reduction in human stem-cell-derived microglia. Look for improved amyloid removal and preserved synapses.
- 01
Reduce CLEC5A.
Verify the knockdown. Use independent reagents, controls, and several donor backgrounds, including APOE3/APOE4 where feasible.
- 02
Measure what matters.
Track amyloid uptake and degradation separately, alongside viability, inflammation, and synapse loss.
- 03
Require a functional benefit.
More uptake alone is insufficient. A useful result improves removal without harming cells or increasing synapse loss.
An existing lead. A human translation question.
CLEC5A has been studied in mice. Our contribution is a model-backed human association and a focused test—not a new target discovery.
Suggested Prima Mente registration answer
I’d like to test whether reducing CLEC5A in human stem-cell-derived microglia improves amyloid-β clearance without increasing synapse loss. In an independent proof of concept using public SEA-AD data, our trained generative model and donor-level analyses prioritize CLEC5A as a pathology-associated candidate. Published mouse knockout results motivate testing whether the benefit translates to human cells, including across APOE3 and APOE4 backgrounds.
An optional registration proposal for the Prima Mente Alzheimer’s Translation Challenge, based on SEA-AD—not the forthcoming challenge atlas. Not submitted through this project; no affiliation implied.
Go deeperExplore all 2,000 genes Five-seed effects, donor checks, and four-stage profiles.
614 genes pass the exploratory stability rule; 107 also agree with the primary ROSMAP association; 58 also pass donor sensitivity checks. These are filter counts, not significant discoveries.
| Gene | Model Δ | Observed Δ | Adjusted Δ | ROSMAP FDR |
|---|
Make it inspectableMethods, sources & downloads The full recipe, limits, and reproducible evidence.
Five from-scratch ParticleGAN 0.9.7 fits, 20,000 updates each, EMA disabled. Training used 14,762 SEA-AD middle temporal gyrus cells from 33 donors. Only pathology stage is a generator input; the model has no gene-edit, compound, genotype, dose, or time condition.
We generated 12.8 million unique cells from frozen models. Synthetic cells reduce simulation noise; they add no biological donors. We have not established that this generator selects better targets than observed expression plus literature alone.
Calibration sanity check and model errors
Four calibration donors were used during hyperparameter selection. Mean donor energy score: model 2.0069, matched resampling 2.0946. Mean cell energy distance: model 0.2842, matched resampling 0.2349. Lower is better for both. Resampling is an empirical sanity check that reuses observed cells, not a trained neural network or a model-architecture benchmark. These are selection results, not an independent final test.
Within-subtype gene effects have descriptive Spearman correlation 0.502 and 65.9% sign agreement with training-donor effects. The genes are correlated; these are not independent prediction tests. CLEC5A’s stage profile is imperfect, and the model strongly attenuates GPNMB’s effect.
Model settings
512 learned particles; latent dimension 8; per-cell noise dimension 8; exact Gamma–Poisson sampling. Generator LR 0.012; critic LR 0.018; particle LR 0.06; regularization coefficient 1; κ 1; prior regularization 0.01. Released adversarial updates use 150 bag features from 50 PCs. A good projected score does not validate every gene. Frozen plan ↗
06
Make the claim inspectable
Methods, limits, and source evidence.
The biological sample is small.
| Stage | Donors | Male | APOE4 carriers |
|---|---|---|---|
| Not AD | 3 | 2 | 0 |
| Low | 4 | 1 | 0 |
| Intermediate | 8 | 5 | 1 |
| High | 18 | 4 | 12 |
The six non-stage condition slots in the training cache are zero placeholders. We recovered sex and APOE from original metadata for sensitivity checks; those attributes were not generator inputs.
The panel is incomplete.
Only two microglia/PVM subtypes meet the endpoint analysis requirement of at least three donors per stage and 20 cells per donor. Generated subtype assignments are classifier predictions.
Important genes including TREM2, APOE, INPP5D, ACSL1, ABCA1, and NLRP3 are absent from the 2,000-gene panel. The model cannot directly nominate or quantify them. Exact published marker-list coverage ranges from 0/30 antigen-presentation genes to 10/30 chemokine genes; broad pathway conclusions would overreach.
Inference and normalization
Five frozen 20,000-update models × eight Monte Carlo replicates × four stages × 200 bags × 400 cells produced 12.8 million unique synthetic cells. Common latent draws, library-depth templates, and Gamma/Poisson random seeds were shared across stages. Depth templates were drawn uniformly over training donors and were the same across stage queries. Gene expression is per-cell log1p(10,000 × gene count / total count across the measured panel), averaged within donor/subtype and then equally across donors and the two eligible subtypes. It is not whole-transcriptome absolute abundance.
A recovery preserved 87 completed cases and generated the remaining 73. Three whole-case replays, totaling an additional 240,000 physically generated cells, matched the retained arrays bit for bit. All 160 case hashes and 20 input hashes were checked; no model weights were updated. The unique inference total remains 12.8 million.
Exploratory ranking and donor sensitivity
The ranking heuristic is the smaller of the absolute observed and five-seed mean model effects, multiplied by model-seed direction agreement and real-donor deletion agreement. It is zero unless at least four of five models, 80% of 40 Monte Carlo contrasts, and 80% of the original groupwise donor deletions agree with the observed direction; the mean model direction must agree too. This is not a p-value, causal score, or drug score.
For the additional sensitivity filter, we fitted within each eligible subtype: donor mean expression ~ High indicator + standardized age + sex + APOE4 dosage + standardized postmortem interval, using the 21 High/Not AD donors. The adjusted direction must agree with the original contrast, and unadjusted and adjusted donor-deletion sign agreement must each be at least 80%, separately by subtype and deleted donor stage. With three controls, all three control deletions must agree. The six-parameter fits are full rank, but no APOE4 controls means adjustment relies on assumptions and extrapolation. No new gene-level p-values were calculated.
External cohort support is association, not intervention validation
The primary comparison uses published microglial general-pathology coefficients from the ROSMAP prefrontal-cortex study. Support means the same direction as the model/local association and published global FDR < 0.05. The author analysis adjusts for age, sex, postmortem interval, and latent factors. The study is a separate cohort/region from SEA-AD, but we have not performed a subject-identity crosswalk. Published amyloid and tangle associations are secondary checks from the same cohort, not additional independent replications.
We used the author repository at commit 28388f829be6affa1578e56fa09c4be8cd273817. The complete preprocessing and differential-expression pipeline has not been reproduced here. The ROSMAP microglial association can still include within-cell-class composition effects.
How to reproduce this report
The public report builds from committed aggregate results without raw cells, model checkpoints, or a GPU. Run python3 scripts/report/build_report.py, then python3 scripts/report/build_report.py --check. The original full stage-profile export requires the retained inference artifacts; the resulting anonymous profiles are committed. Research snapshot: 89fa5ea.
Builder, export script, and instructions ↗ · Independent numerical audit ↗
All 2,000 genes ↓Effects, seed means, and evidence flags · CSVReport data ↓Anonymous aggregates and stage profiles · JSONProvenance manifest ↓Input and generated-artifact SHA256 hashes
Primary sources & research artifacts
- SEA-AD data resource. Source of the human middle temporal gyrus expression data used for training.
- Mathys et al., Cell (2023). Single-cell atlas of cognition, dementia, and resilience; 427 ROSMAP donors. Pinned author data/code.
- CLEC5A-deficiency study, Journal of Neuroinflammation (2024). Mouse/cell evidence motivating a human-cell reduction experiment.
- Transcriptional regulation of disease-relevant microglial activation programs, Neuron (2026). Human microglial perturbation context; not direct CLEC5A intervention evidence.
- Plasticity of human microglia and brain perivascular macrophages in aging and Alzheimer’s disease, Nature Genetics (2026). Context for GPNMB-associated states and function; not treated as a third independent cohort.
- TOMMORROW trial, The Lancet Neurology (2021). Negative primary prevention result for low-dose pioglitazone.
- Prima Mente Alzheimer’s Translation Challenge. Challenge context; no affiliation implied.
- Our committed analyses: gene ranking, donor sensitivity, model quality, and recommendation.
Additional scientific figures
