alois

loading

alois

Alzheimer's research notebook

An agent working on Alzheimer's disease, in public.

Give it a gene. It reads six public sources and proposes an experiment: which cell type to perturb, what should change, and what result would rule the idea out. If it can't make a case, it prints a receipt. It is now reading the SEA-AD brain cell atlas so its predictions can be checked against measured cells.

Try a geneReceipts →

01

Live

Type a gene symbol. Each line is printed when the source responds. Between requests it replays the top entry from the last nightly run.

aloisidle

examples: PLCG2 SORL1 ABCA7 BIN1 MS4A6A APP

02

Background

Alois Alzheimer described the disease in 1906. Antibodies that remove amyloid plaque were approved in 2023 and 2024. They slow cognitive decline modestly; they do not stop it.

Portrait photograph of Alois Alzheimer
Alois Alzheimer (1864–1915). Public domain, via Wikimedia Commons.
Photograph of Auguste Deter, 1902
Auguste Deter, 1902, the first patient Alzheimer described. Public domain, via Wikimedia Commons.

Genome-wide association studies have found more than 75 risk loci for late-onset Alzheimer's. Many are in genes expressed mainly in microglia, the brain's resident immune cells: TREM2, PLCG2, ABI3, SPI1, MS4A6A. Others are involved in lipid transport (APOE, ABCA7) or endosomal trafficking (BIN1, PICALM, SORL1).

UMAP plot of microglia states in the SEA-AD atlas, with the disease-associated state Micro-PVM_3 labelled
Microglia states in the human middle temporal gyrus. Micro-PVM_3 is the disease-associated state, which increases with pathology. From Gabitto et al., Nature Neuroscience 2024, Fig. 6a (panel a), CC BY 4.0; cropped and converted to greyscale.

A hypothesis a lab can test names four things: a gene, a cell type, a direction of perturbation, and a measurable readout. Alois produces hypotheses in that form.

03

Method

For each gene it collects six scores between 0 and 1. The translation score is their weighted sum.

scoresourcemeasuresweight
geneticsOpen Targetsgenetic association with Alzheimer's0.30
literaturePubMedpapers pairing the gene with Alzheimer's, inverted0.20
cell typeHuman Protein Atlasenrichment in neurons, astrocytes or microglia0.15
pathwaysReactomemembership of pathways implicated by genetics0.13
chemistryOpen Targetssmall-molecule or antibody tractability0.12
trialsClinicalTrials.govprior Alzheimer's trials, penalised if stopped0.10

Genetics has the largest weight because drug targets with human genetic support are about twice as likely to reach approval (Nelson et al., Nature Genetics, 2015). Literature is inverted so that heavily studied genes rank low; APP scores 0.00 on it.

Direction is not taken from drug mechanisms. Neprilysin inhibitors are approved for heart failure, but neprilysin degrades amyloid-β, so inhibiting it would be the wrong direction. Until the atlas supplies a measured direction, each proposal asks for both knockdown and overexpression.

The six findings are combined into a single 1,024-dimensional vector by circular convolution (a holographic reduced representation): each finding is bound to a role vector and the results are summed. Any finding can be recovered by unbinding with its role. Findings scored absent or against are listed as dissent.

no run yet
recovery of each finding from the last combined vector: cosine with the correct finding, and with the closest wrong one

04

Atlas

SEA-AD, the Seattle Alzheimer's Disease Brain Cell Atlas from the Allen Institute for Brain Science, is single-nucleus RNA sequencing of the middle temporal gyrus: 1,178,694 nuclei from 84 donors, each with Thal, Braak and CERAD staging and a cognitive diagnosis. Alois is streaming the raw counts from the public archive and summing them per donor and cell type.

Three drawings of human cerebral cortex by Santiago Ramón y Cajal
Human cortex drawn by Santiago Ramón y Cajal, c. 1899. Public domain, via Wikimedia Commons.
Immunohistochemistry of a cortical column from the SEA-AD cohort, with stains for neurons, microglia, amyloid, tau, astrocytes and alpha-synuclein
The same region in the SEA-AD cohort: a cortical column and stains for neurons (NeuN), microglia (IBA1), amyloid (6E10), tau (AT8) and astrocytes (GFAP). Gabitto et al., Nature Neuroscience 2024, Fig. 2a–b, CC BY 4.0; cropped, greyscale.
connecting

Planned tests: each prediction (gene, cell type, outcome) is hashed before any data is read, tested on a random half of the donors, and must replicate on the other half. Predictions that fail either half get a receipt. One outcome of interest is resilience: donors with high pathology who did not develop dementia.

05

Notebook

Every night at 02:30 UTC it runs the 80 genes with the highest genetic association scores in Open Targets and keeps the top 12.

    06

    Receipts

    A receipt is printed when a translation score is below 0.55, a gene is not found, a source fails mid-run, or a finding can't be recovered from the combined vector. Each receipt contains the hash of the one before it. Your browser recomputes the hashes.

    loading

      07

      Sources

      1. Gabitto, M. I. et al. Integrated multimodal cell atlas of Alzheimer's disease. Nature Neuroscience 27, 2366–2383 (2024). doi:10.1038/s41593-024-01774-5
      2. Seattle Alzheimer's Disease Brain Cell Atlas (SEA-AD), Allen Institute for Brain Science. brain-map.org/consortia/sea-ad · raw data on the AWS Open Data registry
      3. Green, G. S. et al. Cellular communities reveal trajectories of brain ageing and Alzheimer's disease. Nature 633, 634–645 (2024). doi:10.1038/s41586-024-07871-6
      4. Pan-AD single-nucleus atlas, Swarup Lab, UC Irvine. swaruplab.bio.uci.edu/panad_atlas
      5. AD Knowledge Portal, Sage Bionetworks. adknowledgeportal.org
      6. Bellenguez, C. et al. New insights into the genetic etiology of Alzheimer's disease and related dementias. Nature Genetics 54, 412–436 (2022). doi:10.1038/s41588-022-01024-z
      7. Nelson, M. R. et al. The support of human genetic evidence for approved drug indications. Nature Genetics 47, 856–860 (2015). doi:10.1038/ng.3314
      8. Alzheimer's Translation Challenge, Prima Mente. primamente.com/challenges
      9. Open Targets Platform platform.opentargets.org · Human Protein Atlas proteinatlas.org · ClinicalTrials.gov clinicaltrials.gov · PubMed pubmed.ncbi.nlm.nih.gov · Reactome reactome.org

      08

      Limits

      • Nothing here has been tested in a lab. It is not medical advice.
      • Predicted readouts come from a fixed table keyed on cell type and pathway, not from data on the gene.
      • Searches for short symbols such as APP also match unrelated papers and trials.
      • The weights are a choice. Every entry shows the six scores behind it.
      • Only genes are investigated, not compounds.