AI-Assisted Bayesian Inference Found Gamma-Secretase Alzheimer Pathways

TL;DR: A 2026 medRxiv preprint applied ChatGPT-4o-assisted Bayesian-frequentist hybrid inference to Alzheimer’s single-nucleus RNA-seq data and identified gamma-secretase and HP1 transcription pathways that a no-evidence frequentist setting did not recover.

Source: medRxiv (2026) | Han et al.

Alzheimer’s single-nucleus RNA sequencing can measure gene expression in specific brain-cell populations, but thousands of genes create a statistical power problem. A conventional model may miss pathways when sample size, technical variation, and multiple-testing correction all work against detection.

This preprint tested whether a structured AI-assisted prior could help. ChatGPT-4o was used to rate existing evidence linking each gene to Alzheimer’s disease, then those ratings were converted into prior distributions for a hybrid Bayesian model.

ChatGPT-4o Ratings Became Alzheimer’s Gene Priors

The method separated the Alzheimer’s case-control parameter from other covariates. The Alzheimer’s parameter received an informative prior, while age, sex, APOE status, intercept, and residual variance were handled as frequentist parameters.

The AI step used a standardized prompt rather than open-ended interpretation. Each gene was rated on an evidence scale, then mapped to a Cohen’s d effect-size prior.

• Level 1: No evidence, mapped to Cohen’s d = 0.
• Level 2: Weak evidence, mapped to Cohen’s d = 0.2.
• Level 3: Moderate evidence, mapped to Cohen’s d = 0.5.
• Level 4: Moderate-to-strong evidence, mapped to Cohen’s d = 0.9.
• Level 5: Strong evidence, mapped to Cohen’s d = 1.5.

That mapping is the core editorial point. The language model did not diagnose Alzheimer’s disease or inspect brain tissue; it converted literature context into a structured statistical input.

ROSMAP Inhibitory Neurons Provided the Alzheimer’s Test Case

The application used public single-nucleus RNA-seq data from 427 postmortem prefrontal cortex samples in ROSMAP. Researchers dichotomized Alzheimer’s pathology by Braak stage, comparing mild/none pathology with severe pathology.

A second prefrontal cortex dataset from Lau et al. supplied variability estimates for the prior construction. The final application focused on inhibitory neurons, a cell type relevant to brain-circuit regulation.

1. ROSMAP analysis set: 203 severe Alzheimer’s samples and 113 mild/none samples in inhibitory neurons.
2. Prior-variance source: 12 severe and 9 mild/none samples from the Lau dataset.
3. Gene filtering: 13,649 inhibitory-neuron genes were narrowed to 3,885 genes for prior generation.

The model adjusted for baseline age, sex, and APOE E4 carrier status. Alzheimer’s gene-expression analysis needs those covariates because age and APOE can strongly shape disease-associated transcription patterns.

No-Evidence Frequentist Analysis Found No FDR-Significant Genes

The comparison across prior settings gives the preprint its practical result. In the no-evidence/frequentist setting, no genes had FDR less than 0.2.

The non-informative prior setting also did not identify significant pathways. The pathway signal appeared only when the analysis used informative AI-assisted priors based on Alzheimer’s evidence.

• No evidence setting: No genes passed the reported FDR threshold.
• Non-informative prior: No significant pathways were identified.
• Informative prior: Alzheimer’s-relevant pathways appeared in the pathway analysis.

This is a sensitivity result, not proof that every AI-assisted gene prior was biologically correct. It shows how prior information can change what becomes detectable in a high-dimensional RNA analysis.

Gamma-Secretase Pathways Emerged With Informative Priors

With the informative prior, two gamma-secretase pathways were highly ranked. Gamma-secretase proteolytic targets had q = 5.75E-04, and gamma-secretase regulation of neuronal cell development and function had q = 6.81E-04.

That is biologically plausible because gamma-secretase cleaves amyloid precursor protein, helping generate amyloid-beta peptides. Amyloid-beta plaques are a core Alzheimer’s pathology, although this RNA analysis does not measure plaque burden directly.

A third pathway, ACM1/ACM3/ACM5 signaling in the brain, reached q = 0.04 with the informative prior and q = 0.07 with the neuron-specific informative prior. Those values put the pathway inside or near the preprint’s pathway significance range.

• Gamma-secretase target pathway: q = 5.75E-04 with the informative prior.
• Gamma-secretase neuronal pathway: q = 6.81E-04 with the informative prior.
• Brain cholinergic signaling pathway: q = 0.04 with the informative prior.

The important boundary is that pathway enrichment is downstream of the model. It can organize gene-level results into biological themes, but it does not prove a pathway is driving disease progression.

Neuron-Specific Priors Highlighted HP1 Transcriptional Silencing

The neuron-specific informative prior changed the pathway ranking. The HP1 family transcriptional-silencing pathway reached q = 3.65E-03, while it was not below 0.1 in the informative-prior column.

HP1 is tied to heterochromatin, a compact DNA-protein state involved in gene regulation. Alzheimer’s research has connected chromatin remodeling and tau-related pathology, so the HP1 result fits a plausible disease-biology context.

1. Cell-type context: Neuron-specific priors were intended to make the literature input more relevant to inhibitory neurons.
2. HP1 pathway: The q value improved to 3.65E-03 under the neuron-specific informative prior.
3. Axon/synapse signal: Neurofilaments in axon growth and synapses moved close to significance at q = 0.1.

The analysis therefore produced two related messages: known Alzheimer’s biology became detectable, and neuron-specific priors shifted attention toward cell-type-relevant regulatory pathways.

AI-Assisted Priors Need Transparent Validation Before Biology Claims

The caution is straightforward. A language model can summarize literature associations, but it can also inherit uneven literature coverage, database bias, and prompt sensitivity.

The preprint partly handles that risk by using a standardized prompt and a defined mapping from evidence levels to priors. Still, any claimed pathway needs independent replication, preferably with preregistered prior rules and external datasets.

• Best use: Hypothesis generation for high-dimensional omics analyses where power is limited.
• Main risk: Overconfident priors can pull estimates toward literature expectations.
• Next test: Repeating the pipeline across additional Alzheimer’s cell types, cohorts, and prompt-locked prior files.

For Alzheimer’s research, the study does not show ChatGPT-4o discovering a new disease mechanism by itself. The stronger claim is narrower: AI-structured priors helped a hybrid statistical model recover plausible Alzheimer’s pathways from inhibitory-neuron RNA data.