Cell-Type-Resolved Pseudobulk Classification Across Independent Cohorts Identifies Microglial PTPRG as a Transcriptional Hub in Alzheimer's Disease

By applying a cross-cohort validated machine learning approach to single-nucleus RNA sequencing data, this study identifies microglial PTPRG as a central transcriptional hub that integrates neuronal signaling and inflammatory dysregulation in Alzheimer's disease.

The Gist

The Big Picture: Finding the "Smoking Gun" in Alzheimer's

Imagine the human brain as a massive, bustling city with millions of different workers (cells) doing specific jobs. Some are the city planners (neurons), some are the sanitation crew (microglia), and others are the maintenance staff (astrocytes).

In Alzheimer's disease (AD), this city starts to break down. For a long time, scientists looked at the city as a whole, taking a "blurry photo" of everything happening at once. This made it hard to tell exactly who was causing the trouble or what specific job they were failing at.

This new study decided to stop taking blurry photos. Instead, they put on high-powered glasses to look at each type of worker individually. They wanted to answer two big questions:

1. Which workers are the best at telling us if the city is in trouble?
2. Is there one specific "boss" worker whose behavior changes the most when the city gets sick?

The Method: The "Pseudobulk" Team Huddle

The researchers used a clever trick called pseudobulk analysis.

Imagine you have a huge stadium full of fans (cells). If you ask every single fan what they are thinking, the data is chaotic and noisy. Instead, this study asked the fans to form small groups based on their team jerseys (cell types: Microglia, Astrocytes, Neurons, etc.). Then, they asked each group to combine their thoughts into one single, clear "team report."

By treating each group as one unit, they could compare the "Microglia Team Report" against the "Astrocyte Team Report" to see which group had the clearest story about the disease.

The Discovery: The Glial "Alarm System"

When they tested every possible combination of these teams to see which could best predict Alzheimer's, they found a winner: The Glial Duo.

• Microglia (the brain's immune cells/sanitation crew) and Astrocytes (the support staff) were the most accurate predictors.
• Even though neurons (the brain's main workers) are the ones dying in Alzheimer's, the support staff were the ones screaming the loudest about the problem.

The computer model built using just these two teams was incredibly accurate. It could tell the difference between a healthy brain and an Alzheimer's brain with about 87-89% accuracy, and it worked just as well on a completely different group of people (a different "city").

The Star Player: PTPRG (The Brake Pedal)

Out of hundreds of genes (the instructions inside the cells) the model looked at, one stood out as the most important. It was a gene called PTPRG, found in the Microglia.

The Analogy:
Think of Microglia as a car.

• In a healthy brain, PTPRG is the brake pedal. It keeps the immune system calm and stops it from revving up unnecessarily.
• In an Alzheimer's brain, the brake pedal (PTPRG) is broken or missing. Without it, the Microglia slam the gas pedal. They go into overdrive, screaming "FIRE!" when there's only a small spark. This causes chronic inflammation that damages the brain.

The study found that in Alzheimer's patients, the "brake" (PTPRG) wasn't just turned down; the entire neighborhood around it changed. In healthy brains, PTPRG was part of a quiet, organized neighborhood focused on maintenance. In Alzheimer's brains, it was surrounded by a chaotic, angry mob of inflammatory signals.

The Connection: Neurons Ringing the Doorbell

The researchers also asked: Who is telling the Microglia to hit the gas?

They found that the Neurons (the city planners) were sending signals to the Microglia.

• Excitatory Neurons (the active workers) were ringing the doorbell loudly, triggering an immune response.
• Inhibitory Neurons (the calm-down workers) were whispering, but their message was different, focusing more on stress and metabolism.

Crucially, the "doorbells" being rung included famous Alzheimer's risk genes like APOE and PSEN1. This suggests that the genetic risks people inherit from their parents cause their neurons to send the wrong signals, which breaks the Microglia's brakes, leading to the disease.

The "Continuum" Insight

One of the coolest findings was about people in the middle stage of the disease (Mild Cognitive Impairment).

• The model didn't just say "Sick" or "Healthy."
• It placed these middle-ground patients on a sliding scale. Their "sickness score" was right in between the healthy people and the severe Alzheimer's patients.
• This proves that Alzheimer's isn't a switch that flips overnight; it's a dimmer switch that slowly gets brighter (or darker, in this case) over time.

Why This Matters

1. New Targets: Instead of trying to fix the neurons directly (which is hard), maybe we can fix the brake pedal (PTPRG) in the Microglia. If we can restore the brakes, we might stop the inflammation from destroying the brain.
2. Better Diagnosis: We might be able to use these specific "team reports" from blood or tissue samples to catch the disease earlier, even before memory loss gets bad.
3. A New Map: This study gives us a new map of the brain, showing us that to understand Alzheimer's, we have to listen to the support staff (glia), not just the main workers (neurons).

In short: The brain's immune system loses its brakes in Alzheimer's. This study found the specific broken part (PTPRG) and showed that the neurons are the ones accidentally stepping on the gas. Fixing the brakes could be the key to slowing down the disease.

Technical Summary

1. Problem Statement

Alzheimer's Disease (AD) is a complex neurodegenerative disorder involving heterogeneous cell populations. While single-nucleus RNA sequencing (snRNA-seq) has revealed cell-type-specific changes, applying supervised machine learning to single-cell data presents a fundamental methodological challenge: the labeling problem. Clinical diagnoses (AD vs. Non-Cognitively Impaired, NCI) are assigned at the sample level, not the cell level.

• Current Limitations: Assigning sample-level labels to individual cells is methodologically unsound because not all cells in a diseased brain are equally affected. Conversely, pooling all cells into a single "bulk" profile obscures cell-type-specific contributions. Existing methods either ignore cell-type identity or analyze cell types independently, failing to capture coordinated cross-cell-type disease mechanisms.
• Goal: To develop a robust, unbiased framework that identifies which specific brain cell types drive AD transcriptional signatures, quantify their relative contributions, and identify key molecular drivers (genes) that generalize across independent cohorts.

2. Methodology

The authors employed a pseudobulk aggregation strategy combined with supervised machine learning and multi-omics network analysis.

• Data Sources:
  • Training/Discovery: ROSMAP cohort (367 subjects).
  • Validation: Seattle Alzheimer's Disease Brain Cell Atlas (SEAD) cohort (48 subjects).
• Preprocessing & Phenotyping:
  • Six major cell types were analyzed: Microglia, Astrocytes, Excitatory Neurons, Inhibitory Neurons, Oligodendrocyte Precursor Cells (OPCs), and Oligodendrocytes.
  • Phenotype Definition: Strict criteria were used to define "extreme" groups to maximize class separation:
    • AD: High cognitive decline (cogdx=4), high Braak stage (≥4), low CERAD (≤2).
    • NCI: No cognitive impairment (cogdx=1), low Braak (≤3), high CERAD (≥3).
    • MCI+P: Mild cognitive impairment with plaques (used for continuum analysis).
  • Pseudobulk Aggregation: Gene expression was summed per subject per cell type (normalized to $10^6$ counts) to create subject-level vectors, preserving cell-type identity while avoiding cell-level labeling errors.
• Classification Framework:
  • Combinatorial Search: All 63 possible combinations of the 6 cell types were tested.
  • Model: L1-regularized (Lasso) Logistic Regression.
  • Feature Selection: Highly Variable Genes (HVGs) were selected within each cross-validation fold. Only genes consistently selected across all 5 folds were retained for the final model.
  • Validation: Performance was evaluated on a held-out ROSMAP test set and the independent SEAD cohort using Balanced Accuracy, F1-score, and AUC.
• Downstream Analyses:
  • WGCNA (Weighted Gene Co-expression Network Analysis): Performed separately on AD and NCI microglia to identify co-expression modules surrounding the top gene.
  • NicheNet: Used to predict neuronal ligands regulating microglial PTPRG.
  • Correlation Analysis: Linear regression models assessed the coupling between neuronal gene expression and microglial PTPRG, controlling for covariates (age, sex, PMI, diagnosis).

3. Key Contributions

1. Methodological Innovation: Introduced a systematic, unbiased pseudobulk classification framework that evaluates all cell-type combinations to determine the most discriminative cellular signatures for AD, avoiding the pitfalls of cell-level labeling.
2. Cell-Type Hierarchy: Established that Microglia and Astrocytes are the primary drivers of the AD transcriptional signal, outperforming neuronal subtypes in classification tasks.
3. Discovery of PTPRG: Identified Microglial PTPRG (Protein Tyrosine Phosphatase Receptor Type G) as the single most predictive gene for AD classification, acting as a central hub.
4. Cross-Cohort Generalizability: Demonstrated that a model trained on ROSMAP data (using only 228 genes) achieves high performance (AUC 0.89) on an independent cohort (SEAD, AUC 0.92), a rarity in computational AD research.
5. Mechanistic Insight: Revealed that PTPRG operates in fundamentally different gene network contexts in AD (inflammatory) vs. NCI (homeostatic) and serves as a convergence point for AD risk genes (APOE, GRN, PSEN1, CLU) originating from neurons.

4. Key Results

• Classification Performance:

  • The optimal model used Microglia + Astrocytes (500 HVGs each).
  • ROSMAP Test Set: Balanced Accuracy = 0.87, AUC = 0.89, Sensitivity (AD) = 93%.
  • SEAD Validation: Balanced Accuracy = 0.86, AUC = 0.92, Specificity (NCI) = 100%.
  • MCI+P Continuum: Intermediate samples fell between AD and NCI probabilities, confirming the model captures a disease gradient rather than a binary switch.

• Feature Importance:

  • The final model selected 72 non-zero genes (47 from Microglia, 25 from Astrocytes).
  • Top Gene: PTPRG (Microglia) had the highest absolute coefficient.
  • Other Key Genes: Microglial CXCR4, FLT1; Astrocytic IGFBP5, SORCS3.

• Biological Characterization:

  • GSEA: Microglial genes were enriched for chronic immune activation and inflammatory signaling. Astrocytic genes showed protein homeostasis stress and HSF1-mediated chaperone pathways.
  • WGCNA:
    • AD Microglia: PTPRG co-expressed with 801 genes in a module dominated by inflammatory pathways (Toll-like receptors, Interleukins).
    • NCI Microglia: PTPRG co-expressed with 1,168 genes in a homeostatic module (mTORC1, oxidative phosphorylation).
    • Overlap: Only 110 genes were shared, indicating a near-complete reorganization of the PTPRG regulatory network in AD.
  • Cell-Cell Communication (NicheNet):
    • Neuronal ligands predicted to regulate PTPRG include major AD risk genes: APOE, GRN, PSEN1, CLU.
    • Excitatory vs. Inhibitory Neurons:
      • Excitatory: Strong coupling to PTPRG via immune activation and antigen presentation (Regression $\beta \approx 2.3$).
      • Inhibitory: Weak coupling via lipid metabolism and stress response (Regression $\beta \approx 0.15$).

5. Significance

• Therapeutic Target: PTPRG is identified as a critical negative regulator of innate immune activation. Its downregulation and network reorganization in AD suggest that restoring PTPRG-mediated phosphatase activity could resolve chronic neuroinflammation.
• Mechanistic Model: The study proposes a coherent model where neuronal dysfunction (driven by AD risk genes) signals to microglia via ligand-receptor interactions, converging on PTPRG. This triggers a shift from homeostatic surveillance to inflammatory dysregulation, while astrocytes simultaneously struggle with proteostasis stress.
• Clinical Utility: The ability to classify AD with high accuracy using a small set of glial genes suggests potential for transcriptomics-informed patient stratification and staging, particularly for identifying intermediate disease states (MCI).
• Generalizability: The pseudobulk classification framework is applicable to other neurodegenerative diseases, offering a robust strategy to dissect complex cellular contributions without the noise of single-cell sparsity or the bias of manual cell-type selection.

In summary, this paper moves beyond descriptive differential expression to provide a predictive, cell-type-resolved mechanistic map of AD, pinpointing Microglial PTPRG as the central node integrating neuronal risk signals and inflammatory pathology.