# Hongkui Zeng Review
## Initial reaction
**Greenlight: yes, with important refinements to make this maximally leverage our infrastructure.**
This proposal asks exactly the right kind of question for the Allen Institute: it takes two major community resources we built — the Connectivity Atlas and the Cell Type Atlas — and tests whether they have predictive power for disease biology. That is what these resources are for: not just description, but prediction and mechanism.
## What I like
- The core hypothesis is strong and testable: **connectivity motifs carry information about selective vulnerability**. This is what Seeley's group showed at the macroscale for human FTD/AD (Seeley et al., *Neuron* 2009), but nobody has tested it at cell-type resolution using a systematic connectivity matrix.
- The choice of 5xFAD is pragmatic: early onset, fast enough for a 3-month pilot, and extensive existing scRNA-seq data to benchmark against.
- The proposal uses our Cre lines exactly as they were designed — to access specific cell populations in defined anatomical locations.
- Using the CCF as the common coordinate system for alignment between connectivity data, dissection, and sequencing is correct and is how we do things here.
## What needs tightening
### 1. Define the connectivity prior more precisely
"Convergent input" is a reasonable starting metric, but the Allen projection matrix contains richer information. I would compute **three** connectivity features and test each:
- **Convergent input strength**: sum of normalized projection densities into a region (your proposed metric).
- **Input diversity**: Shannon diversity of the input distribution — a region receiving equally strong input from many sources may differ from one receiving very strong input from few.
- **Reciprocity**: whether a region has strong reciprocal connections with its inputs, which could indicate feedback regulation that buffers stress.
This enriches the analysis and tests whether the relationship is driven by a specific aspect of connectivity architecture.
### 2. Narrow the Cre-line panel for the pilot
Five Cre lines × 6 regions × 3 timepoints × 2 genotypes = 180 libraries is a lot for a pilot. For a 3-month proof of concept, I would start with **two cell types**:
- **Tlx3-Cre** (excitatory glutamatergic neurons — the primary vulnerable population in early AD)
- **Pvalb-IRES-Cre** (fast-spiking interneurons — implicated in network dysfunction)
That cuts libraries to 72, which is feasible and still tests whether the connectivity-stress relationship is cell-type-specific.
### 3. Use INTACT for nuclear isolation
The Cre-dependent Sun1-sfGFP (INTACT) approach is the right method. But specify the crossing strategy explicitly: use the **B6;129S-Intact2** line (Mo et al., *Cell Reports* 2024) crossed with Cre lines to achieve cell-type-specific nuclear tagging. This is a well-validated pipeline in our lab.
### 4. Add MERFISH validation
After the snRNA-seq screen identifies predicted-vulnerable vs. predicted-resilient cell types, a targeted MERFISH panel in the 6 regions would validate that the stress signatures are spatially coherent and not artifacts of dissociation. The Allen Institute has the Vizgen MERSCOPE platform and standardized MERFISH panels aligned to CCF.
### 5. Be explicit about the plaque-distance control
The hardest control is distinguishing "connectivity predicts vulnerability" from "vulnerable regions simply have more plaques nearby." You need to:
- Quantify plaque load in each of the 6 regions at each timepoint by IHC (anti-Aβ 6E10).
- Include plaque density as a covariate in the connectivity-stress regression.
- If the correlation holds after controlling for plaque density, that is the strongest result.
### 6. Consider the 4-month timepoint carefully
At 4 months, 5xFAD mice have substantial amyloid. The interesting biological signal is at **2-3 months** — the pre-plaque and early-plaque window where connectivity-predicted stress would be most informative. I would allocate more mice to 2 and 3 months.
## My requested changes before finalization
- Compute three connectivity metrics, not just convergent input.
- Reduce to Tlx3-Cre + Pvalb-IRES-Cre for the pilot.
- Specify INTACT2 nuclear tagging.
- Include MERFISH validation on a subset.
- Control for plaque density explicitly in the statistical model.
- Emphasize the 2-3 month pre-symptomatic window.
With those changes, this is exactly the kind of experiment I would want to see proposed: it takes our atlas resources and uses them to answer a disease question that cannot be asked without them.