ronniebasak/ach-cortical-circuits

---
license: mit
task_categories:
- tabular-regression
language:
- en
tags:
- neuroscience
- computational-neuroscience
- acetylcholine
- neuromodulation
- sim-to-real
- jaxley
- allen-brain-observatory
- neuropixels
- biophysical-simulation
size_categories:
- 10K<n<100K
---

# ACh Cortical Circuits: Cross-Cortical Transfer Dataset

15,000 simulated cortical microcircuits with acetylcholine (ACh) modulation + 2,678 real mouse V1 recording epochs from the Allen Brain Observatory.

## Dataset Description

This dataset accompanies the paper **"Cross-Cortical Transfer of Cholinergic Modulation Principles via Biophysical Simulation Surrogates"** (CAISc 2026).

### Simulated Circuits (15,000 total)

Three mechanistic variants of ACh modulation, each with 5,000 circuits × 11 ACh levels = 55,000 simulations per variant:

| Variant | ACh Mechanism | File |
|---------|--------------|------|
| **v14a** | Synaptic suppression only (E2 dose-response) | `v14a_circuits.tar.gz` |
| **v14b** | Depolarization only (OU mean shift) | `v14b_circuits.tar.gz` |
| **v14c** | Both mechanisms combined | `v14c_circuits.tar.gz` |

**Parameters source:** Ramaswamy, Colangelo & Markram (2018), *Frontiers in Neural Circuits* — rat somatosensory cortex (S1).

**Simulator:** [Jaxley](https://github.com/jaxleyverse/jaxley) (JAX-based, GPU-accelerated)

Each circuit JSON contains:
- Circuit parameters (n_neurons, connectivity, synaptic conductances, OU noise params)
- ACh level (0.0 to 1.0 in 0.1 steps)
- 11 population statistics (firing rate, CV ISI, Fano factor, synchrony index, pairwise correlation, peak frequency, spectral power, etc.)

### Allen Brain Observatory Epochs (2,678 total)

Real mouse V1 (area VISp) neural recordings from the Allen Brain Observatory Visual Coding Neuropixels dataset:
- 2,325 running epochs + 353 stationary epochs
- 44 sessions, 16 with both behavioral states
- Quality filtered (presence_ratio > 0.95, ISI violations < 0.5)
- Same 11 population statistics as simulated data

File: `allen_epochs.tar.gz`

### Model Checkpoints

12 trained surrogate models (3 architectures × 3 ACh variants + 3 baselines):
- CircuitTransformer (222K params)
- CircuitMLP (5.6K params)
- XGBoost

File: `checkpoints.tar.gz`

## Key Results

| Statistic | Model S (syn) | Model D (depol) | Model SD (both) |
|-----------|:---:|:---:|:---:|
| Pairwise correlation | **75.0%** | 25.0% | 25.0% |
| Mean firing rate | 12.5% | **87.5%** | **87.5%** |
| Spectral power | **68.8%** | **68.8%** | **68.8%** |

**Key finding:** Synaptic suppression captures correlation structure; depolarization captures rate modulation. The mechanisms are dissociable in sim-to-real transfer.

## Citation

```bibtex
@inproceedings{neuropeek2026crosstransfer,
title={Cross-Cortical Transfer of Cholinergic Modulation Principles via Biophysical Simulation Surrogates},
author={NeuroPeek AI and Basak, Sohan},
booktitle={1st Conference for AI Scientists (CAISc 2026)},
year={2026}
}
```

## License

MIT

## Acknowledgments

- Simulation parameters from [Ramaswamy et al. 2018](https://www.frontiersin.org/articles/10.3389/fncir.2018.00077/full)
- Real data from [Allen Brain Observatory](https://portal.brain-map.org/)
- Simulation engine: [Jaxley](https://github.com/jaxleyverse/jaxley) (Deistler et al., Nature Methods 2025)
- Compute: [Modal](https://modal.com/) cloud GPUs (~$230 total)
- Built with NeuroPeek AI Agents