🧠 Prediction Error Signaling in VIP and SST Neurons

An Event-Locked Calcium Signal Analysis of Stimulus Omission in Familiar and Novel Image Sequences

Neuromatch Academy Computational Neuroscience Project (NMA 2025)

---

🔍 Project Overview

This repository contains the code and analyses from our NMA2025 project, where we investigated prediction-error signaling in inhibitory neurons of mouse visual cortex, using the Allen Brain Observatory Visual Behavior 2-Photon (2P) dataset.

We focused on VIP and SST neurons in the primary visual cortex (V1), exploring their response to:

• ✅ Presented vs. Omitted image flashes
• 🌄 Familiar vs. Novel image exposure
• 🧠 Trial-by-trial calcium signal analysis using event-locked measures
• 🤖 Cell-type classification via decoding models

---

❓ Research Question

How do VIP and SST neurons encode prediction-error signals during stimulus omissions, and how does this differ between familiar and novel visual contexts?

---

🧬 Methodology

Our analysis pipeline was designed to investigate prediction-error signaling in VIP and SST neurons using calcium imaging data from the Allen Visual Behavior 2-Photon (2P) dataset. The pipeline consists of the following key stages:

1. 🧹 Preprocessing

• Baseline Correction: We first applied baseline correction to calcium fluorescence traces to normalize across trials.

• Averaging per Cell: We then averaged the event-locked traces across trials for each unique cell ID, under each combination of:

  • Cell type (VIP or SST)
  • Omission status (True or False)
  • Exposure (Familiar or Novel)

• Result: This resulted in a compact DataFrame with 640 rows and the following fields:

  • mouse_id, cell_id, cell_type, omission, exposure, mean_trace

Each mean_trace was a vector representing the average event-locked calcium signal for that condition.

2. ⚙️ Feature Extraction

From the mean_trace vectors, we computed biologically relevant signal features such as:

• AUC (Area Under Curve)
• Peak Amplitude
• Latency to Peak
• Slope & Timing Windows

These features formed the input space for statistical analysis and supervised classification.

3. 🧪 Statistical Analysis

To test the effects of omission, exposure, and cell type on neural activity:

• We used three-way ANOVA to assess:
  • Main effects of cell_type, exposure, and omission
  • All two-way and three-way interaction effects
• We followed this with post-hoc comparisons (Tukey HSD and Hedges' g) to estimate effect sizes and pairwise differences

This revealed strong prediction-error dynamics in VIP neurons, particularly under familiar + omitted conditions.

4. 🔍 Spike Inference with CASCADE

In parallel, we also analyzed raw calcium traces using CASCADE, a state-of-the-art deep learning model for supervised spike inference (Berens et al., 2021):

• CASCADE output: spike_prob — estimated spike probability per time bin

• We extracted spike-based features (e.g., average spike rate in different bins) and repeated the full factorial ANOVA

• These results corroborated our findings using traditional calcium features

• 

• 

5. 🤖 Machine Learning Pipeline

To determine if neuron types and stimulus conditions could be decoded from event-locked activity:

• We trained several classifiers on the extracted features:
  • Logistic Regression
  • Random Forest
  • Gradient Boosting
  • XGBoost
  • SVM (RBF)
  • k-Nearest Neighbors

🧪 Leave-One-Mouse-Out Cross-Validation (LOMO)

To ensure generalization across individual animals, we used Leave-One-Mouse-Out (LOMO) cross-validation:

• In each fold, data from one mouse was used as the test set
• The model was trained on the remaining 12 mice
• This avoids overfitting and ensures biological generalizability across individuals

The best-performing model (Gradient Boosting) achieved:

• LOMO Accuracy: 94.5%
• Test Accuracy: 97.7%
• F1 Score: 0.976
• ROC-AUC: 0.997

---

📊 Classifier Performance

Model	LOMO Accuracy	Test Accuracy	F1 Score	ROC-AUC
Gradient Boosting	0.945	0.977	0.976	0.997
XGBoost	0.947	0.969	0.969	0.996
Random Forest	0.931	0.969	0.968	0.998
Logistic Regression	0.902	0.953	0.951	0.984
SVM (RBF)	0.880	0.914	0.913	0.977
KNN (k=7)	0.873	0.898	0.896	0.958

• Top features: Early trial (0–500 ms) firing rate and slope
• Gradient Boosting showed strongest generalization across all 13 mice

---

🗂️ Project Structure

├── EDA.ipynb # Exploratory data analysis

├── event-locked-calcium-signal_preprocessing&feature extraction_per-cell.ipynb # Event-locked signal extraction per-cell(preprocessing) and Feature extraction(processing)

├── modeling.ipynb # Classifier training & evaluation

├── event-locked-calcium-signal-MVPA # multivariate pattern analysis on unprocessed dataset(Time Resolved Decoding)

├── event-locked-calcium-signal-Hypothesis_testing_per-cell # testing hypothesis with ANOVA and post-hoc

├── event-locked-calcium-signal-cascade model.ipynb # CASCADE-based spike analysis

├── figures/ # Plots and visualizations

├── README.md # Project overview

---

📁 Dataset

• Source: Allen Institute Visual Behavior 2P Dataset
• Download: Figshare - Allen SDK Visual Behavior 2P
• Includes: Cre lines (Vip, Sst), image familiarity tags, 2P calcium recordings

---

Team

Team Name: Feisty Bes / VIP Gladiators
Contributors:

• Hossein Damavandi
• Mohammad Saeed Soleimani
• Sotoode Aliyari
• Elnaz Abbaszadeh
• Niyayesh Yousefi
• Fatemeh Gholamzadeh

---

📚 References

• Piet et al. (2022) Nature Neuroscience
• Najafi et al. (2023) Nature Neuroscience
• Garrett et al. (2023) eLife
• Homann et al. (2023) Nature Neuroscience

---

📄 License

This project is released under the MIT License.

---

🔗 GitHub

➡️ Project Repository