Virtual Rat

Virtual Rat project is a computaional neuroscience project conducted by Xiangci Li as a research assistant at Erlich lab under the guidance of New York University Shanghai Professor Jeffrey Erlich. We showed that the task switch cost of rat's ProAnti orienting task can be modeled by an Elman Recurrent Neural Network. The experimental data are from Dr. Carlos Brody's lab at Princeton.

There is a video available for this project!

For any question, please contact

Xiangci Li (xiangci.li@nyu.edu)

Jeffrey Erlich (jerlich@nyu.edu)

Briefly about the implementation

From the perspective of implementation, Virtual Rat project consists of two parts: Virtual Rat (RNN model) and data analysis of rats' behavioral and electrophysiology data.

How to reproduce all results:

Environment setup:

Python:

1. Visit https://github.com/dmlc/minpy to setup Minpy and MXNet.
2. Other requirements: Python 2.7, Numpy, Matplotlib, Scikit-learn, Scipy

MATLAB:

MATLAB 2017a or later.

Note:

The codes for the final version of this project are all under publication folder.

figures folder saves figures produced. mats folder saves .mat files and pkls folder saves .pkl files.

Procedures:

Virtual Rat part:

1. Experiments for the most basic RNNs:

1. Run sbatch trainingTime.sh on HPC to evoke TrainingTime.py. All weights after certain training epochs will be saved.
2. Run sbatch testTrainingTime.sh on HPC to evoke TestTrainingTime.py. Similarly run sbatch testTrainingTimeFine.sh to evoke TestTrainingTimeFine.py. The weights saved will be loaded to test the performance of the models and the results will be saved in the folder TrainingTime.
3. Run jupyter notebook TrainingTime.ipynb. This notebook finds out best epochs to plot later figures.
4. Run jupyter notebook VirtualRat.ipynb and DiluteActivation.ipynbto obtain figures.
5. Run jupyter notebook PETH.ipynb to choose sample PETH.

2. Experiments for varying block sizes:

1. Run sbatch trainBlock.sh on HPC to evoke BlockTime.py. All weights after certain training epochs will be saved.
2. Run sbatch testBlockTime.sh on HPC to evoke TestBlockTime.py. The weights saved will be loaded to test the performance of the models and the results will be saved in the folder BlockTime.
3. Run jupyter notebook Block.ipynb.

3. Experiments for varying pro to anti ratio:

1. Run sbatch RatioTime.sh on HPC to evoke TrainRatioTime.py. All weights after certain training epochs will be saved.
2. Run sbatch testTrainingTimeRatio.sh on HPC to evoke TestTrainingTimeRatio.py. The weights saved will be loaded to test the performance of the models and the results will be saved in the folder RatioTime.
3. Run jupyter notebook TrainingTimeRatio.ipynb to check the performace over time.
4. Run jupyter notebook Ratio.ipynb.

Data analysis part (MATLAB & Python):

1. Run convertBehavior.m to extract necessary information from the original huge data file for running Python scripts.
2. Run countSpikes.m to produce spike counts.
3. Run realRatEphysTrainClfs.py and realRatEphysTrainClfsTarget.py
4. Run ratSGDrule.m, ratSGDTarget.m, RNN_SGD_rule.m and RNN_SGD_target.m.
5. Run jupyter notebook realRatEphys.ipynb.
6. Run jupyter notebook realRatEphysTarget.ipynb.
7. Run jupyter notebook RNNEphys.ipynb.
8. Run jupyter notebook realRatBehavior.ipynb.

Documentation to files, funtions and data structures

Data analysis part

convertBehavior.m

• Collect pro anti and other data from original spike data

  • Saved in SessionInfo.mat, a cell type structure.
  • Each cell has double of (N_of_trials, 4)
  • Each trial has (pro, target_right, switch, hit)

• Collect brain area data

  • Saved in BrainArea.mat, a cell type structure.

  • Each cell has double of (1, N_of_cells)

  • Index: brain area ​ 0: left mPFC ​ 1: right mPFC ​ 2: left SC ​ 3: right SC ​ 4: left FOF ​ 5: right FOF

• Collect rat index

  • Saved in RatIndexPerSession.mat, (N_session, 1) double.

• Collect cellid per session

  • Saved in CellIndexPerSession.mat, a cell type structure.
  • Each cell has double of (1, N_of_cells), containing cell indices (not the raw cell ID, but still sufficient)

• Collect Duan's session index and Erlich's session indices.

countSpikes.m

• Count spikes from Duan's ephys data and Erlich's ephys data using different time periods and time steps.
  • Duan's spike count includes delay step, which matches the RNN model.
  • Erlich's does not include delay step.
• Saved in SpikeCountsPerSessionDuan.mat and SpikeCountsPerSessionErlich.mat, cell type structure.
• Each cell is a double with shape (5 steps, num_cells recorded in this session, num_trials)

singleCellRuleEncoding.m

• Separately calculate for Erlich's and Duan's data.
• Save as Duan_single_AUC_p.mat, an array with (N_cells,10)
• Each column corresponds to 'ITI_auc','ITI_p','rule_auc','rule_p','delay_auc','delay_p','target_auc','target_p','choice_auc','choice_p'

realRatEphysTrainClfs.py

• This script trains (overfit) logistic regression classifier to predict each trials' rule encoding score and save them in to an array which will be converted to MATLAB table for fitting GLME.
• Saves results computed in experimentor_SGD_table_brainArea.mat
  • 21 categories: session_index, pro, right, switch, hit, rat_index, score0, score1, score2, score3, score4,accuracy0, accuracy1, accuracy2, accuracy3, accuracy4, encoding0, encoding1, encoding2, encoding3, encoding4