Skip to content

INDEX.md

Latest commit

 

History

History
262 lines (193 loc) · 7.43 KB

INDEX.md

File metadata and controls

262 lines (193 loc) · 7.43 KB

Multi-Condition LOOCV Pipeline - Complete Index

📖 Documentation

Start here based on your needs:

For Quick Copy-Paste Commands

→ QUICK_START.md

  • One-line commands for each step
  • "I want to..." lookup table
  • Troubleshooting guide

For Complete Reference

→ PIPELINE_COMMANDS.md

  • Detailed step-by-step explanation
  • Parameter descriptions
  • Expected performance metrics
  • File structure diagram

🔧 Scripts

1. Regenerate Baseline Weights

File: scripts/regenerate_baseline.sh

Extracts baseline weights from the control condition for use in plot comparison.

bash scripts/regenerate_baseline.sh

Output: /tmp/baseline_weights_intersection.csv

2. Run LOOCV with Plots

File: scripts/run_multicond_loocv.py (existing, enhanced)

Runs leave-one-odor-out cross-validation with visualization.

python scripts/run_multicond_loocv.py \
  --csv /tmp/reaction_rates_no_citral.csv \
  --control-row opto_AIR \
  --conditions opto_AIR,opto_EB,opto_hex,opto_benz_1,opto_3-oct \
  --model elasticnet \
  --feature-set intersection \
  --activation-threshold 0.05 \
  --l1-ratio 0.3 \
  --plot --plot-top-n 13 \
  --plot-baseline-weights /tmp/baseline_weights_intersection.csv \
  --plot-comparison \
  --outdir out/multicond_loocv_best

Outputs:

  • out/multicond_loocv_best/plots/ - 7 comparison plots
  • out/multicond_loocv_best/weights_mean_*.csv - Averaged LOOCV weights
  • out/multicond_loocv_best/conditions_overview.csv - Summary metrics

3. Make Predictions from Averaged Weights

File: scripts/predict_with_avg_weights.py

Uses the mean LOOCV weights to predict ΔPER and compare to true values.

python scripts/predict_with_avg_weights.py \
  --loocv-dir out/multicond_loocv_best \
  --csv /tmp/reaction_rates_no_citral.csv \
  --control-row opto_AIR \
  --conditions opto_EB,opto_hex,opto_benz_1,opto_3-oct \
  --feature-set intersection \
  --activation-threshold 0.05 \
  --outdir out/prediction_plots_best

Outputs:

  • out/prediction_plots_best/predictions_vs_true.png - Per-condition bar plots
  • out/prediction_plots_best/predictions_scatter.png - Scatter plot (overall R²)
  • out/prediction_plots_best/weights_comparison.png - Weight comparison
  • out/prediction_plots_best/predictions.csv - Detailed predictions (24 rows)

📊 Output Files

LOOCV Results

out/multicond_loocv_best/
├── plots/                           ← 7 PNG files
│   ├── weights_deltaper_3-octonol.png
│   ├── weights_deltaper_apple_cider_vinegar.png
│   ├── weights_deltaper_benzaldehyde.png
│   ├── weights_deltaper_ethyl_butyrate.png
│   ├── weights_deltaper_hexanol.png
│   ├── weights_deltaper_linalool.png
│   └── weights_all_conditions.png
├── predictions_opto_*.csv           ← Per-condition LOOCV predictions
├── weights_mean_opto_*.csv          ← Averaged weights (13 receptors)
├── weights_folds_opto_*.csv         ← Fold-specific weights
├── conditions_overview.csv          ← Summary: R², MSE, n_features
└── summary_opto_*.json              ← Detailed metrics per condition

Prediction Results

out/prediction_plots_best/
├── predictions_vs_true.png          ← Bar plots (4 conditions)
├── predictions_scatter.png          ← Scatter plot with R²
├── weights_comparison.png           ← Bar chart (13 receptors)
└── predictions.csv                  ← 24 rows (6 odors × 4 conditions)

Baseline Weights

/tmp/baseline_weights_intersection.csv
├── feature: Or19a, Or22a, ..., ac3_noOr35a (13 receptors)
└── baseline_w: mean weights from opto_AIR control

📈 Performance Summary

Configuration: 6 odors (no citral), 13 receptors (intersection), ElasticNet

Condition R² MSE Folds
opto_EB 0.40 0.009 6
opto_hex 0.44 0.032 6
opto_benz_1 0.30 0.037 6
opto_3-oct 0.27 0.004 6
Overall 0.35 - -

🔑 Key Features Used (13 Receptors)

Or19a, Or22a, Or2a, Or35a, Or47b, Or67b, Or7a, Or83c,
Or85b, Or98a, ac1, ac2, ac3_noOr35a

These are the only receptors active (threshold > 0.05) across all 6 odors.

🎯 How to Use This Pipeline

Option A: Run Everything at Once

# 1. Regenerate baseline
bash scripts/regenerate_baseline.sh

# 2. Run LOOCV (copy from QUICK_START.md)
python scripts/run_multicond_loocv.py ...

# 3. Make predictions
python scripts/predict_with_avg_weights.py ...

Option B: Run Step-by-Step

  1. Read QUICK_START.md for copy-paste commands
  2. Copy Step 1 command and run
  3. Copy Step 2 command and run
  4. Copy Step 3 command and run
  5. View results in output directories

Option C: Reference Only

Use this as a reference while working on your own variations.

📝 Example Output Interpretation

LOOCV Plot (weights_deltaper_3-octonol.png)

Top Subplot (Weights):

  • Purple bars = Baseline weights (opto_AIR control)
  • Blue/Orange/Green/Red bars = Delta weights for each trained condition
  • Height = Receptor contribution to prediction

Bottom Subplot (ΔPER):

  • Mean-centered behavioral response (what we're trying to predict)
  • Shows which condition activates neuron the most

Prediction Plot (predictions_scatter.png)

  • X-axis = True ΔPER (centered)
  • Y-axis = Predicted ΔPER (centered)
  • Diagonal line = Perfect prediction
  • R² value = Goodness of fit (higher is better, max = 1.0)

Predictions CSV

Each row shows:

condition,odor,true_delta_per,predicted_delta_per,true_centered,predicted_centered
opto_EB,3-Octonol,0.1097,0.0894,0.0052,0.0189

This means: For opto_EB + 3-Octonol odor, we predicted 0.0894 when truth was 0.1097

🐛 Troubleshooting

Issue Solution
Baseline weights all zero Make sure you ran bash scripts/regenerate_baseline.sh first
File not found error Check that /tmp/reaction_rates_no_citral.csv exists
Wrong number of receptors Verify --feature-set intersection --activation-threshold 0.05
Plots missing baseline bars Pass --plot-baseline-weights /tmp/baseline_weights_intersection.csv

See QUICK_START.md for more details.

📚 Related Files

Input Data

  • /tmp/reaction_rates_no_citral.csv - PER data (6 odors)
  • data/mappings/door_to_flywire_mapping.csv - Receptor mapping

Source Code

  • src/door_toolkit/multicond_loocv.py - LOOCV implementation
  • src/door_toolkit/glomerulus_features.py - Feature matrix builder
  • src/door_toolkit/encoder.py - DoOR encoder

🔄 Next Steps

  1. Choose which documentation to read:

  2. Run baseline regeneration:

    bash scripts/regenerate_baseline.sh

  3. Run LOOCV + predictions (copy command from docs)

  4. View plots and results in output directories

📧 Questions?

Last Updated: 2026-02-18 Configuration: 6 odors (no citral), 13 receptors (intersection), ElasticNet Status: ✓ Complete and ready to use