Perturb-seq + Deep Learning: Predicting CRISPRi Guide Identity from Expression

Author: Syed Nurul Hasan, Ph.D. Contact: [email protected] | ORCID: 0000-0002-4564-872X

---

Why this project

Two related CV gaps, closed together: Perturb-seq (CRISPR perturbation with single-cell readout) and deep learning (PyTorch/TensorFlow) applied to single-cell data. This repo trains a PyTorch classifier to predict which CRISPRi guide a cell received from its transcriptional response alone — the same modeling task underlying real Perturb-seq QC (confirming a guide produced a detectable on-target effect) and perturbation-signature discovery.

Pipeline

Adamson et al. 2016 Perturb-seq (K562, CRISPRi, UPR pathway targets)
        │
        ▼
1. preprocess.py  (scanpy)
   ├── Drop unassigned-guide cells
   ├── QC filter, normalize, log1p
   ├── Top 2000 HVGs, scale
   └── Export dense matrix + integer labels
        │
        ▼
2. train_classifier.py  (PyTorch)
   ├── MLP (2000 -> 256 -> 64 -> 8), dropout, class-weighted cross-entropy
   ├── 80/20 stratified train/test split
   └── Training curve + confusion matrix + per-class precision/recall/F1

Dataset

One sample (GSM2406675, run 10X001) from Adamson CH, Norman TM, Mount A, et al. "A Multiplexed Single-Cell CRISPR Screening Platform Enables Systematic Dissection of the Unfolded Protein Response." Cell. 2016;167(7):1867-1882.e21. — a CRISPRi Perturb-seq screen of UPR-pathway transcription factors in K562 cells. Fetched as a harmonized .h5ad from the scPerturb project's Zenodo archive (MD5-verified in scripts/download_data.sh), rather than reprocessed from raw FASTQ — same "downstream of a public processed artifact" pattern as the Cell Ranger disclosure in cite-seq-wnn-trajectory.

5,768 cells x 35,635 genes; after dropping 6 cells with no confidently assigned guide, 8 perturbation groups remain: 7 targeted genes (SPI1, EP300, SNAI1, ZNF326, BHLHE40, CREB1, DDIT3) plus one large group labeled 62(mod)_pBA581 — by far the most abundant single guide identity (1,769/5,762 cells), consistent with a non-targeting or reference-guide population in this design, though I haven't independently confirmed that assignment against the original paper's supplementary tables, so I'm stating it as a reasonable inference rather than a confirmed fact.

Results (this run)

• Train/test split: 4,601 / 1,151 cells (stratified by guide identity).
• Majority-class baseline (always predict 62(mod)_pBA581): 30.8%.
• Model test accuracy after 40 epochs: 44.0% — a clear lift over the majority-class baseline, i.e. the network is picking up real perturbation-specific transcriptional signal, not just learning class frequencies (class-weighted loss specifically guards against the latter).
• The training curve shows classic small-N overfitting — train loss drops to ~0.02 while test loss climbs back up after ~epoch 5 — expected with 2,000 features and a few hundred cells per class; an early-stopping checkpoint around epoch 5 (test_acc ≈ 0.46) would likely generalize slightly better than the final epoch. Reported as observed rather than tuned away.
• Per-class performance is uneven and biologically legible: SPI1 knockdown is the easiest to detect (0.84 precision / 0.91 recall) — consistent with SPI1 being a dominant myeloid master transcription factor whose loss produces a large, distinctive expression shift — while CREB1 and DDIT3 knockdowns are barely separable from background (~0.15 F1), consistent with more subtle or redundant regulatory roles. Full breakdown: results/metrics.json, results/training_and_confusion.png.

Running it

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

bash scripts/download_data.sh
python3 scripts/preprocess.py
python3 scripts/train_classifier.py

Key Features

• Real, published Perturb-seq data (MD5-verified download)
• A genuine multi-class deep learning classification task on single-cell expression, not a toy dataset
• Class-weighted loss to handle real perturbation-group imbalance
• Baseline comparison (majority-class) so the accuracy number means something

References

1. Adamson CH, Norman TM, Mount A, et al. A Multiplexed Single-Cell CRISPR Screening Platform Enables Systematic Dissection of the Unfolded Protein Response. Cell. 2016;167(7):1867-1882.e21.
2. Peidli S, Green TD, Shen C, et al. scPerturb: harmonized single-cell perturbation data. Nature Methods. 2024;21:531-540.

License

MIT — see LICENSE.