CRISPR SignalEngine: Cell Classifier produced in partnership with LifeEdit (aq. NovoNordisk)

Telling edited cells from unedited ones using single-cell DNA expression alone

An ML classifier by DIIG Data for Life Edit to detect edited vs unedited cells using single-cell DNA expression data.

---

Contents

• Overview
• Method
• Datasets
• Repository Layout
• Getting Started
• Tech Stack & Techniques
• Useful Links
• Important Notes
• For Future Reference
• Team
• License

Overview

Gene editing leaves a signature in the transcriptome, but it is not a single marker you can look up — it is a diffuse shift spread across thousands of genes, most of which are noise. This project asks whether that signature is separable from expression data alone, without knowing in advance which genes matter.

The answer the pipeline arrives at is yes, by way of an aggressive feature-selection step. Starting from 39,376 genes across 38 samples, a statistical filter reduces the space to roughly four thousand genes whose behaviour actually differs between edited and untreated populations. Those genes are then clustered, characterised, and handed to a random forest.

Method

1. Normalisation. Raw NCBI counts are log-transformed with log2(x + 1), then z-scored per gene with StandardScaler, so genes with wildly different absolute expression become comparable.

2. Relevance filtering. A gene survives if the edited and untreated populations differ meaningfully in either centre or spread:

|median_edited − median_untreated| ≥ min_median_difference
    OR  std_edited / std_untreated ≥ 1 + min_std_percent_difference / 100
    OR  std_untreated / std_edited ≥ 1 + min_std_percent_difference / 100

Three thresholds were swept, and the gene sets are committed so results are reproducible without rerunning the filter:

Gene set	Median difference	Std difference	Genes retained
relevant_genes_1.2_275.txt	1.2	275%	4,122
relevant_genes_1.6_250.txt	1.6	250%	4,188
relevant_genes_3.0_250.txt	3.0	250%	4,050

Each file is a flat comma-separated list of NCBI GeneIDs.

3. Dimensionality and structure. PCA over the retained expression columns establishes how much variance the surviving genes carry, and an elbow sweep on the resulting space settles on nine clusters.

4. Cluster characterisation. Each cluster is annotated against NCBI gene descriptions and written up in data/elbowClusterData/clusterScience/ — for example, cluster 0 resolves to a "Divergent Pseudogene-Enriched Immunoglobulin and Transcriptomic Regulatory Cluster." This is the step that turns a cluster index into biology.

5. Classification. A random forest is trained over the nine-cluster representation at the 1.6 / 250 threshold. The fitted model is committed at random_forest_gene_classifier_9Clusters_16_250.pkl, with per-cluster outputs in data/elbowClusterData/clusterTestresults/.

Datasets

File	Contents
GSE218462_raw_counts_GRCh38.p13_NCBI.tsv	Raw counts, GEO accession GSE218462, GRCh38.p13
GSE218463_raw_counts_GRCh38.p13_NCBI.tsv	Raw counts, GEO accession GSE218463, GRCh38.p13
Human.GRCh38.p13.annot.tsv	NCBI gene annotations used to join descriptions onto GeneIDs
labeled_gene_data.csv	Labelled matrix produced by the pipeline

Counts are indexed by GeneID with one column per GSM sample. Edited and untreated groups are split by mechanism within the normalised frame rather than by source file.

Repository Layout

LifeEdit-geneclassifier/
├── code/
│   ├── EDA/                    exploratory analysis
│   │   ├── filtering_sj.ipynb        gene relevance filtering
│   │   ├── graphs_mz.ipynb           expression visualisation
│   │   └── pca_amy.ipynb             principal component analysis
│   ├── Elbow_mz/
│   │   ├── elbowTest_filtered/       elbow sweeps at each filter threshold
│   │   └── elbowClassifier/          nine-cluster random forest and model
│   ├── NLP_mz/                 NCBI description parsing and gene-text analysis
│   └── classifier_imt/
│       ├── forest_classifier_amy.ipynb
│       ├── gene_relevance_imt.ipynb
│       └── streamlit_app_imt.py      interactive PCA and filtering dashboard
├── data/
│   ├── 5000 Gene Combinations/ retained gene sets and filtered descriptions
│   ├── elbowClusterData/       cluster outputs, write-ups, and test results
│   ├── nlpClusterData/         NLP-derived cluster notes
│   └── test/                   scratch space for user-generated output
└── LICENSE

Getting Started

The project targets a virtualenv named lifeedit (see .python-version).

python -m venv .venv && source .venv/bin/activate
pip install pandas numpy scikit-learn plotly streamlit jupyter

Run the notebooks in dependency order — EDA and filtering first, then the elbow sweep, then the classifier:

jupyter lab code/

Or launch the interactive dashboard, which performs normalisation, gene filtering and PCA on an uploaded counts file:

streamlit run code/classifier_imt/streamlit_app_imt.py

Tech Stack & Techniques

• Python
• Jupyter
• pandas and NumPy for the expression matrices
• scikit-learn for StandardScaler, PCA, and the random forest
• Streamlit and Plotly for the interactive dashboard
• Statistical gene relevance filtering, elbow-method cluster selection, and NLP over NCBI gene descriptions

Useful Links

• Google Drive

Important Notes

• All User Generated Data will be created in the data/localData folder

For Future Reference

• Figure out a cleaner data pipeline so we don't need 10 different files for similar data

Team

Built by DIIG Data for Life Edit. File suffixes throughout code/ mark authorship.

Contributor	Focus
Mohammad Zoraiz — project lead	Elbow clustering, nine-cluster random forest, NLP description pipeline, visualisation
Islam M. Tayeb	Gene relevance analysis, Streamlit dashboard
Amy Liu	Principal component analysis, forest classifier
Soham Jinsi	Gene filtering and exploratory analysis

License

Released under the GNU General Public License v3.0. See LICENSE for the full text.