Skip to content
VeriHelix by NexagenLabs

Colab Notebook Starter

Pick an analysis and download a commented starter notebook.

Runs in your browser. Nothing you enter is uploaded.

Options

19 cells, ready to run

Open Colab
  1. text
    Single-cell RNA-seq: QC to clusters A complete first-pass analysis with every threshold visible and justified. Before you run this 1. Open in Google Colab and choose a GPU runtime if the notebook asks for one. 2. Replace every `# EDIT:` line with your own paths and parameters. 3. Run cells in order. Each section ends with a check that tells you whether the step worked. Generated by VeriHelix (verihelix.nexagenlabs.com). Background and interpretation: *AI for Biotech* (Tripathy, 2026), Chapter 12.
  2. In [1]
    !pip -q install scanpy==1.10.* leidenalg scrublet celltypist
    import scanpy as sc, numpy as np, pandas as pd, matplotlib.pyplot as plt
    sc.settings.verbosity = 1
    sc.set_figure_params(dpi=100, facecolor="white")
    print("scanpy", sc.__version__)
  3. text
    1. Load the data Edit `DATA_DIR` to point at your own data.
  4. In [3]
    DATA_DIR = "/content/filtered_feature_bc_matrix"  # EDIT: your path
    adata = sc.read_10x_mtx(DATA_DIR, var_names="gene_symbols", cache=True)
    adata.var_names_make_unique()
    print(adata)
    assert adata.n_obs > 0 and adata.n_vars > 0, "Nothing loaded: check DATA_DIR"
  5. text
    2. Quality control Starting thresholds for pbmc: at least 200 genes per cell, at most 5000, and under 10% mitochondrial reads. These are defaults, not rules. Look at the distributions below and move them to fit your data.
  6. In [5]
    adata.var["mt"] = adata.var_names.str.startswith(("MT-", "mt-"))
    sc.pp.calculate_qc_metrics(adata, qc_vars=["mt"], inplace=True, log1p=False)
    sc.pl.violin(adata, ["n_genes_by_counts", "total_counts", "pct_counts_mt"], jitter=0.4, multi_panel=True)
    
    MIN_GENES, MAX_GENES, MAX_MT = 200, 5000, 10  # EDIT after looking at the plots
    before = adata.n_obs
    adata = adata[(adata.obs.n_genes_by_counts >= MIN_GENES) & (adata.obs.n_genes_by_counts <= MAX_GENES) & (adata.obs.pct_counts_mt < MAX_MT)].copy()
    sc.pp.filter_genes(adata, min_cells=3)
    print(f"Kept {adata.n_obs} of {before} cells ({adata.n_obs/before:.1%})")
    assert adata.n_obs > 100, "Fewer than 100 cells left: your thresholds are probably too strict"
  7. text
    3. Doublets Two cells captured in one droplet look like a hybrid cell type and create fake "intermediate" clusters.
  8. In [7]
    sc.pp.scrublet(adata)
    print(adata.obs.predicted_doublet.value_counts())
    adata = adata[~adata.obs.predicted_doublet].copy()
    print(f"{adata.n_obs} cells after removing doublets")
  9. text
    4. Normalise and find variable genes The raw counts are kept in `adata.layers["counts"]` because differential expression needs them later.
  10. In [9]
    adata.layers["counts"] = adata.X.copy()
    sc.pp.normalize_total(adata, target_sum=1e4)
    sc.pp.log1p(adata)
    adata.raw = adata
    sc.pp.highly_variable_genes(adata, n_top_genes=2000, batch_key=None)
    sc.pl.highly_variable_genes(adata)
    adata = adata[:, adata.var.highly_variable].copy()
    sc.pp.scale(adata, max_value=10)
  11. text
    5. Dimensionality reduction Choose the number of PCs from the elbow in the variance plot.
  12. In [11]
    sc.tl.pca(adata, svd_solver="arpack", n_comps=50)
    sc.pl.pca_variance_ratio(adata, n_pcs=50, log=True)
    N_PCS = 30  # EDIT: read the elbow off the plot above
    sc.pp.neighbors(adata, n_neighbors=15, n_pcs=N_PCS)
    sc.tl.umap(adata)
  13. text
    6. Clustering Resolution controls how many clusters you get. There is no correct value: try several and keep the one whose clusters correspond to markers you can defend.
  14. In [13]
    for res in [0.3, 0.5, 1.0]:
        sc.tl.leiden(adata, resolution=res, key_added=f"leiden_{res}", flavor="igraph", n_iterations=2)
        print(res, adata.obs[f"leiden_{res}"].nunique(), "clusters")
    sc.pl.umap(adata, color=["leiden_0.3", "leiden_0.5", "leiden_1.0"], wspace=0.3)
    adata.obs["leiden"] = adata.obs["leiden_0.5"]  # EDIT: pick the resolution you trust
  15. In [14]
    sc.tl.rank_genes_groups(adata, "leiden", method="wilcoxon", layer="counts", use_raw=False)
    sc.pl.rank_genes_groups_dotplot(adata, n_genes=5)
    pd.DataFrame(adata.uns["rank_genes_groups"]["names"]).head(10)
  16. text
    7. Automatic annotation CellTypist gives a first pass. Always check its labels against the marker genes above: automatic annotation is a hypothesis.
  17. In [16]
    import celltypist
    from celltypist import models
    models.download_models(model=["Immune_All_Low.pkl"])  # EDIT: pick a model for your tissue
    pred = celltypist.annotate(adata.raw.to_adata(), model="Immune_All_Low.pkl", majority_voting=True)
    adata.obs["celltypist"] = pred.predicted_labels.majority_voting.values
    sc.pl.umap(adata, color=["celltypist"], legend_loc="on data", legend_fontsize=7)
  18. text
    Save and record what you did Writing versions and thresholds into the file is what makes the analysis reproducible six months from now.
  19. In [18]
    adata.uns["analysis_params"] = dict(min_genes=MIN_GENES, max_genes=MAX_GENES, max_mt=MAX_MT, n_pcs=N_PCS, scanpy=sc.__version__)
    adata.write("analysis_result.h5ad")
    print("Saved. Cells:", adata.n_obs, "| Clusters:", adata.obs.leiden.nunique())
After downloading, open Colab, choose File → Upload notebook, and edit every line marked # EDIT.

Go deeper: The full workflow this notebook sketches: single-cell omics and foundation models — AI for Biotech, Ch. 12 (Tripathy, 2026) Chapter 12: Single-Cell Omics and Foundation Models

These are starting points with the common mistakes designed out, not finished pipelines. Library APIs change; if a cell errors, check that package’s current documentation. AI for Biotech explains what each step does and how to judge the output.

What makes these different from a tutorial

Each notebook has the decisions written into it: why a threshold is set where it is, what to look at before changing it, and a check after each step that tells you whether it worked. The mistakes that most often spoil these analyses are designed out — fitting preprocessing before splitting, normalising counts twice, using the wrong background for enrichment, trusting a cluster resolution because it was the default.

Every line you need to change is marked # EDIT.