Popv Cell Annotation

popV Multi-Method Cell Type Transfer

Overview

popV (Population Voting for single-cell annotation) annotates a query scRNA-seq dataset by running 10+ independent classification algorithms against a labeled reference atlas and aggregating results via majority voting. Each method produces its own label; the final popv_prediction is the consensus across all methods, and the popv_agreement score quantifies how many methods agree. This ensemble strategy is robust to individual method failures on unusual datasets and provides a principled uncertainty estimate: low agreement highlights novel cell states or annotation gaps.

When to Use

• Annotating a query dataset by transferring labels from a well-curated reference atlas when you want a consensus rather than a single model's judgment
• Identifying novel or ambiguous cell states as cells where methods disagree (low popv_agreement score)
• Benchmarking annotation reliability by comparing per-method labels to detect systematic disagreements
• Annotating large atlas datasets (>100k cells) where batch effects between reference and query are substantial
• Producing annotation for downstream analyses that require high-confidence labels (clinical data, regulatory submissions)
• Use CellTypist (celltypist-cell-annotation) instead when speed matters and a pre-trained model matches your tissue; popV is slower because it trains multiple models on your reference
• Use scANVI (scvi-tools-single-cell) instead when you need a single probabilistic deep generative model with formal uncertainty quantification and do not require the ensemble

Prerequisites

• Python packages: popv>=0.6, scanpy>=1.9, anndata, scvi-tools>=1.0, harmonypy, bbknn, celltypist
• Data requirements: Two AnnData objects — a labeled reference (adata_ref) with cell type labels in obs, and an unlabeled query (adata_query). Both must be from the same species and have overlapping gene sets. Raw counts in adata.X (popV applies its own normalization internally)
• Environment: Python 3.9+; GPU recommended for scVI/SCANVI methods (falls back to CPU); 32 GB RAM recommended for >200k reference cells

pip install popv scvi-tools harmonypy bbknn celltypist

Quick Start

Minimal pipeline from labeled reference and unlabeled query to annotated result:

import popv
import scanpy as sc

# Load reference (labeled) and query (unlabeled) AnnData objects
adata_ref = sc.read_h5ad("reference_atlas.h5ad")  # adata_ref.obs["cell_type"] exists
adata_query = sc.read_h5ad("query_dataset.h5ad")

# Prepare combined object with popV preprocessing
adata = popv.preprocessing.Process_Query(
    adata_ref,
    adata_query,
    ref_labels_key="cell_type",
    ref_batch_key="batch",
    query_batch_key="batch",
    unknown_celltype_label="unknown",
    save_path_trained_models="./popv_models/",
    n_epochs_unsupervised=50,
)

# Run all annotation methods
popv.annotation.annotate_data(adata)

# Inspect consensus results for query cells
query_mask = adata.obs["_dataset"] == "query"
print(adata[query_mask].obs[["popv_prediction", "popv_agreement"]].head(10))

Core API

Module 1: Reference and Query Data Setup

Both AnnData objects must share a gene space and have required metadata columns. popV will subset to the intersection of genes automatically.

import anndata as ad
import scanpy as sc
import numpy as np

# Reference: must have cell type labels and (optionally) batch metadata
adata_ref = sc.read_h5ad("reference_atlas.h5ad")
print(f"Reference: {adata_ref.n_obs} cells x {adata_ref.n_vars} genes")
print(f"Cell types: {adata_ref.obs['cell_type'].nunique()} unique labels")
print(f"Reference cell type counts:\n{adata_ref.obs['cell_type'].value_counts().head(10)}")

# Query: no labels required; batch metadata optional
adata_query = sc.read_h5ad("query_dataset.h5ad")
print(f"\nQuery: {adata_query.n_obs} cells x {adata_query.n_vars} genes")

# Check gene overlap (popV will handle subsetting but >70% overlap is recommended)
shared_genes = adata_ref.var_names.intersection(adata_query.var_names)
pct_shared = len(shared_genes) / adata_ref.n_vars
print(f"\nShared genes: {len(shared_genes)} ({pct_shared:.1%} of reference genes)")
if pct_shared < 0.5:
    print("WARNING: <50% gene overlap — annotation quality may be reduced")

# Verify required fields before popV setup
assert "cell_type" in adata_ref.obs.columns, "Reference needs cell type labels"

# Add batch column if absent (popV requires it even for single-batch data)
if "batch" not in adata_ref.obs.columns:
    adata_ref.obs["batch"] = "ref_batch"
if "batch" not in adata_query.obs.columns:
    adata_query.obs["batch"] = "query_batch"

print("Reference obs columns:", adata_ref.obs.columns.tolist())
print("Query obs columns:    ", adata_query.obs.columns.tolist())

Module 2: POPV Object Creation (Process_Query)

Process_Query combines reference and query, normalizes counts, selects HVGs, and prepares the joint embedding needed by all annotation methods.

import popv

# Create processed combined AnnData
adata = popv.preprocessing.Process_Query(
    adata_ref,
    adata_query,
    ref_labels_key="cell_type",      # obs column with reference labels
    ref_batch_key="batch",           # obs column with reference batch info
    query_batch_key="batch",         # obs column with query batch info
    unknown_celltype_label="unknown",# label to use for query cells before annotation
    save_path_trained_models="./popv_models/",  # directory for scVI/SCANVI model checkpoints
    n_epochs_unsupervised=50,        # scVI training epochs (increase to 100–200 for large datasets)
    n_epochs_semisupervised=20,      # scANVI fine-tuning epochs
    use_gpu=True,                    # GPU for scVI/SCANVI (falls back to CPU if unavailable)
    hvg=4000,                        # number of highly variable genes to use
)

print(f"Combined object: {adata.n_obs} cells x {adata.n_vars} genes")
print(f"Dataset labels: {adata.obs['_dataset'].value_counts().to_dict()}")
# Expected: {'ref': N_ref, 'query': N_query}

Module 3: Running the Method Ensemble

annotate_data runs all selected methods sequentially and adds per-method label columns plus the consensus to adata.obs.

import popv

# Run annotation with default set of methods
popv.annotation.annotate_data(
    adata,
    methods=[
        "knn_harmony",    # KNN on Harmony-corrected embedding
        "knn_bbknn",      # KNN on BBKNN cross-batch graph
        "knn_scvi",       # KNN on scVI latent space
        "scanvi_popv",    # Semi-supervised scANVI label transfer
        "celltypist_popv",# CellTypist logistic regression
        "rf",             # Random Forest on HVG expression
        "xgboost",        # XGBoost classifier
        "svm",            # Support Vector Machine
        "onclass",        # ONCLASS (ontology-guided)
    ],
)

# Inspect per-method result columns (all end in "_popv")
query_mask = adata.obs["_dataset"] == "query"
popv_cols = adata.obs.filter(like="_popv").columns.tolist()
print(f"Per-method columns: {popv_cols}")
print(adata[query_mask].obs[popv_cols + ["popv_prediction", "popv_agreement"]].head(10))

Module 4: Consensus Results and Agreement Scoring

popv_prediction is the majority-vote consensus; popv_agreement is the fraction of methods that agreed on the winning label.

import pandas as pd

query_mask = adata.obs["_dataset"] == "query"
query_obs = adata[query_mask].obs.copy()

# Consensus label distribution
print("Consensus cell type distribution:")
print(query_obs["popv_prediction"].value_counts().head(15))

# Agreement score statistics
print(f"\npopv_agreement statistics:")
print(query_obs["popv_agreement"].describe())
# agreement = 1.0 → all methods agree; agreement = 0.2 → only 2/10 methods agree

# Cells with high confidence (>80% method agreement)
high_conf = query_obs["popv_agreement"] >= 0.8
print(f"\nHigh-confidence cells (agreement >= 0.8): {high_conf.sum()} ({high_conf.mean():.1%})")

# Cells with low confidence — candidate novel states or annotation gaps
low_conf = query_obs["popv_agreement"] < 0.5
print(f"Low-confidence cells  (agreement <  0.5): {low_conf.sum()} ({low_conf.mean():.1%})")

Module 5: Visualization

popV provides built-in UMAP and heatmap visualization of per-method agreement and consensus labels.

import popv
import scanpy as sc
import matplotlib.pyplot as plt

# Compute UMAP on the joint reference+query embedding (if not already present)
if "X_umap" not in adata.obsm:
    sc.tl.umap(adata)

# popV built-in visualization: UMAP panel showing consensus + agreement
popv.visualization.predict_celltypes_umap(
    adata,
    save="popv_annotation_umap.png",
)
print("Saved popv_annotation_umap.png")

# Custom UMAP panels
fig, axes = plt.subplots(1, 3, figsize=(21, 6))
sc.pl.umap(adata, color="popv_prediction", ax=axes[0],
           title="popV Consensus", legend_loc="on data",
           legend_fontsize=6, show=False)
sc.pl.umap(adata, color="popv_agreement", ax=axes[1],
           cmap="RdYlGn", vmin=0, vmax=1,
           title="Method Agreement Score", show=False)
sc.pl.umap(adata, color="_dataset", ax=axes[2],
           title="Reference vs Query", show=False)
plt.tight_layout()
plt.savefig("popv_custom_umap.png", dpi=150, bbox_inches="tight")
print("Saved popv_custom_umap.png")

Key Concepts

Method Ensemble and Majority Voting

popV runs each method independently; the final prediction is determined by plurality vote across all methods. The popv_agreement score equals the fraction of methods that voted for the winning label (e.g., 0.7 = 7/10 methods agreed). This design has several properties:

• Robustness: if one method fails or produces outlier labels, the consensus is unaffected if the remaining methods agree
• Uncertainty signal: low agreement does not mean the annotation is wrong — it often flags biologically interesting cells (transitional states, rare populations) that differ from all reference cell types
• Method independence: KNN-based methods depend on the embedding quality; tree-based methods (RF, XGBoost) work directly on expression; SVM works in feature space; CellTypist uses a separate logistic regression. Together they span multiple algorithmic families

Method Comparison

ONCLASS and Ontology-Aware Annotation

ONCLASS uses the Cell Ontology (CL) to represent cell types as nodes in a knowledge graph and predict unseen cell types by propagating similarity through the ontology. Unlike other methods, ONCLASS can predict a cell type that was not present in the training reference if it is ontologically adjacent to known types. Enable it by including "onclass" in the methods list.

Reference Quality Requirements

popV annotation quality scales directly with reference quality:

• Minimum cell count per type: 50–100 cells per label; rare types with <20 cells may be missed by KNN methods
• Balanced representation: highly imbalanced references (one type is 80% of cells) cause tree methods to be biased toward the majority class
• Label granularity: coarse labels (10 types) annotate reliably; fine-grained labels (100+ types) require a larger, matched reference

Common Workflows

Workflow 1: Standard Reference-Query Annotation

Goal: Annotate an unlabeled query dataset using a curated reference atlas end-to-end.

import popv
import scanpy as sc
import pandas as pd

# 1. Load data
adata_ref = sc.read_h5ad("reference_atlas.h5ad")   # has obs["cell_type"] and obs["batch"]
adata_query = sc.read_h5ad("query_dataset.h5ad")   # no cell type labels
if "batch" not in adata_query.obs.columns:
    adata_query.obs["batch"] = "query"

# 2. Preprocess: build joint normalized object
adata = popv.preprocessing.Process_Query(
    adata_ref,
    adata_query,
    ref_labels_key="cell_type",
    ref_batch_key="batch",
    query_batch_key="batch",
    unknown_celltype_label="unknown",
    save_path_trained_models="./popv_models/",
    n_epochs_unsupervised=100,
    n_epochs_semisupervised=30,
    use_gpu=True,
    hvg=4000,
)
print(f"Prepared: {adata.n_obs} total cells")

# 3. Run ensemble annotation
popv.annotation.annotate_data(adata)

# 4. Extract query results
query_mask = adata.obs["_dataset"] == "query"
query_annotations = adata[query_mask].obs[[
    "popv_prediction", "popv_agreement",
    "knn_harmony_popv", "scanvi_popv", "rf_popv", "xgboost_popv"
]].copy()

# 5. Transfer back to original query object
adata_query.obs = adata_query.obs.join(
    query_annotations, how="left"
)
print(f"Annotated {query_mask.sum()} query cells")
print(query_annotations["popv_prediction"].value_counts().head(10))

# 6. Save annotated query
adata_query.write_h5ad("annotated_query.h5ad", compression="gzip")
query_annotations.to_csv("popv_annotations.csv")
print("Saved annotated_query.h5ad and popv_annotations.csv")

Workflow 2: Confidence Filtering and Novel Cell State Detection

Goal: Separate high-confidence annotations from ambiguous cells; flag candidate novel or transitional states for manual review.

import popv
import scanpy as sc
import pandas as pd
import matplotlib.pyplot as plt

# Assume adata has been annotated (as in Workflow 1)
query_mask = adata.obs["_dataset"] == "query"
query_obs = adata[query_mask].obs.copy()

# Tier cells by agreement score
bins = [0.0, 0.5, 0.8, 1.01]
labels = ["low (<0.5)", "medium (0.5–0.8)", "high (≥0.8)"]
query_obs["confidence_tier"] = pd.cut(
    query_obs["popv_agreement"], bins=bins, labels=labels, right=False
)
print("Cells per confidence tier:")
print(query_obs["confidence_tier"].value_counts())

# High-confidence subset: use popv_prediction directly
high_conf_mask = query_obs["popv_agreement"] >= 0.8
print(f"\nHigh-confidence annotations ({high_conf_mask.mean():.1%} of query cells):")
print(query_obs[high_conf_mask]["popv_prediction"].value_counts().head(10))

# Low-confidence subset: inspect per-method disagreement
low_conf = query_obs[query_obs["popv_agreement"] < 0.5]
popv_method_cols = [c for c in query_obs.columns if c.endswith("_popv") and
                    c not in ("popv_prediction", "popv_agreement")]
print(f"\nLow-confidence cells sample (showing per-method labels):")
print(low_conf[popv_method_cols + ["popv_prediction"]].head(10).to_string())

# Visualize agreement distribution
fig, axes = plt.subplots(1, 2, figsize=(14, 5))
query_obs["popv_agreement"].hist(bins=20, ax=axes[0], color="steelblue", edgecolor="white")
axes[0].axvline(0.8, color="red", linestyle="--", label="High-confidence threshold")
axes[0].set_xlabel("Method Agreement Score")
axes[0].set_ylabel("Cell Count")
axes[0].set_title("popV Agreement Distribution")
axes[0].legend()

query_obs["confidence_tier"].value_counts().plot.bar(ax=axes[1], color="steelblue")
axes[1].set_title("Cells by Confidence Tier")
axes[1].set_xlabel("Confidence Tier")
axes[1].set_ylabel("Cell Count")
plt.tight_layout()
plt.savefig("popv_confidence_distribution.png", dpi=150, bbox_inches="tight")
print("Saved popv_confidence_distribution.png")

Key Parameters

Best Practices

1. Check gene overlap before running: popV performs best with >70% gene overlap between reference and query. If overlap is <50%, annotation quality degrades significantly — consider using a different reference or imputing missing genes.

   shared = adata_ref.var_names.intersection(adata_query.var_names)
   print(f"Gene overlap: {len(shared) / adata_ref.n_vars:.1%}")
   

2. Use raw counts as input: pass raw (un-normalized) counts in adata.X to Process_Query. popV internally applies its own normalization. Pre-normalized data can distort the scVI/SCANVI latent space.

3. Match reference granularity to query biology: if your query contains subtypes not in the reference, no method will correctly assign them — they will appear as low-agreement cells. Either add them to the reference or accept that the consensus will assign the nearest parent type.

4. Exclude slow methods when speed matters: scanvi_popv and onclass are the slowest. For a quick first-pass, run only knn_harmony, knn_bbknn, rf, xgboost, and celltypist_popv.

   popv.annotation.annotate_data(adata, methods=["knn_harmony", "knn_bbknn", "rf", "xgboost", "celltypist_popv"])
   

5. Save trained models for repeated queries: Process_Query stores scVI/SCANVI models in save_path_trained_models. Reuse these when annotating additional query batches against the same reference to avoid retraining.

Common Recipes

Recipe: Subset to High-Confidence Annotations Only

When to use: downstream analyses (DE, trajectory) require clean labels; exclude ambiguous cells.

import scanpy as sc

# Annotate as in Workflow 1 first
query_mask = adata.obs["_dataset"] == "query"
adata_query_annotated = adata[query_mask].copy()

# Keep only high-confidence cells
high_conf = adata_query_annotated[adata_query_annotated.obs["popv_agreement"] >= 0.8].copy()
print(f"High-confidence cells: {high_conf.n_obs} / {adata_query_annotated.n_obs} "
      f"({high_conf.n_obs/adata_query_annotated.n_obs:.1%})")
print(high_conf.obs["popv_prediction"].value_counts())

# Recompute UMAP on high-confidence subset for visualization
sc.pp.neighbors(high_conf, use_rep="X_scVI")  # use scVI embedding stored by popV
sc.tl.umap(high_conf)
sc.pl.umap(high_conf, color="popv_prediction", save="_high_conf_celltypes.png")

Recipe: Per-Method Label Comparison Heatmap

When to use: understanding where methods disagree to identify systematic biases or novel populations.

import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

query_mask = adata.obs["_dataset"] == "query"
query_obs = adata[query_mask].obs.copy()

# Collect per-method columns
method_cols = [c for c in query_obs.columns
               if c.endswith("_popv") and c not in ("popv_prediction", "popv_agreement")]

# Cross-tabulate two key methods
ct = pd.crosstab(
    query_obs["knn_harmony_popv"],
    query_obs["scanvi_popv"],
    margins=False,
)
# Normalize rows
ct_norm = ct.div(ct.sum(axis=1), axis=0)

plt.figure(figsize=(12, 10))
sns.heatmap(ct_norm, cmap="Blues", vmin=0, vmax=1,
            xticklabels=True, yticklabels=True,
            cbar_kws={"label": "Fraction of cells"})
plt.title("knn_harmony vs scanvi label agreement")
plt.xlabel("SCANVI label")
plt.ylabel("KNN-Harmony label")
plt.tight_layout()
plt.savefig("popv_method_agreement_heatmap.png", dpi=150)
print("Saved popv_method_agreement_heatmap.png")

Recipe: Fast Annotation Without Deep Learning Methods

When to use: quick annotation without GPU or when scVI/SCANVI training is prohibitively slow (>500k cells).

import popv

# Process without training deep generative models (scVI not needed for KNN-Harmony)
adata = popv.preprocessing.Process_Query(
    adata_ref,
    adata_query,
    ref_labels_key="cell_type",
    ref_batch_key="batch",
    query_batch_key="batch",
    unknown_celltype_label="unknown",
    save_path_trained_models="./popv_models/",
    n_epochs_unsupervised=0,   # skip scVI training
    n_epochs_semisupervised=0, # skip scANVI training
    use_gpu=False,
    hvg=3000,
)

# Run only fast non-DL methods
popv.annotation.annotate_data(
    adata,
    methods=["knn_harmony", "knn_bbknn", "knn_scanorama", "rf", "xgboost", "svm", "celltypist_popv"],
)

query_mask = adata.obs["_dataset"] == "query"
print(adata[query_mask].obs[["popv_prediction", "popv_agreement"]].describe())

Troubleshooting

Related Skills

• celltypist-cell-annotation — single-model annotation with pre-trained logistic regression; faster but lacks ensemble uncertainty
• scanpy-scrna-seq — preprocessing pipeline (QC, normalization, clustering) that produces AnnData inputs for popV
• scvi-tools-single-cell — scANVI for probabilistic label transfer with a single deep generative model; use when you prefer a formal variational framework over ensemble voting
• harmony-batch-correction — Harmony embedding used by knn_harmony method internally; understand it to tune popV's KNN-based methods

References

• GitHub: YosefLab/popV — official source code, installation instructions, and example notebooks
• popV documentation — API reference and tutorials
• Ergen et al., bioRxiv 2023 — "Population-level integration of single-cell datasets enables multi-scale analysis across samples", original popV preprint
• ONCLASS paper — Wang et al., Nature Methods 2021 — ontology-aware cell type classification underlying the ONCLASS method in popV