Harmonypy

harmonypy

Use this skill for fast exploratory batch correction of single-cell embeddings with Harmony. Harmony operates on a low-dimensional representation, usually PCA, and returns corrected coordinates for neighbor graph construction and visualization.

When To Use

• Multi-sample scRNA-seq integration where batches dominate PCA/UMAP.
• Quick exploratory integration before deeper modeling.
• Scanpy pipelines that need a lightweight correction step between PCA and neighbors.

Prefer scvi-tools when the project needs a probabilistic model, covariates, transfer learning, or robust multi-modal integration.

Scanpy Pattern

[!WARNING] Scanpy 1.12.1's scanpy.external.pp.harmony_integrate still writes harmony_out.Z_corr.T. That is the right orientation for harmonypy 0.0.x, but harmonypy >=0.1.0 already returns Z_corr as cells × PCs. With current harmonypy, use the direct pattern below and store ho.Z_corr in adata.obsm["X_pca_harmony"]; otherwise pin harmonypy 0.0.10 before using the Scanpy wrapper.

import scanpy as sc
import scanpy.external as sce

sc.pp.normalize_total(adata)
sc.pp.log1p(adata)
sc.pp.highly_variable_genes(adata, batch_key="batch")
adata = adata[:, adata.var["highly_variable"]].copy()
sc.pp.scale(adata, max_value=10)
sc.tl.pca(adata)

sce.pp.harmony_integrate(adata, key="batch")
sc.pp.neighbors(adata, use_rep="X_pca_harmony")
sc.tl.umap(adata)
sc.tl.leiden(adata)

Direct harmonypy Pattern

Current version: harmonypy 2.0.0 (April 2026, C++ backend rewrite matching R harmony2). Pre-built wheels for Linux and macOS (Python 3.9–3.13); Windows is not supported.

This no-transpose pattern requires harmonypy >=0.1.0. In reproducible environments pinned to harmonypy 0.0.x, Z_corr is PCs × cells and the old ho.Z_corr.T orientation is still required.

import harmonypy as hm

ho = hm.run_harmony(pca_matrix, metadata, vars_use=["batch"])
corrected = ho.Z_corr  # harmonypy >=0.1.0: cells × PCs; use ho.Z_corr.T on 0.0.x

Checks

• Confirm that the correction variable is technical, not the biology of interest.
• Plot UMAPs colored by both batch and expected biological labels.
• Compare marker genes before and after correction to catch overcorrection.
• Do not run Harmony on raw counts; use PCA or another embedding.
• Keep uncorrected PCA and raw expression available for differential expression.