ATACformer — Foundation Model Tutorial¶

ATACformer — ATAC-seq-native transformer, peak-based (not gene-based) input, chromatin accessibility specialist

Property	Value
Tasks	embed, integrate
Species	human
Gene IDs	custom (peak-based)
GPU Required	Yes
Min VRAM	16 GB
Embedding Dim	512
Repository	https://github.com/Atacformer/Atacformer

Important: ATACformer requires scATAC-seq data (peak x cell matrix). It is not compatible with RNA expression data.

This tutorial demonstrates how to use ATACformer through the unified ov.fm API.

Cite: Zeng, Z. et al. (2024). OmicVerse: a framework for bridging and deepening insights across bulk and single-cell sequencing. Nature Communications, 15(1), 5983.

In [ ]:

Copied!





import omicverse as ov
import scanpy as sc
import os
import warnings
warnings.filterwarnings('ignore')

ov.plot_set()
import omicverse as ov
import scanpy as sc
import os
import warnings
warnings.filterwarnings('ignore')

ov.plot_set()

scATAC-seq preprocessing for ATACformer¶

Before running ATACformer, preprocess your scATAC-seq data:

import muon as mu
import scanpy as sc

# Load scATAC-seq data
adata_atac = sc.read_h5ad('atac_peaks.h5ad')

# Standard scATAC preprocessing
mu.atac.pp.tfidf(adata_atac)          # TF-IDF normalization
mu.atac.tl.lsi(adata_atac, n_comps=50) # LSI dimensionality reduction

# Save for ov.fm
adata_atac.write_h5ad('atac_preprocessed.h5ad')

# Run ATACformer
result = ov.fm.run(
    task='embed', model_name='atacformer',
    adata_path='atac_preprocessed.h5ad',
    output_path='atac_atacformer.h5ad',
)

Step 1: Inspect Model Specification¶

Use ov.fm.describe_model() to get the full spec for ATACformer.

In [ ]:

Copied!





info = ov.fm.describe_model("atacformer")

print("=== Model Info ===")
print(f"Name: {info['model']['name']}")
print(f"Version: {info['model']['version']}")
print(f"Tasks: {info['model']['tasks']}")
print(f"Species: {info['model']['species']}")
print(f"Embedding dim: {info['model']['embedding_dim']}")
print(f"Differentiator: {info['model']['differentiator']}")

print("\n=== Input Contract ===")
print(f"Gene ID scheme: {info['input_contract']['gene_id_scheme']}")
print(f"Preprocessing: {info['input_contract']['preprocessing']}")

print("\n=== Output Contract ===")
print(f"Embedding key: {info['output_contract']['embedding_key']}")
print(f"Embedding dim: {info['output_contract']['embedding_dim']}")
info = ov.fm.describe_model("atacformer")

print("=== Model Info ===")
print(f"Name: {info['model']['name']}")
print(f"Version: {info['model']['version']}")
print(f"Tasks: {info['model']['tasks']}")
print(f"Species: {info['model']['species']}")
print(f"Embedding dim: {info['model']['embedding_dim']}")
print(f"Differentiator: {info['model']['differentiator']}")

print("\n=== Input Contract ===")
print(f"Gene ID scheme: {info['input_contract']['gene_id_scheme']}")
print(f"Preprocessing: {info['input_contract']['preprocessing']}")

print("\n=== Output Contract ===")
print(f"Embedding key: {info['output_contract']['embedding_key']}")
print(f"Embedding dim: {info['output_contract']['embedding_dim']}")

Step 2: Prepare Data¶

Load a dataset and save it for the ov.fm workflow. Most foundation models expect raw counts (non-negative values).

In [ ]:

Copied!





# ATACformer requires scATAC-seq data (peak matrix, not gene expression).
# Replace with your own scATAC-seq dataset:
# adata = sc.read_h5ad('your_atac_data.h5ad')
#
# Typical scATAC preprocessing:
# import muon as mu
# mu.atac.pp.tfidf(adata)
# mu.atac.tl.lsi(adata, n_comps=50)

# For demonstration, we show the API pattern with PBMC RNA data.
# The validation step will correctly flag the modality mismatch.
adata = sc.datasets.pbmc3k()
sc.pp.filter_cells(adata, min_genes=200)
sc.pp.filter_genes(adata, min_cells=3)
adata.write_h5ad('pbmc3k_atacformer.h5ad')
print(f'Dataset: {adata.n_obs} cells x {adata.n_vars} genes')
print('Note: This is RNA data — ATACformer will flag incompatibility.')
# ATACformer requires scATAC-seq data (peak matrix, not gene expression).
# Replace with your own scATAC-seq dataset:
# adata = sc.read_h5ad('your_atac_data.h5ad')
#
# Typical scATAC preprocessing:
# import muon as mu
# mu.atac.pp.tfidf(adata)
# mu.atac.tl.lsi(adata, n_comps=50)

# For demonstration, we show the API pattern with PBMC RNA data.
# The validation step will correctly flag the modality mismatch.
adata = sc.datasets.pbmc3k()
sc.pp.filter_cells(adata, min_genes=200)
sc.pp.filter_genes(adata, min_cells=3)
adata.write_h5ad('pbmc3k_atacformer.h5ad')
print(f'Dataset: {adata.n_obs} cells x {adata.n_vars} genes')
print('Note: This is RNA data — ATACformer will flag incompatibility.')

Step 3: Profile Data & Validate Compatibility¶

Check whether your data is compatible with ATACformer before running inference.

In [ ]:

Copied!





profile = ov.fm.profile_data("pbmc3k_atacformer.h5ad")

print("=== Data Profile ===")
print(f"Species: {profile['species']}")
print(f"Gene scheme: {profile['gene_scheme']}")
print(f"Modality: {profile['modality']}")
print(f"Cells: {profile['n_cells']:,}")
print(f"Genes: {profile['n_genes']:,}")

# Validate compatibility
validation = ov.fm.preprocess_validate("pbmc3k_atacformer.h5ad", "atacformer", "embed")
print(f"\n=== Validation: {validation['status']} ===")
for d in validation.get("diagnostics", []):
    print(f"  [{d['severity']}] {d['message']}")
if validation.get("auto_fixes"):
    print("\nSuggested fixes:")
    for fix in validation["auto_fixes"]:
        print(f"  - {fix}")
profile = ov.fm.profile_data("pbmc3k_atacformer.h5ad")

print("=== Data Profile ===")
print(f"Species: {profile['species']}")
print(f"Gene scheme: {profile['gene_scheme']}")
print(f"Modality: {profile['modality']}")
print(f"Cells: {profile['n_cells']:,}")
print(f"Genes: {profile['n_genes']:,}")

# Validate compatibility
validation = ov.fm.preprocess_validate("pbmc3k_atacformer.h5ad", "atacformer", "embed")
print(f"\n=== Validation: {validation['status']} ===")
for d in validation.get("diagnostics", []):
    print(f"  [{d['severity']}] {d['message']}")
if validation.get("auto_fixes"):
    print("\nSuggested fixes:")
    for fix in validation["auto_fixes"]:
        print(f"  - {fix}")

Step 4: Run ATACformer Inference¶

Execute ATACformer through ov.fm.run(). The function handles preprocessing, model loading, inference, and output writing.

In [ ]:

Copied!





result = ov.fm.run(
    task="embed",
    model_name="atacformer",
    adata_path="pbmc3k_atacformer.h5ad",
    output_path="pbmc3k_atacformer_out.h5ad",
    device="auto",
)

if "error" in result:
    print(f"Error: {result['error']}")
    if "suggestion" in result:
        print(f"Suggestion: {result['suggestion']}")
else:
    print(f"Status: {result['status']}")
    print(f"Output keys: {result.get('output_keys', [])}")
    print(f"Cells processed: {result.get('n_cells', 0)}")
result = ov.fm.run(
    task="embed",
    model_name="atacformer",
    adata_path="pbmc3k_atacformer.h5ad",
    output_path="pbmc3k_atacformer_out.h5ad",
    device="auto",
)

if "error" in result:
    print(f"Error: {result['error']}")
    if "suggestion" in result:
        print(f"Suggestion: {result['suggestion']}")
else:
    print(f"Status: {result['status']}")
    print(f"Output keys: {result.get('output_keys', [])}")
    print(f"Cells processed: {result.get('n_cells', 0)}")

Step 5: Visualize & Interpret Results¶

Load the output, compute UMAP from ATACformer embeddings, and evaluate quality.

In [ ]:

Copied!





if os.path.exists("pbmc3k_atacformer_out.h5ad"):
    adata_out = sc.read_h5ad("pbmc3k_atacformer_out.h5ad")
    emb_key = "X_atacformer"
    
    if emb_key in adata_out.obsm:
        print(f"Embedding shape: {adata_out.obsm[emb_key].shape}")
        
        # UMAP visualization
        sc.pp.neighbors(adata_out, use_rep=emb_key)
        sc.tl.umap(adata_out)
        sc.tl.leiden(adata_out, resolution=0.5)
        sc.pl.umap(adata_out, color=["leiden"],
                   title="ATACformer Embedding (PBMC 3k)")
        
        # QA metrics
        interpretation = ov.fm.interpret_results("pbmc3k_atacformer_out.h5ad", task="embed")
        if "embeddings" in interpretation["metrics"]:
            for k, v in interpretation["metrics"]["embeddings"].items():
                print(f"\n{k}: dim={v['dim']}", end="")
                if "silhouette" in v:
                    print(f", silhouette={v['silhouette']:.4f}", end="")
                print()
    else:
        print(f"Embedding key {emb_key} not found.")
        print(f"Available keys: {list(adata_out.obsm.keys())}")
else:
    print("Output file not found — check model installation and adapter status.")
    print("See the Guide page for installation instructions.")
if os.path.exists("pbmc3k_atacformer_out.h5ad"):
    adata_out = sc.read_h5ad("pbmc3k_atacformer_out.h5ad")
    emb_key = "X_atacformer"
    
    if emb_key in adata_out.obsm:
        print(f"Embedding shape: {adata_out.obsm[emb_key].shape}")
        
        # UMAP visualization
        sc.pp.neighbors(adata_out, use_rep=emb_key)
        sc.tl.umap(adata_out)
        sc.tl.leiden(adata_out, resolution=0.5)
        sc.pl.umap(adata_out, color=["leiden"],
                   title="ATACformer Embedding (PBMC 3k)")
        
        # QA metrics
        interpretation = ov.fm.interpret_results("pbmc3k_atacformer_out.h5ad", task="embed")
        if "embeddings" in interpretation["metrics"]:
            for k, v in interpretation["metrics"]["embeddings"].items():
                print(f"\n{k}: dim={v['dim']}", end="")
                if "silhouette" in v:
                    print(f", silhouette={v['silhouette']:.4f}", end="")
                print()
    else:
        print(f"Embedding key {emb_key} not found.")
        print(f"Available keys: {list(adata_out.obsm.keys())}")
else:
    print("Output file not found — check model installation and adapter status.")
    print("See the Guide page for installation instructions.")

Summary¶

Step	Function	What it does
1	`ov.fm.describe_model("atacformer")`	Inspect model spec and I/O contract
2	`sc.datasets.pbmc3k()`	Prepare input data
3	`ov.fm.profile_data()` + `preprocess_validate()`	Check compatibility
4	`ov.fm.run()`	Execute ATACformer inference
5	`ov.fm.interpret_results()`	Evaluate embedding quality

For the full model catalog, see ov.fm.list_models() or the ov.fm API Overview. For detailed ATACformer specifications, see the ATACformer Guide.