You can deploy this model and start earning money today!

Genentech/
human-atac-catlas-model

Name: human-atac-catlas-model
Availability: InStock

pytorch-lightning

biology

genomics

tabular-classification

Genentech/human-atac-catlas-data

Genentech/enformer-model

finetune

10.57967/hf/8609

mit

Views

No views yet

human-atac-catlas-model

Model Description

This model is a multi-task classifier trained to predict the binary accessibility of genomic DNA sequences in 204 cell types. It was trained by fine-tuning the Enformer model using the grelu library on the human ATAC CATlas dataset.

Architecture: Fine-tuned Enformer (EnformerPretrainedModel)
Input: Genomic sequences (hg38)
Output: Probability of accessibility in 204 cell types
Parameters: 72M total (all trainable)

Performance

Metrics are computed per cell type and averaged across all 204 cell types.

Test Set

Metric	Mean	Std	Min	Max
Accuracy	0.9416	0.0175	0.8959	0.9743
AUROC	0.9053	0.0167	0.8634	0.9467
Average Precision	0.6097	0.0374	0.4545	0.7008
Best F1	0.5716	0.0289	0.4704	0.6395

Validation Set

Metric	Mean	Std	Min	Max
Accuracy	0.9482	0.0172	0.9071	0.9789
AUROC	0.8935	0.0190	0.8350	0.9379
Average Precision	0.5524	0.0370	0.4168	0.6888
Best F1	0.5253	0.0299	0.4285	0.6309

Training Details

Parameter	Value
Task	Binary classification
Loss	Binary Cross-Entropy
Optimizer	Adam
Learning rate	0.0001
Batch size	3072
Max epochs	10
n_transformers	1
crop_len	0
grelu version	1.0.4.post1.dev39

Repository Content

model.ckpt: The trained model weights and hyperparameters (PyTorch Lightning checkpoint).
2_train.ipynb: Jupyter notebook containing the training logic, architecture definition, and evaluation loops.
output.log: Training logs.

How to use

To load this model for inference or fine-tuning, use the grelu interface:

python

from grelu.lightning import LightningModel
from huggingface_hub import hf_hub_download

ckpt_path = hf_hub_download(
    repo_id="Genentech/human-atac-catlas-model",
    filename="model.ckpt"
)

model = LightningModel.load_from_checkpoint(ckpt_path, weights_only=False)
model.eval()