Deep learning scaffolding for context-conditioned protein sequence generation and rule-based safety checks, as described in the Project Celery specification.
Note: The PyPI package name is
celery-xenoand the CLI iscelery-xeno, notcelery, to avoid conflicting with the widely used Celery distributed task queue.
cd "/path/to/Celery Project"
pip install -e ".[dev]"fair-esm is a core dependency; ESM-2 weights are fetched when the encoder first loads (network required unless cached).
Build training batches from real .h5ad + FASTA, then train (config must set batches_path, or pass --batches):
python scripts/preprocess_all.py --h5ad path/to/cells.h5ad --fasta path/to/proteins.fasta \
--cell-type keratinocyte --output data/processed/training_batches.pt
celery-xeno train --config configs/training/adversarial.yaml --batches data/processed/training_batches.pt --dry-runDesign requires a checkpoint on disk, the same scRNA matrix used for training (or one preprocessed with matching QC / HVG settings), and n_genes in the checkpoint (or --n-genes):
celery-xeno design --cell-type keratinocyte --checkpoint checkpoints/best_model.pt \
--scrna-h5ad path/to/cells.h5ad --n-candidates 4 --output outputs/Evaluate a protein + RNA pair:
celery-xeno evaluate --sequence MKVLWAIVLFGDWGCVGC --cell-type keratinocyte --rna GGCCUGCUGAG...celery/— library code (models, training, inference, safety)configs/— YAML configsscripts/— Python entry scriptstests/— pytest
This repository is research software. It does not validate therapeutic safety, regulatory compliance, or wet-lab feasibility. Any use toward therapeutics requires independent expert review.
Apache-2.0 (see LICENSE).