sbol-torch

A PyTorch library for synthetic biology and biodesign automation.

Installed as sbol-torch, imported as sboltorch (commonly import sboltorch as st).

sbol-torch pulls designs from a running sbol-db instance (or local SBOL/FASTA files), normalizes them into a single record type, and trains transformer models against them. The input modality, tokenizer, and training objective are all set in configuration, so trying a new combination never means forking the pipeline.

Capabilities

Install

pip install sbol-torch

For development:

uv venv
uv pip install -e '.[dev]'

Quickstart

A run is fully specified by one YAML config. From the command line:

# Materialize a corpus to the local Parquet cache (offline, reproducible).
sboltorch ingest examples/configs/finetune_expression.yaml

# Train. Resolved config, per-epoch metrics.jsonl, and best.pt land in output_dir.
sboltorch train examples/configs/finetune_expression.yaml

# Resume an interrupted run from its rolling checkpoint (needs checkpoint_every_n_steps).
sboltorch train examples/configs/pretrain_mlm.yaml --resume runs/pretrain_mlm/last.pt

# Generate from a trained causal backbone — point model.backbone at the run's
# backbone/ (with from_scratch: false), then complete a design from a prompt.
sboltorch generate my_causal_run.yaml --prompt ATGCGT --max-new-tokens 200 --temperature 0.8

Train multi-GPU with torchrun and train.distributed.strategy: ddp:

torchrun --nproc_per_node=<gpus> -m sboltorch.cli train examples/configs/pretrain_causal_long.yaml

Or from Python:

import sboltorch as st

config = st.RunConfig.from_yaml("examples/configs/train_graph.yaml")
metrics = st.run_training(config)

Example configs

Experiment tracking

The two synthetic-data configs (train_graph.yaml and finetune_structure_aware.yaml) ship with Weights & Biases enabled. Set WANDB_API_KEY in a .env at the repo root and run both:

python examples/run_wandb_examples.py

Each run logs per-step loss and learning rate, per-epoch train/val metrics, the resolved config, the corpus fingerprint and split sizes as lineage, and the best checkpoint as a model artifact.

Graph transformer	Structure-aware sequence

Documentation

Release history is in CHANGELOG.md.

Develop

uv run pytest
pre-commit run --all-files