Bioverse

Machine learning on biomolecules

Bioverse is a standardized framework for machine-learning experiments on biomolecules — proteins, RNA, small molecules, and related structures. It combines reusable benchmarks, transforms, and evaluation components with a config-driven CLI so you can train and evaluate models on common tasks without reimplementing data loading, splitting, or metrics.

What you can do

  • Run an experiment from a YAML config and the bioverse CLI

  • Browse built-in datasets, benchmarks, metrics, tasks, and transforms

  • Extend the library with new components and contribute them upstream

Get started

pip install bioverse-ml
pip install torch lightning   # required for training
bioverse train experiment.yaml

See Quickstart for a complete minimal example (model + config + command).

Choose your path

Run experiments with existing components

Start with Quickstart, then read the User Guide for configuration, CLI workflows, and troubleshooting.

Understand the architecture and add components

Read the Developer Guide, then follow the How to tutorials for datasets, benchmarks, metrics, and other component types.

Contribute to the repository

See the Contributor Guide for naming conventions, tests, and pull-request expectations.

How it works

Adapter  →  Dataset  →  Benchmark  →  Trainer + Model
                       (sampler,
                        task,
                        metric)

  1. An adapter downloads or generates raw data.

  2. A dataset stores versioned shards and splits (optionally via transforms).

  3. A benchmark wires the dataset to a sampler, task, and metric.

  4. The trainer runs your model and logs results.

Details: Developer Guide and Code Structure.

Built-in library

Bioverse ships with a growing catalog of ready-to-use components. Browse the Implementations section for auto-generated reference pages with docstrings and configs.

Featured benchmarks:

  • B_AFCATH — inverse folding on CATH domains with recovery and BLOSUM metrics

  • B_PROGYM — protein mutational effect prediction (ProteinGym)

  • B_INVC42 — inverse folding on CATH structures

See Benchmarks and Datasets for the full lists.

Citation

@software{bioverse2026,
  author = {Kucera, Tim and Bioverse Contributors},
  title = {Bioverse: A standardized framework for machine learning on biomolecules},
  year = {2026},
  url = {https://github.com/BorgwardtLab/bioverse}
}

See Citation for the full reference and author contact details.

Guides

How to

API Reference

Implementations