# Eliminating Conda/CUDA dependency hell in computational biology pipelines

> Source: <https://github.com/evo-design/proto-tools>
> Published: 2026-07-21 04:27:27+00:00

Welcome! This repository contains the open-source implementation of `proto-tools`

, a Python package containing a large suite of computational biology and biological AI tools, all accessible through a single, consistent Python interface. Language models, structure predictors, inverse folding, sequence analysis, gene annotation, conformational dynamics, genomic scoring, and more are all available through a single `pip install`

command.

Every tool runs in its own automatically managed isolated environment, so all dependency wrangling is handled for you. In addition, `proto-tools`

implements extensive infrastructure for features such as device management and GPU fan-out, making it easy to call tools in quick succession. You can use it as a standalone Python library, as part of the broader [proto-language](https://github.com/evo-design/proto-language) optimization system, or through the [proto-client](https://github.com/evo-design/proto-client) Python SDK for hosted access over the Proto Bio API.

Proto-tools is open source under an MIT license. Contributions are welcome!

Proto-tools requires Python 3.10+:

```
pip install git+https://github.com/evo-design/proto-tools.git
```

Note

A direct PyPI install (`pip install proto-tools`

) will be available soon.

Note

If you are developing or contributing to this project, follow the setup instructions in [CONTRIBUTING.md](/evo-design/proto-tools/blob/main/CONTRIBUTING.md) instead.

All persistent data (model weights and tool environments) is cached under the `PROTO_HOME`

directory on first use (defaults to `~/.proto/`

).

To customize the storage location, you can specify a path via the following environment variable:

```
# Add to your shell profile:
export PROTO_HOME=/path/to/your/proto_home
```

For shared filesystems, model weights can be reused to avoid downloading duplicate copies. The `PROTO_MODEL_CACHE`

environment variable lets you point just the weights at that shared location (sharing tool environments is not recommended): `export PROTO_MODEL_CACHE=/path/to/shared/weights`

. See [notes/storage.md](/evo-design/proto-tools/blob/main/notes/storage.md) for all details and options.

A few tools use gated models or software that require accepting a license / terms-of-use first (e.g. ESM3, AlphaGenome, AlphaFold3, X3DNA). See [notes/gated-models.md](/evo-design/proto-tools/blob/main/notes/gated-models.md) for the full list and per-model access steps.

Tip

**You're all set up!** To learn what features are available in the library, check out the [guides](/evo-design/proto-tools/blob/main/guides) — four short notebooks covering tool environments, persistent execution, device management, and parallel multi-GPU runs.

```
binder_design/                  # De novo antibody / binder design pipelines
├── bindcraft/
├── freebindcraft/
└── germinal/
causal_models/                 # Autoregressive sequence models
├── evo1/
├── evo2/
├── progen2/
└── progen3/
database_retrieval/             # Sequence and structure database access
├── alphafold_db/
├── alphamissense_db/
├── ccd_lookup/
├── ensembl/
├── interproscan/
├── ncbi/
├── pdb/
├── pubchem/
├── sequence_fetch/
└── uniprot/
gene_annotation/                # Sequence annotation
├── crispr_tracr_rna/
├── meme/
├── minced/
├── miranda/
├── promoter_calculator/
└── pyhmmer/
inverse_folding/                # Sequence design from structures
├── esm_if1/
├── fampnn/
├── ligandmpnn/
└── proteinmpnn/
masked_models/                  # Masked language models
├── ablang/
├── esm2/
├── esm3/
└── esmc/
mutagenesis/                    # Random sequence mutagenesis
├── random_nucleotide/
└── random_protein/
orf_prediction/                 # Open reading frame detection
├── orfipy/
└── prodigal/
rna_splicing/                   # RNA splice site prediction
├── pangolin/
├── splice_transformer/
└── spliceai/
sequence_alignment/             # Sequence search and multiple sequence alignment
├── blast/
├── mafft/
└── mmseqs2/
sequence_scoring/               # Genomic and regulatory scoring
├── alphagenome/
├── borzoi/
├── deeppbs_specificity/
├── enformer/
├── malinois/
├── na_mpnn_specificity/
├── puffin/
└── segmasker/
structure_alignment/            # Structure comparison
├── foldmason/
├── foldseek/
├── pymol_rmsd/
├── tmalign/
└── usalign/
structure_design/               # De novo structure generation
└── rfdiffusion3/
structure_dynamics/             # Conformational dynamics
└── bioemu/
structure_prediction/           # 3D structure prediction
├── alphafold2/
├── alphafold3/
├── boltz2/
├── chai1/
├── esmfold/
├── esmfold2/
├── protenix/
├── rf3/
├── viennarna/
└── x3dna/
structure_scoring/              # Structure quality scoring
├── dssp/
├── ipsae/
├── metal3d/
├── pdockq2/
├── pyrosetta/
└── structure_metrics/
```

Runnable walkthroughs of the core framework features live in [ guides/](/evo-design/proto-tools/blob/main/guides) and are also available on our

[docs page](https://proto.evodesign.org/docs/tools/introduction):

[Tool Environments](/evo-design/proto-tools/blob/main/guides/tool_environments.ipynb)— how isolated environments are built and cached on first call.[Tool Persistence](/evo-design/proto-tools/blob/main/guides/tool_persistence.ipynb)— keep models warm across calls[Device Management](/evo-design/proto-tools/blob/main/guides/device_management.ipynb)— GPU allocation, LRU eviction, CPU offload[Parallel Execution](/evo-design/proto-tools/blob/main/guides/parallel_execution.ipynb)— fan out work across every GPU with`ToolPool`

Each specific tool also ships a minimal `examples/example.ipynb`

under `proto_tools/tools/{category}/{tool}/examples/`

.

Run tools through natural language with any coding agent (Claude Code, Gemini CLI, OpenAI Codex CLI, etc.). Point the agent at `proto-tools agent-context`

: it prints a primer covering the `Input → Config → run_*() → Output`

pattern, the offline CLI discovery verbs, persistence and parallel execution, and links to the long-form notes on GitHub. The command ships in the wheel, so it works on a plain `pip install`

with no repo checkout.

If you've cloned the repo for contributing, agents also pick up `CLAUDE.md`

(symlinked as `AGENTS.md`

/`GEMINI.md`

) and the task-specific guides in [ .claude/skills/](/evo-design/proto-tools/blob/main/.claude/skills) automatically.

See [CONTRIBUTING.md](/evo-design/proto-tools/blob/main/CONTRIBUTING.md) for full developer setup, storage
configuration, PR format, code style, and testing conventions.

If you use Proto in your research, please cite our preprint:

Merchant AT, Guo D, Viggiano B, Brennan-Almaraz LE, Hur E, Mai T, Yin P, King SH, Ashley E, Hie BL.

A high-level programming language for generative biology with Proto.bioRxiv(2026). doi:[10.64898/2026.06.22.733870]

```
@article{Merchant2026.06.22.733870,
  author = {Merchant, Aditi T and Guo, Daniel and Viggiano, Ben and Brennan-Almaraz, Lucas Emmanuel and Hur, Evelyn and Mai, Tina and Yin, Peter and King, Samuel H and Ashley, Euan and Hie, Brian L},
  title = {A high-level programming language for generative biology with Proto},
  elocation-id = {2026.06.22.733870},
  year = {2026},
  doi = {10.64898/2026.06.22.733870},
  publisher = {Cold Spring Harbor Laboratory},
  URL = {https://www.biorxiv.org/content/10.64898/2026.06.22.733870},
  journal = {bioRxiv}
}
```