novoweave

本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

NovoWeave

图片：Concept 图片：Python 图片：License: MIT 图片：Responsible use

A conceptual generative protein-design framework. NovoWeave is a research-oriented software blueprint for weaving backbone generation, sequence design, evaluation, and provenance into one end-to-end deep-learning workflow. It presents the architecture, interfaces, configuration system, and engineering conventions that a production project could use

Why this repository exists

Protein-design projects often mix data ingestion, geometric modeling, sequence generation, ranking, and experiment tracking in one code path. This blueprint separates those concerns behind small, testable contracts so researchers can discuss system design before committing to an implementation.

Conceptual workflow

flowchart LR
    A[Design brief] --> B[Constraint parser]
    B --> C[Backbone generator]
    C --> D[Sequence designer]
    D --> E[In-silico filters]
    E --> F[Ranked candidates]
    F --> G[Human review]

The conceptual pipeline has four replaceable stages:

1. Backbone proposal — a diffusion-style geometric model interface.
2. Sequence design — a structure-conditioned transformer interface.
3. Evaluation — confidence, geometry, and diversity metric contracts.
4. Selection — auditable multi-objective ranking with human review gates.

Repository layout

configs/                 Example experiment configurations
docs/                    Architecture, model card, and research scope
src/novoweave/           Typed Python package and pseudocode interfaces
tests/                   Contract tests for implemented scaffolding
.github/                 CI and community health files

Illustrative usage

The following snippet documents the intended API. It is not an executable protein-design example.

from novoweave import DesignBrief, DesignPipeline

brief = DesignBrief(
    name="example_scaffold",
    length=120,
    objective="Demonstrate the software contract only",
)

pipeline = DesignPipeline.from_config("configs/base.yaml")
result = pipeline.design(brief)  # intentionally raises NotImplementedError

The planned command-line surface is similarly illustrative:

novoweave validate-config configs/base.yaml
novoweave design --config configs/base.yaml --brief examples/brief.yaml

Project status

This is a conceptual scaffold / design document in code form, not a model release. The following are deliberately absent:

• trained weights, datasets, and downloadable checkpoints;
• working tensor kernels, loss functions, or samplers;
• experimentally validated scoring functions;
• synthesis instructions or biological claims;
• benchmark results.

Implemented pieces are limited to configuration/data contracts, boundary validation, logging conventions, and repository tooling. See the roadmap for a hypothetical implementation sequence.

Design principles

• Explicit scientific boundaries: interfaces distinguish hypotheses from validated outputs.
• Reproducibility by construction: every planned run records configuration, code revision, seed, and provenance.
• Modular research: geometry, sequence, evaluation, and ranking components can evolve independently.
• Responsible defaults: no autonomous wet-lab handoff and no claims of safety or efficacy.

Development

For repository and documentation work only:

python -m pip install -e ".[dev]"
pytest
ruff check .

These checks validate the scaffold; they do not test protein-design capability. See CONTRIBUTING.md before proposing changes.

Citation

If this blueprint is useful in a discussion or teaching context, cite the repository metadata in CITATION.cff.

License

Released under the MIT License. Scientific validity, fitness for a particular purpose, and biological safety are explicitly not warranted.