AF3-NA-plus

AF3-NA+

AF3-NA+ is a sequence-native, DNA+RNA neural-template sidecar for AlphaFold 3. It converts nucleotide sequences into four structural-template slots and supplies those features to the separately installed AF3 runtime via the included integration patch. The released weight is not a standalone coordinate predictor without dependency of AF3.

This repository contains the BF16 sidecar weights, tokenizer and model configuration, a small PyTorch inference runtime, and the pinned AF3 integration contract.

Enhancement

The experiments are run followed complete literature-matched suite and its entry-mean lDDT estimator on the percentage scale. AF3-NA+ candidates use Main-25 generation; the SelfRank-Top1 is selected before reference structures are opened for scoring. Published AF3 values are taken from the official Nature 2023 paper.

Quick start

Python 3.10 or newer is recommended. Runtime dependencies are PyTorch, NumPy, and safetensors.

python -m pip install .
python hub_predict.py \
  --model-dir . \
  --input examples/input.json \
  --output output/ \
  --device cuda \
  --dtype bfloat16

The input may be an AlphaFold 3 JSON or a compact chains JSON like examples/input.json. The command writes a K=4 neural-template artifact as manifest.json plus arrays.npz. It does not run AF3 itself.

The Python loader is also available directly:

from hub_loader import load_release
release = load_release(".", device="cuda", dtype="bfloat16")

AlphaFold 3 integration

AF3-NA+ expects a separately installed AlphaFold 3 runtime and model-parameter directory. It does not include AlphaFold 3 source or weights. See integrations/alphafold3/README.md for the pinned upstream revision, AF3 parameter download, patch application, and resolver binding details.

Scope and limitations

• The packaged model produces neural-template features; AF3 is required for coordinate generation.
• The runtime accepts canonical DNA/RNA sequences and explicit chain identity. Modified residues and unsupported chemistries require separate validation.
• Memory and runtime grow with sequence length and pair coverage.
• AF3 is independently distributed and has its own terms. Obtain it from its official source when needed.
• This package is research software and is not validated for clinical use.

Licensing

Hugging Face metadata uses license: other because no single license covers every file. Project-authored code and AF3-NA+ weights are under GPL-3.0-only; the AF3-derived patch is covered by the bundled CC-BY-NC-SA-4.0 text from the pinned AF3 revision. See NOTICE.md for the file-level boundary.