ZZL2005/msa-pairing-agent-results

MSA Pairing Agent: experimental inputs and results

Experimental inputs and score tables for MSA Pairing Agent: Blind inference on FoldBench-v1 (FB1), FoldBench-v2 (FB2), and HD439, plus GT-guided inference on FB1.

Each Agent archive contains the starting AF3 input for every included episode and the recorded builder plus actual AF3 input for every round, including the native Default control. The baseline archives contain the actual AF3 inputs for mMSA, pMSA, and uMSA. The score tables identify targets, interfaces, and selected submissions.

Files and coverage

Archive	Targets / episodes	Recorded rounds or baseline inputs
agent/fb1_blind.tar.zst	238	1,558 rounds
agent/fb2_blind.tar.zst	283	2,068 rounds
agent/hd439_blind.tar.zst	438	3,094 rounds
agent/fb1_gt_guided.tar.zst	239	2,617 rounds
baselines/fb1.tar.zst	239	717 inputs: three methods per target
baselines/fb2.tar.zst	283	849 inputs: three methods per target
baselines/hd439.tar.zst	438	1,314 inputs: three methods per target

In total: 1,198 starting inputs, 9,337 builder/input pairs, 2,880 baseline inputs, and 12 score CSV files. The seven archives total 5.27 GiB to download; their extracted data total approximately 82.53 GiB.

FB1 Blind uses the reported merged cohort, retaining the earlier batch for overlapping targets. Target 8BVK has no completed Blind selection and is excluded from that archive; it is included in the FB1 GT-guided and baseline archives. The published Blind versus GT-guided comparison uses their 238 shared completed targets. HD439 excludes 7ACW_1, which lacks the required MSA. FB2 uses the main repeat_01 experiment; its score tables use the corrected OST evaluations.

Archive layout

Extracted Agent archives preserve the episode-relative paths used by the score tables:

fb1_blind/                         # or fb2_blind, hd439_blind, fb1_gt_guided
  <CASE>/
    input/source.json
    artifacts/
      001/builder.py
      001/input.json
      002/builder.py
      002/input.json
      ...

• input/source.json is the prepared input provided to the Agent before its MSA changes, including the starting sequences, MSAs, and templates. The host normalized its name to target when creating the episode. It is an episode input snapshot, rather than an unprocessed upstream dataset download.
• artifacts/NNN/builder.py is the recorded builder for submission NNN.
• artifacts/NNN/input.json is the actual AF3 input used for that submission, after the candidate's MSA changes and AF3 seeds have been applied. It contains the MSA organization needed to rerun that submission.
• Submission numbers in the CSV files are one-based: submission 5 maps to artifacts/005/. Round 001 is the native AF3 Default control.

Baseline archives extract as follows:

fb1/                              # or fb2, hd439
  mmsa/<CASE>/input.json
  pmsa/<CASE>/input.json
  umsa/<CASE>/input.json

These are the final baseline AF3 inputs. Preserve both pairedMsa and unpairedMsa channels exactly as supplied. The project README describes the baseline construction and experiment entry points.

Download and extract

Install the Hugging Face CLI, then download the complete release:

python -m pip install -U huggingface_hub
hf download ZZL2005/msa-pairing-agent-results \
  --repo-type dataset --local-dir msa-pairing-results

To download just one archive, supply its filename:

hf download ZZL2005/msa-pairing-agent-results agent/fb1_blind.tar.zst \
  --repo-type dataset --local-dir msa-pairing-results

With zstd and GNU tar installed, extract it into a directory of your choice:

mkdir -p expanded
tar --zstd -xf msa-pairing-results/agent/fb1_blind.tar.zst -C expanded

Each archive can be downloaded and extracted independently. Scores are small CSV files that can also be read directly on the Hub; see their documentation for aggregation and selection definitions.

Rerun a prepared AF3 input

The study used AF3 3.0.2, seeds [42, 66, 101, 2024, 8888], five diffusion samples per seed, and ten recycles. Seeds are already included in the released round and baseline inputs. With an AF3 installation and its model parameters available, for example:

/path/to/af3-env/bin/python /path/to/alphafold3/run_alphafold.py \
  --json_path=expanded/fb1_blind/7QUV/artifacts/007/input.json \
  --model_dir=/path/to/af3-model-parameters \
  --output_dir=predictions/fb1_blind/7QUV/007 \
  --norun_data_pipeline \
  --resolve_msa_overlaps=false \
  --num_recycles=10 \
  --num_diffusion_samples=5

Use --resolve_msa_overlaps=true for the native Default control (001). Use false for Agent MSA candidates and all three fixed baselines. Use a separate output directory for each target, method, and round. The data pipeline is disabled because the inputs already contain the prepared MSAs and templates.

Builders accept --source, --output, and --seed. The experiment's builder seed was 0. A builder produces a candidate description; the project backend merges it into the starting source to obtain the final AF3 input. The corresponding released input.json can be passed directly to AF3 without rebuilding that candidate.

Predicted structures, model weights, reference structures for evaluation, and full Agent conversation logs are outside this release. Repeating Agent exploration also requires a model provider, the experiment prompts, prepared dataset metadata, and evaluator setup described in the project repository. AF3 and evaluation software must be installed separately.

Dataset sources and terms

• FB1: FoldBench official repository, including its original target and reference-structure download links.
• HD439: MSApair official download page.
• FB2: an internal dataset from a collaborating team, awaiting its own public release. This repository provides the experimental inputs and derived results used in our study.

Third-party sequences, alignments, and template contents retain the terms of their original sources. No additional blanket license for third-party content is granted by this release.