Methodology

Deep-learning co-folding models are becoming increasingly adopted in drug discovery. However, benchmark results accompanying model releases can be hard to replicate, access, or inspect. CoFold Arena aims to provide a centralized, maximally transparent, and regularly updated benchmark for co-folding models on therapeutically relevant tasks.

Compared to existing benchmarks (e.g. FoldBench, PXMeter), CoFold Arena has three differentiating aims:

• Availability and transparency. All model predictions upstream of headline metrics can be inspected in-browser or easily downloaded for reanalysis.

• Up-to-date with the current PDB. All models are run on weekly PDB releases to provide a maximally recent, common, and evolving benchmark set. This (1) enables head-to-head comparisons of all available models instead of excluding models with later training cutoffs, and (2) eliminates the possibility of bias in retrospective evaluations by model developers.

• Customization. Instead of rigidly defined test sets, arbitrary PDB release ranges can be selected in-browser for evaluation.

Master cutoff: Evaluations are available for structures released after September 9, 2024, past the self-reported training cutoff of all models (except Protenix-v1-20250630).

Below we describe the methodology for the antibody-antigen co-folding leaderboard. This page will be updated with methodology for other leaderboards as they become available.

1. Target selection

Models are evaluated on SAbDab entries meeting all of the following, applied in order:

1. Release date after September 9, 2024.
2. One entry per antibody (by SABDAB_ID), taking its earliest-released structure.
3. Antigen in UniProt: the antigen is a protein with a UniProt accession and is not itself an antibody. Peptide, small-molecule, and accession-less antigens are dropped.
4. Copy selection: pick one physical copy of the antibody in the mmCIF and collect every valid antigen chain it contacts (excluding other antibody chains).
5. Antigen de-duplication: at most one antibody per antigen UniProt accession (earliest first).
6. Resolvable chains: Chains under labels that do not map to real auth_asym_ids in the deposited mmCIF are skipped.

2. Input preparation

1. Each target's chains and sequences are extracted from the deposited mmCIF; noncanonical or modified residues are replaced with X tokens, so every model sees the same query sequence.
2. Targets larger than 1024 tokens (total chain length) are dropped.
3. One MSA is built per unique chain sequence with a single ColabFold (MMseqs2) call, and the same unpaired MSA is reused by every method. We do not request the server's paired MSAs.
4. Each model is given taxonomy information in the unpaired MSA and performs MSA pairing with its own native logic.

3. Inference

• 10 trunk passes for every model (each model's recycle parameter set to match).
• Diffusion sampling. AF3-style models draw 5 diffusion samples and keep the top-1 by the model's own confidence (0.8·ipTM + 0.2·pTM). AF-Multimer runs the v3 weights (model_1_multimer_v3) and produces a single structure.
• A single seed. For AF3-style models it is well documented that drawing more seeds and keeping the best improves accuracy. To reduce benchmarking cost, however, we only run with a single seed. Differences in performance and relative rankings may emerge if running with more seeds.

4. Scoring

Each prediction is compared to the trimmed native with DockQ over the antibody–antigen interface(s) using tinyprot. The leaderboard summarizes the per-target DockQ two ways:

• Mean DockQ over the panel.
• Success rate — the fraction of targets scoring above a CAPRI quality threshold: acceptable (≥ 0.23), medium (≥ 0.49), or high (≥ 0.80).

Turning on Show uncertainty adds a paired bootstrap over the target panel: targets are resampled with replacement and every model is re-scored on the same resample, giving each metric a 95% confidence interval and each model a rank range.