---
license: cc-by-4.0
tags: [protein, binder-design, alphafold2, alphafold3, ensemble, uncertainty]
configs:
- config_name: ensemble
  default: true
  data_files: manifest.csv
- config_name: per_sample
  data_files: af3_k5_persample.csv
---

# litscrape-af2af3-k5: AF2 + AF3 metrics with a K=5 AF3 diffusion ensemble

Per-design AlphaFold2 and AlphaFold3 confidence metrics for the binder-target
pairs in [`yk0/litscrape`](https://huggingface.co/datasets/yk0/litscrape), extended with a
**K=5 AlphaFold3 diffusion ensemble** for uncertainty quantification.
4020 rows (one per binder-target pair); `af3_k5_persample.csv` holds all 5
samples per design.



## Two views (config selector in the viewer)
- **ensemble** (default): one row per design, with AF2, AF3 top-1, and the K=5 ensemble aggregates (`af3k5_*` mean/std/min/max/best).
- **per_sample**: the raw **5 AF3 diffusion samples per design** (`sample-0..4`), each with iptm/ptm/binder_plddt/ranking_score, plus the design metadata and `label`. This is where the 5 generations are visible.

## How metrics were computed (sequences only; no input structures)
1. **Target** folded once with AF2 monomer (ColabDesign hallucination) to get a template.
2. **AF2** (`af2_*`): de-novo AF2-Multimer (model_1_multimer_v3, 3 recycles, no MSA, no initial guess) on templated target + binder sequence.
3. **AF3 top-1** (`af3_*`): AF3Score with **init_guess=true** — scores the AF2-predicted complex (anchored, deterministic, 1 sample).
4. **AF3 K=5 ensemble** (`af3k5_*`): AF3Score with **init_guess=false, num_samples=5** — genuine de-novo diffusion sampling. For ptm, iptm, binder_plddt, ranking_score we report `_mean / _std / _min / _max / _best` (best = highest ranking_score sample). **`_std` is the diffusion-noise uncertainty.**

## Key caveats
- **Do not compare `af3_*` (top-1) with `af3k5_*` directly**: they use different modes (init_guess=true vs false). The K=5 ensemble *requires* init_guess=false, since the anchored scoring mode is deterministic (1 sample). Comparisons should stay within-mode.
- **Diffusion spread is small** here (iptm_std median 0.015); the 5 samples agree closely for these mostly-low-confidence de-novo predictions.
- These are **de-novo predictions from sequence**, not refolds of designed structures, so absolute confidences are far below initial-guess pipelines and are not comparable to them.
- As classifiers of the experimental `label`, `binder_plddt` (AF2 or AF3) is the most predictive (~0.60-0.68 AUC); the ensemble mean gives a small lift on proteinbase. The uncertainty (`iptm_std`) is itself not a reliable binding classifier.



## Ensemble structures (for Rosetta / energy computation)
The 5 AF3 diffusion structures per design are provided:
- `litscrape_k5_structures_cif.tar.gz` — AF3 native mmCIF
- `litscrape_k5_structures_pdb.tar.gz` — same models as legacy PDB (gemmi-converted; chain **A = target, B = binder**)
- `structures_index.csv` — maps every structure file -> (description, sample_idx) + that sample's AF3 metrics, so externally-computed energies (e.g. Rosetta) can be joined back to the ensemble.

Each tarball extracts to `<description>/<description>_sample-{0..4}.{cif,pdb}`. Structures and the `af3k5_*` / per-sample metrics come from the **same run** (regenerated together with `write_cif_model=true`), so coordinates and scores correspond exactly.

## Columns
- Source metadata + `label` carried over from `yk0/litscrape`.
- `af2_*`: complex_plddt, ptm, iptm, pae, ipae, min_ipae.
- `af3_*`: top-1 init_guess metrics (ptm, iptm, per-chain plddt/pae/ptm/iptm).
- `af3k5_*`: ensemble mean/std/min/max/best for iptm, ptm, binder_plddt, ranking_score; plus k_samples, has_clash_frac.