---
license: cc-by-4.0
tags:
- protein
- binder-design
- alphafold2
- alphafold3
configs:
- config_name: default
  data_files: manifest.csv
---

# litscrape: AlphaFold2 + AlphaFold3 metrics

Per-design AlphaFold2 and AlphaFold3 confidence metrics for the binder–target
pairs in [`yk0/litscrape`](https://huggingface.co/datasets/yk0/litscrape).
The source dataset provides **sequences only** (no structures); these metrics
were produced by predicting/scoring each complex from sequence.

**4020 rows**, one per binder–target pair (4020 scored ok).

## How the metrics were computed
1. **Target structure** — each unique target *sequence* was folded once with
   AF2 monomer (ColabDesign hallucination protocol) to obtain a target template.
2. **AF2 metrics** (`af2_*`) — *de novo* AF2-Multimer (ColabDesign binder
   protocol, single model `model_1_multimer_v3`, 3 recycles, no MSA, no initial
   guess): the templated target + the binder *sequence* are folded into a
   complex, and confidence metrics are read out.
3. **AF3 metrics** (`af3_*`) — AF3Score run on the AF2-predicted complex,
   conditioned on those coordinates (`init_guess=true`, no MSA). These are
   AF3's confidence in the AF2 structure, **not** an independent de-novo AF3
   prediction.

## Columns
- Metadata / labels carried over from the source dataset: binder_id, source_publication, target, label, binding_affinity_nm, target_sequence, binder_sequence
- `label` — binary experimental binding label (1 = binder, 0 = non-binder).
- AF2 (`model_1_multimer_v3`): af2_complex_plddt, af2_ptm, af2_iptm, af2_pae, af2_ipae, af2_min_ipae
  (note: `af2_complex_plddt`/`af2_iptm` etc. are on a 0–1 scale; PAE values are
  normalised as in ColabDesign).
- AF3: af3_ptm, af3_iptm, af3_chain_A_plddt, af3_chain_A_pae, af3_chain_A_ptm, af3_chain_A_iptm, af3_chain_B_plddt, af3_chain_B_pae, af3_chain_B_ptm, af3_chain_B_iptm, af3_iptm_A_B
  (note: AF3 pLDDT is 0–100; `af3_chain_B_*` = binder chain; the per-chain
  `iptm` fields equal `af3_iptm` for a 2-chain complex).
- `status`, `error`.

## Important caveats
- These are **de-novo predictions from sequence**, not refolds of designed
  structures, so confidence values are markedly lower than initial-guess
  pipelines (e.g. Proteina-Complexa evals) and are **not** directly comparable
  to them in absolute terms.
- The target template is an AF2-*predicted* monomer, so target-fold error
  propagates into every complex for that target.
- As classifiers of the experimental `label`, the metrics are weak-to-modest
  (binder pLDDT is the most predictive; AUC ~0.60–0.68 on litscrape, lower on
  proteinbase). AF3 closely tracks AF2 and adds little.