btnaughton/proteinbase-confidence-benchmark

ProteinBase binder/non-binder panels: predicted complexes and confidence scores

632 public ProteinBase designs with measured binding labels, predicted with four structure predictors (ESMFold2, Protenix2, OpenDDE, Mol-1). Designs, sequences and binding labels come from ProteinBase by Adaptyv Bio (ODC-BY license).

Attribution

This work used Proteinbase by Adaptyv Bio under the ODC-BY license. Source: https://proteinbase.com (dataset: https://huggingface.co/datasets/adaptyvbio/proteinbase). The structures, PAE matrices, arrays and confidence scores in this repository are derived from those designs. Please keep this attribution when you reuse them.

Two panels

designs.csv marks which panel each design belongs to:

• in_blog_400: 400 designs over 33 targets, with 200 binders and 200 non-binders (the panel of the blog analysis).
• in_report_450: 450 designs over 24 targets, with 225 binders and 225 non-binders. Every target has equal numbers of binders and non-binders.

218 designs are in both panels. Select a panel by filtering on these two columns. In every structure, chain A is the binder and chain B is the target whose sequence is in targets.csv.

Files

• targets.csv: one row per target (34 targets). Columns: target key, ProteinBase target name and page URL, source of the target sequence, target length, target sequence, zinc and calcium ion counts, and the numbers of binders and non-binders in each panel. The target key joins targets.csv, designs.csv and scores.csv.
• designs.csv: one row per design. Columns: case id, ProteinBase id and URL, target key and target name, binding label, design class and method, author, length, binder sequence, panel flags.
• scores.csv: one row per design, 37 columns.
  • <predictor>__ipsae, __ipdae, __iptm, __actiptm: four confidence scores per predictor.
  • <predictor>__plddt, __pae, __ipae: mean pLDDT, mean PAE and mean interface PAE of the predicted complex.
  • rmsd_<predictor>_<predictor>: binder RMSD between two predictors, in angstroms, after aligning the target C-alpha atoms. Six predictor pairs.
• structures/<predictor>/<case_id>.cif: predicted binder-target complex, one mmCIF file per design and predictor. Chain A is the binder and chain B is the target.
• pae_full/<predictor>/<case_id>.npz: the full predicted aligned error matrix (array pae, in angstroms, float32). Rows and columns are the binder residues, then the target residues, then any ions, in the predictor's output order.
• pae_blocks/<predictor>/<case_id>.npz: the interface blocks of that matrix. Array forward holds binder rows by target columns. Array reverse holds target rows by binder columns.
• actiptm_arrays/<predictor>/<case_id>.npz: the arrays that define actifpTM. pair_tm is the expected TM-score term for each residue pair. contact_probs is the predicted probability that two residues are within 8 angstroms. chain_ids is the chain index of each row. All are float32 or int8.

Recomputing actifpTM

For each row i, weight every residue j in the other chain by contact_probs[i, j]. Normalize: w[i, j] = c[i, j] / (1e-8 + sum over j of c[i, j]). The row score is the sum over j of w[i, j] * pair_tm[i, j]. actifpTM is the maximum row score. Arrays are stored in float32, so recomputed values agree with scores.csv to about 1e-5.