which PLINDER release matches plinder_valid_ids.txt?

Hi Proteina-Complexa team,

Thank you for releasing the paper/code/assets. I am trying to reproduce the PLINDER subset used by Proteina-Complexa and ran into a dataset version / system ID mismatch.

What I used

I used the public files from this repository:

• assets/data/plinder_valid_ids.txt
• assets/data/plinder_valid_dataset.csv

The ID list has:

rows:        78,368
unique IDs: 67,749

The CSV paths look like this:

protein_path: diffdock_plinder_2404_v0_data/systems/4rek__1__1.A__1.B/receptor.pdb
ligand_paths: diffdock_plinder_2404_v0_data/systems/4rek__1__1.A__1.B/ligand_files/1.B.sdf
path:         ./PFM_data/plinder_valid_systems_cif/4rek__1__1.A__1.B.cif

This suggests that the released list may correspond to a processed diffdock_plinder_2404_v0_data dataset, but I am not sure where that exact dataset/release can be downloaded.

What I tried

I downloaded PLINDER from the public GCS bucket using:

https://storage.googleapis.com/plinder/2024-06/v2/systems/{two_char}.zip

I generated the {two_char} groups from plinder_valid_ids.txt, downloaded all required zip groups, and extracted the systems. I also downloaded the lightweight v2 metadata directories:

2024-06/v2/index/
2024-06/v2/splits/
2024-06/v2/entries/
2024-06/v2/ligands/
2024-06/v2/ligand_scores/
2024-06/v2/dbs/

The download completed successfully.

Mismatch observed

Exact matching against extracted 2024-06/v2/systems gives:

valid rows:          78,368
valid unique IDs:    67,749

exact match rows:    70,327 / 78,368  = 89.74%
exact match unique:  60,910 / 67,749  = 89.91%

missing rows:         8,041
missing unique IDs:   6,839

I also scanned 2024-06/v2/entries/*.zip for systems keys. This did not recover the missing IDs:

entry system IDs unique:        990,269
valid IDs found in entries:      60,898
valid IDs found in systems/entries union: 60,910
missing after union:              6,839

So this does not look like an incomplete download. It looks more like a release/version or system naming difference.

Example missing IDs

A few IDs from plinder_valid_ids.txt that do not exist exactly in 2024-06/v2/systems:

10gs__1__1.B__1.E
12gs__1__1.B__1.E
19hc__1__1.A__1.G
1a0g__1__1.B__1.D
1a0h__1__1.B__1.G
1a0r__1__1.A__1.D
1a3l__1__1.A__1.C
1a4i__1__1.A__1.C
1a4k__2__1.D__1.J
1a7x__1__1.B__1.C

For many missing IDs, same-PDB candidates exist but the system ID naming differs. For example:

missing: 10gs__1__1.B__1.E
candidates in 2024-06/v2:
  10gs__1__1.A_1.B__1.C
  10gs__1__1.A_1.B__1.E

missing: 12gs__1__1.B__1.E
candidates in 2024-06/v2:
  12gs__1__1.A_1.B__1.C
  12gs__1__1.A_1.B__1.E

missing: 1a4k__2__1.D__1.J
candidates in 2024-06/v2:
  1a4k__1__1.A_1.B__1.G
  1a4k__2__1.C_1.D__1.J

I do not want to auto-map these because that may silently change the protein/ligand system.

Questions

1. Which exact PLINDER release/version should be used for assets/data/plinder_valid_ids.txt and assets/data/plinder_valid_dataset.csv?
2. Is diffdock_plinder_2404_v0_data a separate processed PLINDER release? If yes, is it publicly downloadable?
3. Is there a mapping from the released plinder_valid_ids.txt IDs to current PLINDER 2024-06/v2 system IDs?
4. Are the duplicated rows in plinder_valid_ids.txt expected?
5. For reproducibility, should users use only exact matches, or should same-PDB candidate systems be remapped according to some rule?

Any pointer to the intended PLINDER release or processed dataset path would be very helpful.

Thanks!

Start by comparing assets/data/plinder_valid_ids.txt and assets/data/plinder_valid_dataset.csv with the downloaded PLINDER 2024-06/v2 systems, entries, index, and splits archives. Trace the referenced diffdock_plinder_2404_v0_data paths and determine which release, mapping, duplicate-row behavior, and remapping rule reproduce the published subset. Done means documenting the exact source and reproducibility guidance.