different searching reulst by same query in the foldseek search server online

[Huilin-Li]

I used A.pdb as my query and searched in foldseek search server online, I can get a match X.pdb with prob=0.97, evalu=0.0358 months ago.
But today, I re-searched it again with the exactly same A.pdb, I didn't find the match X.pdb anymore.

[martin-steinegger]

Could you please share the PDB files or link to the result?

[Huilin-Li]

Could you please share the PDB files or link to the result?

query protein is Q92508_Cterm0.pdb

My past search result is We4pYiQEtEacSXXbcSZrQFtLPzuUqGkIQ3JpzQFoldseek_2025_09_13_19_44_25.json with 3Di/AA mode.

You can find the match of

"372":[{"query":"job_A","target":"AF-A0A7J4EBV3-F1-model_v4","seqId":7.9,"alnLength":138,"missmatches":75,"gapsopened":8,"qStartPos":2,"qEndPos":128,"dbStartPos":113,"dbEndPos":209,"prob":"0.97","eval":"3.58e-2","score":75,"qLen":222,"dbLen":212,

My latest search result is b2L-8-Bkjzd6jKpeazvNNxPbNzVJfVIMPJq_9g_Foldseek_2025_12_28_14_23_01.json without the match of A0A7J4EBV3

Thank you very much.

[martin-steinegger]

Thank you so much! The database was updated to AFDB v6. Might this explain the difference?

[Huilin-Li]

Thank you so much! The database was updated to AFDB v6. Might this explain the difference?

Hi, thanks, honestly, I don't think so. I saved A0A7J4EBV3_v4.pdb

AF-A0A7J4EBV3-F1-model_v4.pdb

You can see there is no big difference with the latest https://www.uniprot.org/uniparc/UPI0017E25646/entry/A0A7J4EBV3#structure

In my case, I might leave it out, because its z-score of dali-score is below than 2.

But, still curious to know if there is a kind of random seed thing in foldseek model? I think it is very important to repeat my research conclusion somehow.

I was thinking if I need to maintain a local database. However, 1) foldseek easy-search doesn't have a random seed-like parameter to set up. I might also have a slight different results even I run the same command line. 2) My local database will be outdated. Others who want to repeat my conclusion cannot have a totally same database. 3) This might be the common problem in the methods using heuristics. Higher thresholds might help, but not sure if its enough for remote homologues detection.

In my another case, in foldseek result, its prob=1, eval=0.0514, but its z-score to the query is 11 which is definitely homologues protein.

[martin-steinegger]

I cannot find https://alphafold.ebi.ac.uk/search/text/A0A7J4EBV3 on the AFDB website. I assume it is not in v6 anymore.

[Huilin-Li]

I cannot find https://alphafold.ebi.ac.uk/search/text/A0A7J4EBV3 on the AFDB website. I assume it is not in v6 anymore.

Hi, if you click the download logo, you could find the .cif version of this protein.