yifanyang993/PLMCaliper-extras

PLM-Caliper — archived experiment arms

Cold storage for parts of the PLM-Caliper study that are rarely needed for reproducing the headline results. Everything here uses the same layout, naming and column format as the main dataset — https://huggingface.co/datasets/yifanyang993/PLMCaliper — so a path from one repo is a path in the other.

Total 4144 files, 193 GB.

What is archived here

arm	files	size	what it is
PGmsk25pct	87	154.6 GB	ProteinGenerator 25 %-masked decoy — one arm of the decoy-construction ablation
extended_shuf (ASTRAL)	82	38.1 GB	shuffle-extended decoy on the main ASTRAL experiment, the comparison arm to extended_mkv2
bagel	51	0.1 GB	BAGEL bacteriocin search (class_all, class_1)
putative	3924	0.2 GB	putative-bacteriocin queries and the discovery tables/figures built from them

By tier: raw/ 66 files / 100.5 GB · intermediate/ 40 files / 92.1 GB · plot/ 2784 files / 0.3 GB · figures/ 1254 files / 0.05 GB.

raw/
  sequences/                              25 files    34 MB
  scores/ablation_tau0.25/{blastp,dhr,plm,tmvec}/       PGmsk25pct arm, 73.3 GB
  scores/astral_main/                      4 files    27.0 GB   extended_shuf arm
  scores/bagel/                           16 files    96.7 MB
  scores/search_data/                     14 files    19.5 MB   putative → class_all
intermediate/
  calibrated_decoy/  ablation_tau0.25/ (3 × ~13 GB), astral_main/ (15 files, 21.9 GB), search_data/
  target_noisy/      ablation_tau0.25/ (3 × ~10 GB), astral_main/, search_data/
plot/
  bagel/, bagel_discovery/, core/         2784 files   0.3 GB
figures/
  bagel/, core/                           1254 PDFs    50 MB

Format

Identical to the main dataset. Score tables are result_<method>_<query>_<db>.txt[.gz], tab-separated, columns qid tid score homo_type rank (+ evalue for blastp, + rep_id for calibrated / target-noisy tables). homo_type is the SCOP level shared by query and target: 2 same family, 1 same superfamily, 0 same fold, -1 unrelated; the BAGEL tables use a binary same-class label (1 / -1).

What is not here

extended_shuf is the paper's default decoy, so it was archived only where an alternative decoy remains in the main dataset. These stay in the main repo, because moving them would split a target/decoy pair across two repos:

• raw/scores/ur50/result_*_astral4f_extended_shuf_ur50.txt — the only decoy of the UniRef50 / ColabFold experiment (12.5 GB)
• raw/scores/search_data/result_*_astral_s3000_seed123_extended_shuf_*.txt — decoy side of the s3000 subset (1.0 GB)
• raw/sequences/{astral4f,astral_s3000_seed123,astral_filtered,casp13,astral_tmvec_fail}_extended_shuf.*

Sequence files shared with retained experiments are duplicated rather than moved, so each repo stands on its own.

Code: https://github.com/batmen-lab/PLMCaliper