Induced-fit benchmark: Stage 1 data
The benchmark asks whether a protein's binding pocket changes shape with the ligand. For each target it compares crystal structures with different ligands (signal) against structures with the same ligand (noise floor), and asks whether the signal clearly exceeds the floor.
This dataset holds the data behind Stage 1. The code and the reports are in the project's GitHub repository (branch
stage1-full); start with reports/stage1_report.md there. Access is gated: request access on the dataset page
(approved automatically). The results are not yet published.
Layout
results/<folder>/ full pipeline output of each of the 25 analysable targets (50 files per folder)
structures/ the 3,507 PDBe updated mmCIF files (<id>_updated.cif.gz) used by those runs
structures/index.csv target folder -> PDB ID -> file
gallery_sessions/ 26 PyMOL sessions S1_<target>.pse (25 targets + the DOT1L noise control)
| Folder | What | Size |
|---|---|---|
results/ |
pipeline output, including the .npy matrices |
≈ 4.1 GB |
structures/ |
mmCIF files + index.csv |
≈ 650 MB |
gallery_sessions/ |
PyMOL sessions of the Stage 1 example pictures (open in PyMOL; with the repository: uv sync --extra viz); the PNGs and notes are in the repository, reports/gallery/ |
≈ 90 MB |
The 25 result folders
Folder prefixes: EXP_* = LIT-PCBA targets expanded to all PDB entries, except EXP_ABL1 = ABL1 kinase
reference; LIT_* = literature set; KREF_* = EGFR / MK14 kinase references. Units, entries and final class are
from reports/tables/stage1_verdicts.csv in the repository; the number of entries equals the folder's rows in
structures/index.csv.
| Folder | Target | Family | UniProt | Set | Units / entries | Final class |
|---|---|---|---|---|---|---|
EXP_ABL1 |
ABL1 | protein kinase | P00519 | kinase reference | 90 / 45 | clear pass |
EXP_ADRB2 |
ADRB2 | GPCR | P07550 | LIT-PCBA | 40 / 37 | not above floor |
EXP_ESR1 |
ESR1 | nuclear receptor | P03372 | LIT-PCBA | 778 / 347 | pass, flagged |
EXP_GBA |
GBA | glycosidase | P04062 | LIT-PCBA | 60 / 31 | not above floor |
EXP_IDH1 |
IDH1 | oxidoreductase | O75874 | LIT-PCBA | 61 / 24 | clear pass |
EXP_IDH1_site2 |
IDH1_site2 | oxidoreductase | O75874 | LIT-PCBA | 45 / 15 | pass, flagged |
EXP_MAPK1 |
MAPK1 | protein kinase | P28482 | LIT-PCBA | 156 / 142 | clear pass |
EXP_PKM2 |
PKM2 | metabolic kinase (pyruvate kinase) | P14618 | LIT-PCBA | 47 / 12 | not above floor |
EXP_PPARG |
PPARG | nuclear receptor | P37231 | LIT-PCBA | 486 / 332 | not above floor |
EXP_TP53 |
TP53 | transcription factor (p53 core domain) | P04637 | LIT-PCBA | 119 / 62 | not above floor |
EXP_VDR |
VDR | nuclear receptor | P11473 | LIT-PCBA | 48 / 48 | not above floor |
KREF_EGFR |
EGFR | protein kinase | P00533 | kinase reference | 489 / 304 | pass, flagged |
KREF_MK14 |
MK14 | protein kinase | Q16539 | kinase reference | 240 / 222 | clear pass |
LIT_ACHE_MOUSE |
ACHE_MOUSE | serine hydrolase | P21836 | literature | 164 / 79 | not above floor |
LIT_ADORA2A |
ADORA2A | GPCR | P29274 | literature | 173 / 169 | pass, flagged |
LIT_AKR1B1 |
AKR1B1 | oxidoreductase | P15121 | literature | 156 / 153 | not above floor |
LIT_BACE1 |
BACE1 | protease | P56817 | literature | 650 / 402 | clear pass |
LIT_DOT1L |
DOT1L | epigenetic | Q8TEK3 | literature | 45 / 32 | clear pass |
LIT_GRIA2 |
GRIA2 | ion channel LBD | P19491 | literature | 362 / 154 | clear pass |
LIT_HSP90A |
HSP90A | chaperone ATPase | P07900 | literature | 384 / 334 | not above floor |
LIT_INHA |
INHA | oxidoreductase | P9WGR1 | literature | 324 / 109 | clear pass |
LIT_KRAS |
KRAS | small GTPase | P01116 | literature | 543 / 303 | pass, flagged |
LIT_MCL1 |
MCL1 | PPI groove | Q07820 | literature | 172 / 75 | pass, flagged |
LIT_PTP1B |
PTP1B | phosphatase | P18031 | literature | 48 / 47 | not above floor |
LIT_RORC |
RORC | nuclear receptor | P51449 | literature | 34 / 30 | not above floor |
structures/index.csv has 3,508 rows for 3,507 files: entry 9XZ1 is used by two targets (LIT_KRAS and LIT_HSP90A).
Files in each result folder
A unit is one protein chain with its one site ligand. Frames: core (superposed on the stable core) and
pocket (superposed on the pocket). Cutoffs: 4, 5, 6 Ã…. Subsets: all atoms, ca, sidechain.
| File | What |
|---|---|
units.csv |
one row per unit (PDB ID, chain, ligand, scaffold, resolution, space group, covariates); row i = row/column i of every N×N matrix |
pairs.csv |
one row per unit pair: i, j, key_i, key_j, pair_type. Noise: same_entry_diff_chain, same_ligand_diff_entry. Signal: diff_scaffold. same_scaffold_diff_ligand pairs (close analogues) are neither noise nor signal and are not used in S/F |
pairwise_rmsd_<core|pocket>_<all|ca|sidechain>_<4|5|6>.npy |
N×N pocket RMSD between units, Å |
residue_pair_<core|pocket>_<4|5|6>.npy |
(R, N, N) float32, per-residue displacement between units, Ã…; residue r = row r of per_residue_<frame>_<cut>.csv |
per_residue_<core|pocket>_<4|5|6>.csv |
per pocket residue (UniProt number): floor p95, signal p90, excess, moved call, and more |
pocket_{union,intersection}_<4|5|6>.json |
pocket residues (UniProt numbers) |
core.json |
the stable core and its convergence |
summary.json |
counts, S/F and moved residues per frame and cutoff |
chi1.csv, chi1_stats.csv |
side-chain χ1 per unit and residue; circular statistics |
volume.csv, volume_summary.json |
pyKVFinder pocket volume per unit |
contacts.csv, contacts_summary.csv, contacts_summary.json |
PLIP protein-ligand contacts |
disorder_5.csv |
missing pocket residues and ligand-linked disorder (5 Ã… pocket) |
run.log |
the run log; every dropped chain, entry or ligand is a DROP line with its reason |
command.txt |
the exact command, the git commit and the start time |
Exact definitions: docs/PIPELINE_SPEC.md ("Scripts and outputs") in the repository.
Provenance
command.txtrecords git commitbf20d54(EXP_ABL1, EXP_MAPK1, EXP_ESR1: launched from the proof-of-concept code as part of the pre-meeting extras, commitb790736) or1ae4fb5(the other 22 folders).1ae4fb5is not in the repository history (the history was rewritten). The core pipeline code (pipeline.py,run_target.py,align,fetch,structure,ligands,pocket,metrics,noise,subsample) is the same in commitb790736and on branchstage1-full(only a docstring and help-text examples differ), so usestage1-full.- Paths in
command.txtare cluster paths.
Not included: the non-analysable runs, the run logs, and the PoC and apo test runs.
Download
uvx --from huggingface_hub hf auth login # once, with a token of an account that has access
# everything (≈ 4.8 GB), from the repository root
uvx --from huggingface_hub hf download Honzus24/inducedfit-stage1 --repo-type dataset --local-dir data/hf
# one folder only
uvx --from huggingface_hub hf download Honzus24/inducedfit-stage1 --repo-type dataset --local-dir data/hf \
--include "results/LIT_RORC/*"
# without the matrices
uvx --from huggingface_hub hf download Honzus24/inducedfit-stage1 --repo-type dataset --local-dir data/hf \
--exclude "*.npy"
Use with the repository
# from the repository root
# scripts that need the matrices, e.g.
uv run python scripts/stage1_supplement.py --results data/hf/results
# re-running a target offline from the stored structures (other options as in its command.txt)
IDS=$(awk -F, '$1=="LIT_RORC"{print $2}' data/hf/structures/index.csv | tr '\n' ' ')
uv run python scripts/run_target.py --name LIT_RORC --uniprot P51449 --pdb-ids $IDS --ref-ligand 5C4O:4F1 \
--cutoffs 4 5 6 --cache-dir data/hf/structures --out-dir results
Re-running LIT_RORC from these structures alone gave a summary.json identical to the stored one.
Sources and licence
The mmCIF files are PDBe "updated" mmCIF files of PDB entries (wwPDB data, CC0). The result files are derived analyses of the project; not yet published.
- Downloads last month
- 48