Delete README_SERVER.md

README_SERVER.md

@@ -1,505 +0,0 @@
-# Portable rDock Pipeline
-
-## List known-good PDB complexes
-```bash
-python scripts/prepare_pdb_ligand_dataset.py --list-known-good
-```
-
-## Environment and rDock installation
-The portable bundle no longer ships `rdock/` binaries.
-The runtime is rDock-only. Removed secondary and legacy backends are not part of this portable pack.
-
-`setup_server.sh` creates or updates the conda environment with:
-- `conda-forge` + `bioconda`
-- `--strict-channel-priority`
-- platform-independent Python/runtime packages:
-  - `python=3.11`
-  - `numpy pandas scipy scikit-learn matplotlib pyyaml tqdm joblib biopython pytest requests`
-  - `rdkit`
-  - `openbabel`
-  - `plip`
-
-PLIP is part of the portable environment. rDock still produces the poses and native
-SCORE terms; PLIP is used after docking to count biological interaction classes
-such as hydrophobic contacts, hydrogen bonds, salt bridges, pi-stacking and related
-contacts. If PLIP fails for a specific pose, the row records the failure and falls
-back to the rDock energy/pose interaction proxy instead of silently dropping the
-ligand.
-
-For rDock itself, the setup script first tries the conda/bioconda `rdock=24.04.204_legacy`
-package. That package is available on common Linux server platforms but may not exist for
-every local development platform, including some Apple Silicon setups. If conda cannot
-install rDock, the script falls back to an existing `rbdock`/`rbcavity` installation on
-`PATH` and validates the detected `RBT_ROOT`.
-
-On Apple Silicon, a typical local setup is:
-
-```bash
-brew install rdock
-```
-
-On Linux x86_64 servers, use `setup_server.sh`; it will install the conda package when
-available through the configured channels.
-
-Setup exports:
-- `RDOCK_ROOT=$CONDA_PREFIX`
-- `RBT_ROOT=$CONDA_PREFIX/share/rdock` or `$CONDA_PREFIX/share` when that layout is used by the installed package
-- `RBT_HOME=$RBT_ROOT`
-- `PATH=$CONDA_PREFIX/bin:$PATH`
-- `LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH`
-
-`check_environment.sh` hard-fails if any required dependency is missing, including:
-- `rdkit`
-- `rdkit.Chem`
-- `rdkit.Chem.AllChem`
-- `rdkit.Chem.Descriptors`
-- `rdkit.Chem.rdMolDescriptors`
-- `plip`
-- `plip.structure.preparation`
-- `obabel`
-- `rbdock`
-- `rbcavity`
-- `RBT_ROOT`
-- `dock.prm`
-
-so the pipeline code does not need to be patched on the server.
-
-## Setup on server
-```bash
-bash setup_server.sh portable-rdock-pipeline
-conda activate portable-rdock-pipeline
-bash check_environment.sh
-```
-
-## Inspect hetero ligands for a PDB
-```bash
-python scripts/prepare_pdb_ligand_dataset.py \
-  --list-hetero \
-  --pdb-id 1A52
-```
-
-For `1A52`, the parser should confirm that the correct reference ligand is `EST`, not `MK1`. `AU` is treated as an ion/metal and ignored as a docking reference ligand.
-
-## Prepare a dataset for one PDB complex
-```bash
-python scripts/prepare_pdb_ligand_dataset.py \
-  --pdb-id 4HG7 \
-  --receptor-chain A \
-  --reference-ligand-resname NUT \
-  --reference-ligand-chain A \
-  --n-ligands 1000 \
-  --ligand-source smiles_file \
-  --smiles-file data/examples/example_smiles_1000.smi \
-  --out /data/datasets/rdock_4hg7_1000 \
-  --force
-```
-
-## Prepare a dataset by fetching real similar SMILES from PubChem
-```bash
-python scripts/prepare_pdb_ligand_dataset.py \
-  --pdb-id 4HG7 \
-  --receptor-chain A \
-  --reference-ligand-resname NUT \
-  --reference-ligand-chain A \
-  --n-ligands 1000 \
-  --ligand-source pubchem_similarity \
-  --identity-threshold-start 99 \
-  --identity-threshold-stop 85 \
-  --allow-partial-ligand-set \
-  --out /data/datasets/rdock_4hg7_pubchem \
-  --force
-```
-
-This fetches real PubChem compounds via the official PUG REST `fastsimilarity_2d` endpoint. If PubChem returns fewer than the requested count within the threshold range, the script keeps the available ligands, writes them into the dataset, and records a warning in `dataset_manifest.json`, `qc/preparation_report.md`, and `logs/pubchem_diagnostics.json`.
-
-You can also request larger libraries, for example `25000`, when `--ligand-source pubchem_similarity` is set explicitly. In that case the script still keeps a partial real result set if PubChem does not return enough compounds within the threshold range.
-
-## Auto-detect the reference ligand
-```bash
-python scripts/prepare_pdb_ligand_dataset.py \
-  --pdb-id 1A52 \
-  --receptor-chain A \
-  --auto-reference-ligand \
-  --n-ligands 1000 \
-  --ligand-source smiles_file \
-  --smiles-file data/examples/example_smiles_1000.smi \
-  --out /data/datasets/rdock_1a52_1000 \
-  --force
-```
-
-## Validate an existing dataset directory
-```bash
-python scripts/prepare_pdb_ligand_dataset.py \
-  --validate-only \
-  --out /data/datasets/rdock_4hg7_1000
-```
-
-or
-
-```bash
-python -m docking_pipeline validate-dataset \
-  --dataset-dir /data/datasets/rdock_4hg7_1000
-```
-
-## Resume ligand 3D preparation from an interrupted dataset
-If `ligands/all_ligands.smi` already exists and the build stopped during OpenBabel 3D conversion, resume only the ligand preparation step:
-
-```bash
-python scripts/prepare_pdb_ligand_dataset.py \
-  --prepare-ligands-only \
-  --out /data/datasets/rdock_4hg7_pubchem_50k \
-  --ligand-batch-size 250 \
-  --ligand-jobs auto \
-  --ligand-cpu-fraction 0.85
-```
-
-This mode:
-- uses the existing `ligands/all_ligands.smi`
-- writes batch outputs under `ligands/obabel_batches/`
-- updates `ligands/ligand_preparation_progress.json`
-- skips completed batches on rerun
-- if only a partial `ligands/all_ligands.sdf` exists and there are no batch files yet, it resumes from the prepared prefix and builds the missing suffix
-
-Check progress with:
-
-```bash
-cat /data/datasets/rdock_4hg7_pubchem_50k/ligands/ligand_preparation_progress.json
-```
-
-## Prepare a real 1000-SMILES example file
-```bash
-python scripts/download_example_smiles.py \
-  --n 1000 \
-  --out data/examples/example_smiles_1000.smi
-```
-
-This file contains real compounds derived from the benchmark ligand library already present in this repository. It is not a synthetic or random SMILES set.
-
-## Large libraries
-For large libraries such as `50000` ligands, use a real SMILES file:
-
-```bash
-python scripts/prepare_pdb_ligand_dataset.py \
-  --pdb-id 1A52 \
-  --receptor-chain A \
-  --auto-reference-ligand \
-  --n-ligands 50000 \
-  --ligand-source smiles_file \
-  --smiles-file /data/libraries/real_50k_library.smi \
-  --out /data/datasets/rdock_1a52_50k \
-  --force
-```
-
-`PubChem similarity` may fail to collect 1000 analogs for some reference ligands. When that happens, the generator writes `logs/pubchem_diagnostics.json` and you should switch to `--ligand-source smiles_file`.
-
-## Run multi-fidelity docking
-```bash
-python -m docking_pipeline benchmark-adaptive \
-  --dataset-dir /data/datasets/rdock_1a52_50k \
-  --strategy multifidelity_adaptive_rdock \
-  --fidelity-levels 5,10,15,30,50 \
-  --cost-budget-runs 250000 \
-  --jobs auto \
-  --cpu-fraction 0.85 \
-  --out /data/results/rdock_1a52_50k_mf \
-  --resume
-```
-
-## Production screening without a reference ligand
-```bash
-python -m docking_pipeline screen-reference-free \
-  --dataset-dir /data/datasets/my_prepared_pocket_dataset \
-  --strategy reference_free_active_learning_v2 \
-  --triage-model classifier \
-  --triage-controller auto_recall \
-  --triage-retain-fraction 0.02 \
-  --triage-target-recall 0.98 \
-  --classifier-top-percentile 0.05 \
-  --classifier-threshold-mode recall_target \
-  --model-fallback-if-worse union_with_cluster_only \
-  --survivor-combination-policy union \
-  --max-retain-fraction-before-not-useful 0.5 \
-  --calibration-size 1000 \
-  --fidelity-validation-size 200 \
-  --fidelity-levels 5,10,15,30,50 \
-  --jobs 64 \
-  --chunk-size 50 \
-  --cpu-fraction 0.85 \
-  --out /data/results/my_screen \
-  --resume
-```
-
-This mode:
-- does not require `reference_ligand.sdf`
-- treats any reference ligand present in the dataset as optional benchmark-only metadata
-- fails early with `RDKit_REQUIRED_FOR_REFERENCE_FREE_MODEL` if RDKit is unavailable
-- writes triage artifacts:
-  - `tables/initial_prefilter_decisions.csv`
-  - `tables/triage_scores.csv`
-  - `tables/triage_survivors.csv`
-  - `tables/triage_rejected.csv`
-  - `metrics/triage_metrics.json`
-  - `metrics/fidelity_reliability.json`
-
-## Compare strategies on one shared sampled reference
-```bash
-python -m docking_pipeline benchmark-triage-strategies \
-  --dataset-dir /data/datasets/test_1500 \
-  --reference-mode sampled \
-  --reference-sample-size 300 \
-  --reference-sample-seed 42 \
-  --strategies cluster_only_triage,cheap_descriptor_filter_only,diverse_random_cost_balanced,single_fidelity_cost_balanced,reference_free_triage_bandit_v1,reference_free_active_learning_v2 \
-  --triage-target-recall 0.98 \
-  --triage-retain-fraction 0.02 \
-  --calibration-size 300 \
-  --fidelity-validation-size 100 \
-  --cost-budget-runs 3000 \
-  --seeds 1,2,3 \
-  --jobs 10 \
-  --chunk-size 20 \
-  --rdock-timeout-seconds 14400 \
-  --out /data/results/test_1500_strategy_comparison \
-  --resume
-```
-
-## Small benchmark with full reference and the new reference-free strategy
-```bash
-python -m docking_pipeline benchmark-adaptive \
-  --dataset-dir /data/datasets/test_1500 \
-  --strategy reference_free_active_learning_v2 \
-  --reference-mode full \
-  --evaluation-pool-mode same_pool \
-  --triage-model classifier \
-  --triage-controller auto_recall \
-  --triage-retain-fraction 0.05 \
-  --triage-target-recall 0.98 \
-  --calibration-size 300 \
-  --fidelity-validation-size 100 \
-  --balanced-baselines true \
-  --jobs 16 \
-  --chunk-size 25 \
-  --out /data/results/test_1500_ref_free_strategy \
-  --resume
-```
-
-## Large 50k sampled benchmark with the new reference-free strategy
-```bash
-python -m docking_pipeline benchmark-adaptive \
-  --dataset-dir /data/datasets/rdock_50k \
-  --strategy reference_free_active_learning_v2 \
-  --reference-mode sampled \
-  --reference-sample-size 5000 \
-  --reference-sample-seed 42 \
-  --evaluation-pool-mode same_pool \
-  --triage-model classifier \
-  --triage-controller auto_recall \
-  --triage-retain-fraction 0.02 \
-  --triage-target-recall 0.98 \
-  --calibration-size 1500 \
-  --fidelity-validation-size 300 \
-  --balanced-baselines true \
-  --jobs 64 \
-  --chunk-size 50 \
-  --cpu-fraction 0.85 \
-  --out /data/results/rdock_50k_ref_free_strategy \
-  --resume
-```
-
-## Repeated triage safety benchmark
-```bash
-python -m docking_pipeline benchmark-triage-repeated \
-  --dataset-dir /data/datasets/test_1500 \
-  --reference-mode full \
-  --strategy reference_free_triage_bandit_v1 \
-  --baselines cluster_only_triage,cheap_descriptor_filter_only,diverse_random_cost_balanced \
-  --triage-model classifier \
-  --classifier-top-percentile 0.05 \
-  --classifier-threshold-mode recall_target \
-  --triage-controller auto_recall \
-  --model-fallback-if-worse union_with_cluster_only \
-  --survivor-combination-policy union \
-  --triage-target-recall 0.98 \
-  --triage-retain-fraction 0.02 \
-  --calibration-size 300 \
-  --fidelity-validation-size 100 \
-  --seeds 1,2,3 \
-  --jobs 8 \
-  --chunk-size 20 \
-  --cpu-fraction 0.85 \
-  --out /data/results/test_1500_triage_repeated \
-  --force
-```
-
-## Results layout
-- `target/`, `ligands/`, `rdock/`, `poses/`, `tables/`, `metrics/`, `plots/`, `checkpoints/`
-- `report.md`, `manifest.json`, `config.yaml`, `commands.log`
-
-## Candidate search for a 500k run
-```bash
-python scripts/find_pdb_ligand_candidates.py \
-  --out configs/benchmarks/pdb_ligand_candidates_500k.tsv \
-  --max-candidates 50 \
-  --min-heavy-atoms 15 \
-  --max-heavy-atoms 60 \
-  --min-estimated-hits 500000
-```
-
-## Prepare a 500k dataset
-```bash
-export PDB_ID=<BEST_PDB>
-export RECEPTOR_CHAIN=<CHAIN>
-export REF_LIGAND=<RESNAME>
-export REF_CHAIN=<CHAIN>
-export DATASET_DIR=/data/datasets/rdock_${PDB_ID}_pubchem_500k
-
-python scripts/prepare_pdb_ligand_dataset.py \
-  --pdb-id $PDB_ID \
-  --receptor-chain $RECEPTOR_CHAIN \
-  --reference-ligand-resname $REF_LIGAND \
-  --reference-ligand-chain $REF_CHAIN \
-  --n-ligands 500000 \
-  --ligand-source pubchem_similarity \
-  --pubchem-threshold-ladder 99,95,90,85,80,75,70,65,60,55,50,45,40,35,30 \
-  --allow-partial-ligand-set \
-  --ligand-batch-size 100 \
-  --ligand-jobs 60 \
-  --ligand-cpu-fraction 0.85 \
-  --out $DATASET_DIR \
-  --force
-```
-
-## Validate the prepared dataset
-```bash
-python -m docking_pipeline validate-dataset \
-  --dataset-dir $DATASET_DIR
-```
-
-## Smoke ablation benchmark
-```bash
-export SMOKE_OUT=/data/results/${PDB_ID}_500k_smoke_ablation
-
-python -m docking_pipeline benchmark-regressor-ablation \
-  --dataset-dir $DATASET_DIR \
-  --reference-mode sampled \
-  --reference-sample-size 1000 \
-  --reference-sample-seed 42 \
-  --fidelity-levels 5,10,15,30,50 \
-  --cost-budget-runs 3000 \
-  --calibration-size 500 \
-  --fidelity-validation-size 150 \
-  --seeds 1 \
-  --jobs 32 \
-  --chunk-size 50 \
-  --cpu-fraction 0.85 \
-  --out $SMOKE_OUT \
-  --resume
-```
-
-## Full regressor ablation benchmark
-```bash
-export ABLATION_OUT=/data/results/${PDB_ID}_500k_regressor_ablation
-
-python -m docking_pipeline benchmark-regressor-ablation \
-  --dataset-dir $DATASET_DIR \
-  --reference-mode sampled \
-  --reference-sample-size 5000 \
-  --reference-sample-seed 42 \
-  --fidelity-levels 5,10,15,30,50 \
-  --cost-budget-runs 200000 \
-  --calibration-size 5000 \
-  --fidelity-validation-size 1000 \
-  --seeds 1,2,3 \
-  --jobs 64 \
-  --chunk-size 100 \
-  --cpu-fraction 0.85 \
-  --rdock-timeout-seconds 14400 \
-  --out $ABLATION_OUT \
-  --resume
-```
-
-## Full strategy benchmark
-```bash
-export STRATEGY_OUT=/data/results/${PDB_ID}_500k_strategy_benchmark
-
-python -m docking_pipeline benchmark-triage-strategies \
-  --dataset-dir $DATASET_DIR \
-  --reference-mode sampled \
-  --reference-sample-size 5000 \
-  --reference-sample-seed 42 \
-  --strategies cluster_only_triage,cheap_descriptor_filter_only,random_cost_balanced,diverse_random_cost_balanced,single_fidelity_cost_balanced,reference_free_active_learning_v2 \
-  --adaptive-policies cluster_bandit \
-  --triage-target-recall 0.98 \
-  --triage-retain-fraction 0.02 \
-  --calibration-size 5000 \
-  --fidelity-validation-size 1000 \
-  --cost-budget-runs 200000 \
-  --fidelity-levels 5,10,15,30,50 \
-  --seeds 1,2,3 \
-  --jobs 64 \
-  --chunk-size 100 \
-  --cpu-fraction 0.85 \
-  --rdock-timeout-seconds 14400 \
-  --out $STRATEGY_OUT \
-  --resume
-```
-
-## Comparable adaptive benchmark
-Use this when you need a frozen evaluation universe and cost-comparable reporting.
-```bash
-export COMPARABLE_OUT=/data/results/${PDB_ID}_500k_comparable_benchmark
-
-python -m docking_pipeline benchmark-comparable-adaptive \
-  --dataset-dir $DATASET_DIR \
-  --evaluation-universe-size 25000 \
-  --evaluation-universe-seed 42 \
-  --reference-mode full \
-  --strategies adaptive_classifier_regressor,classifier_only,diverse_random_cost_balanced,cluster_only,single_fidelity_cost_balanced \
-  --cost-budget-runs 100000 \
-  --fidelity-levels 5,10,15,30,50 \
-  --promotion-policy quota_ladder \
-  --min-final-ligands 200 \
-  --model-validation-split cluster \
-  --jobs 120 \
-  --chunk-size 10 \
-  --cpu-fraction 0.85 \
-  --rdock-timeout-seconds 1800 \
-  --out $COMPARABLE_OUT \
-  --resume
-```
-
-Prosty wrapper shell:
-```bash
-export DATASET_DIR=$SCRATCH/plgproma/data/rdock_${PDB_ID}_pubchem_500k
-export STRATEGY_OUT=$SCRATCH/plgproma/results/${PDB_ID}_500k_strategy_benchmark
-export JOBS=120
-export CHUNK_SIZE=30
-export CALIBRATION_SIZE=50000
-export FIDELITY_VALIDATION_SIZE=50
-export COST_BUDGET_RUNS=100000
-export REFERENCE_SAMPLE_SIZE=500
-export RDOCK_TIMEOUT_SECONDS=4000
-
-bash scripts/run_strategy_benchmark_server.sh
-```
-
-## Audits
-```bash
-python -m docking_pipeline audit-strategy-benchmark \
-  --run-dir $STRATEGY_OUT
-
-python -m docking_pipeline audit-regressor-ablation \
-  --run-dir $ABLATION_OUT
-```
-
-## Resume
-- Re-run the same `benchmark-adaptive` command with `--resume`.
-- Checkpoints are written under `checkpoints/`.
-
-## Output interpretation
-- `full_docking_scores.csv`: exhaustive final-fidelity docking reference.
-- `random_baseline_scores.csv`: random baseline at matched cost budget.
-- `single_fidelity_adaptive_scores.csv`: direct final-fidelity model-priority comparator.
-- `multifidelity_trace.csv`: per-ligand trace across fidelity levels.
-- `promotion_decisions.csv`: promotion and rejection reasons.
-- `final_hits.csv`: hit table after multi-fidelity promotion logic.
-- `adaptive_benchmark_metrics.json`: aggregate benchmark metrics and cost accounting.