small_molecule_binding: replace AF2 with Boltz-2 co-folding, port to Delta

[drawadiagram]

Splits the small-molecule workflow out of #59. Third of four.

Based on #60 (needs find_gpus and IMPRESS_SESSION_DIR); GitHub retargets this to main when #60 merges. #63 stacks on this one — the guided-RFD3 fix series is split out separately so this PR stays reviewable.

The cut point is the merge commit where the Boltz-2 migration and the Delta portability work met, before the guided-RFD3 fixes began. Everything here is that first half.

Fold prediction: AlphaFold2 → Boltz-2 co-folding

STEP_AF2 keeps its name for compatibility, but the step now co-folds the protein together with the ligand (supplied as SMILES) and reports ligand_iptm — a protein-ligand interface confidence AF2 could never provide, since it never folded the ligand at all. AF2 here ran in single-sequence mode with zero ligand awareness, so this is a real capability upgrade rather than a swap.

The new fold_min_ligand_iptm gate defaults to None in PROD so the switch doesn't silently tighten production thresholds. scripts/af2.sh → scripts/boltz.sh, which holds the CCD cache lock across a completeness check so concurrent tasks can't read a half-extracted mols/ (the same race later ported to the protein workflow in #61).

Ligand SMILES

A Rosetta .params file contains no SMILES, so scripts/derive_ligand_smiles.py derives one offline via RDKit bond-order perception from the params connectivity plus a reference structure's 3D coordinates. Output committed as p{1..8}_in/ALR.smiles.

RFD3

Drops the scaffoldguided.* CLI argument: RFD3 has no such namespace (unlike older RFDiffusion versions) and guidance lives entirely inside the InputSpecification JSON, so the argument was silently inert. Also sets PYTHONNOUSERSITE and clears PYTHONPATH/PYTHONUSERBASE so the host's ~/.local can't shadow packages inside the container.

⚠️ Worth a close look

filter_shape.py changes the RosettaScripts ScoreTermValueBased score_fxn from sfxn to sfxn_clean. That is a scoring change sitting inside what otherwise reads as a sys.argv → argparse refactor. Calling it out so it doesn't slip through as cleanup.

Configuration and retries

The loose module-level threshold constants become a RunConfig dataclass with PROD and TEST instances selected by IMPRESS_TEST_MODE; backend chosen by IMPRESS_BACKEND (dragon|local); GPUs assigned round-robin from find_gpus().

_stage_metrics_improving() gives the fastrelax and interface stages a metric-agnostic "is this retry actually helping" check, so an attempt whose every still-failing metric stayed flat or regressed short-circuits instead of burning the task budget.

Portability and hygiene

Every scripts/*.sh re-activates $VIRTUAL_ENV instead of a hardcoded /anvil path; mpnn_run.py restores the numpy aliases NumPy 1.24 removed but LigandMPNN's bundled openfold still uses; packmin.py/fastrelax.py use pyrosetta.rosetta.* imports and mute Rosetta's tracer.

Adds delta_env_setup.sh, delta_gpu_run.sh, pull_foundry.sh, and scripts/validate_run.py (post-run output-tree validator).

@AymenFJA — the --mail-user address you flagged on the protein-binding script was still present in this workflow's delta_gpu_run.sh; it's a <your e-mail> placeholder here now. One more remains in examples/discontinuous_scaffolds/rfd3_benchmark/run_islands.slurm:10, pre-existing on main and untouched by any of these four PRs — worth a separate cleanup.

Replaces part of #59.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Lo8DwSbyvdWZRkkkka6gA2