Update with GNINA, Karmadock fix, multi chain

[AJPreto]

guild_upgrade — sync from guild-internal

Branch: guild_upgrade. Nothing committed — all changes are uncommitted in the working tree for your review.

This brings the public guild repo up to date with the latest engineering work in the
internal guild-internal repo, without exposing any internal secrets, infrastructure,
database details, or personal paths. The two repos have separate git histories (public
was a sanitized import), so this was done by comparing working trees file-by-file, not by
merging git history.

Guiding rule

Internal-only material was excluded entirely, not just commented out. Where a shared
file mixed good engineering with internal coupling (e.g. bulk.py, the Dockerfile,
the Makefile), the improvements were ported and the internal parts removed — including
indirect tells like internal function names, the DB API shape, internal CI workflows,
private registry URLs, and example compound names.

A full token sweep over the result found zero new internal tokens introduced by this
upgrade (see "Verification" below).

What was brought in (sanitized)

New features

• GNINA docking — guild/docking/gnina.py, guild/constants/gnina.py, plus tests
  (tests/single_run/test_gnina_*.py, tests/bulk_run/test_gnina_orchestration.py).
  The Python code only references container-internal paths (/opt/gnina/...), no secrets.
• GNINA bundle build — Dockerfile.gnina-bundle builds a curated runtime bundle from
  the public gnina/gnina image. The private envedancusacr01.azurecr.io registry
  reference was removed; the bundle image is now an overridable build arg
  GNINA_BUNDLE_IMAGE (default gnina-bundle:latest).
• Multi-chain binding pocket — comma-separated protein_chain (A,B) support, threaded
  through run.py, bulk.py, tools/preparation.py, transformers/pdb.py, and
  scripts/vina_rescore_all.py (multi-chain port applied onto the public, relative-path
  version of that script).
• Per-subprocess troubleshooting logs — guild/tools/subprocess_log.py + integration.
• KarmaDock compatibility patches — scripts/apply_karmadock_patches.py, applied at
  Docker build time (rdkit/torch ABI fixes).
• run_guild.py CLI — gained --methods gnina, --box, --n-workers, --no-gpu,
  --plip-only, --no-plip, --gnina-input-mode. The --no-database-update flag and the
  database_update= wiring were stripped.
• SHELL_SILENCER bugfix — &>/dev/null → > /dev/null 2>&1 (dash-safe).

Updated wholesale (verified clean of internal refs)

guild/run.py, guild/docking/{boltz,diffdock,karmadock,vina}.py,
guild/tools/{preparation,bulk,utils,scores}.py,
guild/transformers/{pdb,msa,converters,chembl}.py,
guild/constants/{bulk,guild,p2rank,visualization}.py, docs/adfrsuite_class.py.

Config / build (sanitized merges)

• Dockerfile — added GNINA bundle stage (genericized) + KarmaDock patches. Kept the
  public repo's clean uv sync (no GH_TOKEN secret) and omitted the Azure CLI block —
  both existed only to serve the excluded enveda-toolkit/keyvault/database layer.
• Makefile — added run-gnina, run-plip, a USE_GPU toggle, gnina/box/workers flags,
  and gnina-bundle auto-build in docker-local. Excluded all azurecr.io push targets,
  the dev/run targets that pulled internal images, the ~/.azure credential mounts, and
  the github-init/init targets (gh repo create enveda/guild, internal remote URL).
• pyproject.toml — version bump 1.0.0 → 1.1.4, added one author. Did not add
  azure-identity, azure-keyvault-secrets, enveda-toolkit, or the [tool.uv.sources]
  git source.
• uv.lock — public lock was already free of internal packages; only the guild version
  pin was updated to 1.1.4 (no dependency changes were made, so the lock stays consistent).
• .gitignore — added safety rules: /notebooks/database, /guild/OLD, /temp_data, and
  a scripts/* allowlist (prevents committing local experiment scripts that carry
  machine-specific paths). Kept the public allowlist for shipped support data + logo.
• .dockerignore — un-ignore scripts/apply_karmadock_patches.py (now COPY'd by Docker).
• .pre-commit-config.yaml — ruff/uv hook version bumps only.

What was deliberately EXCLUDED (internal-only)

• Database / Azure layer: guild/azure/database.py, guild/azure/get_secret.py,
  guild/constants/database.py, DATABASE_SCHEMA.md, notebooks/database/*,
  scripts/test_db_real_submission.py, scripts/test_db_upload_dummy.py. The public repo
  keeps its existing "no database mode". In bulk.py the database_update param, the
  retrieve_existing_rp_scores/upload_rp_scores import and all call sites, and the
  DB-merge/upload code paths were fully removed (not just disabled).
• Internal CI: .github/workflows/deploy-dev.yml, .github/workflows/build-gnina-bundle.yml
  (both use enveda/reusable-workflows + ACR + Prefect). The public test.yml/lint.yml
  were kept as-is.
• Internal tooling/meta: CLAUDE.md (internal dev notes — references ACR, Databricks,
  internal compound names), .cookiecutter.json (internal template + it@envedabio.com),
  .claude/settings.json, .gitmodules (empty), pyptoject_cpu.toml (typo-named CPU
  variant carrying azure/prefect deps), environments/installations.sh (hardcoded
  /home/ec2-user/... paths).
• Azure/Prefect constants in constants/system.py (AZURE_KEYVAULT_URI,
  AZURE_CLIENT_ID, AZURE_ACR_IMAGE_URI, etc.) — used only by excluded code.
• New analysis notebooks / plot scripts / result figures under notebooks/ — these are
  analysis artifacts whose output cells and hardcoded paths (several contained
  /home/antonio.gomes/...) are a high indirect-leak risk for low value. Existing public
  notebooks were left untouched. Worth a selective, manual revisit if you want any of
  these published.

Pre-existing references (NOT introduced by this upgrade — flagged for your call)

• .gitignore:55 → /.databrickscfg — a protective ignore rule (prevents committing a
  Databricks credential file). Left in place; removing it would reduce safety.
• renovate.json:4 → local>enveda/renovate-config — was already in the public repo,
  unchanged. References an internal renovate config repo.

Verification

• All 15 ported modules import cleanly in the repo venv.
• Full internal-token sweep (enveda, azurecr, databricks, keyvault, enveda_toolkit,
  enveda_prod, cdd-benchling, the Azure client GUID, /home/antonio, /home/ec2-user,
  reusable-workflows, prefect, …) over all source/config files → only the two
  pre-existing references above; zero new tokens.
• Docker image rebuild + empirical docking results: see the "Empirical results" section below.

Empirical results

Image rebuilt from scratch with --build-arg GNINA_BUNDLE_IMAGE=gnina-bundle:local
(guild:latest, 38.2 GB). gnina v1.3.1 runs from the genericized local bundle.

Test suite (docker run --rm guild-test python -m pytest tests/ -q): 100 passed.

Empirical docking — single combination 3pbl-A-ETQ-A / lig1 (D3 receptor + a small
ligand), box-driven pocket, --no-decoys, run on an A100 80GB. Every engine produced a
score and completed scoring + PLIP (exit 0):

Engine	Mode	Result
Vina	CPU	vina_score = -4.892
GNINA	GPU	gnina_score = -5.09, gnina_cnn_score = -0.28
Karmadock	GPU	karmadock_score = 13.86 (build-time rdkit/torch patches applied OK)
DiffDock	GPU	diffdock_score = -1.7; auto-rescore vina_rescore_diffdock_score = -0.029
Boltz	GPU	boltz_score = 0.327 (ipTM); auto-rescore vina_rescore_boltz_score = -4.336

Notes:

• The DiffDock→Vina and Boltz→Vina auto-rescore tracks both fired correctly, confirming the
  multi-method scoring wiring survived the database-strip of bulk.py.
• One benign Boltz PLIP warning (Binding site LIG:Z:1 not found) — a known ligand-resname
  quirk, not a failure (run exited 0); unrelated to this upgrade.
• A pandas DataFrameGroupBy.apply FutureWarning appears across runs — pre-existing in
  tools/scores.py, not introduced here.

How to reproduce the build + runs

# Build the gnina bundle once (from the public gnina/gnina image), then the app image:
make docker-local              # auto-builds gnina-bundle:local if missing, then guild:latest

# Empirical docking (workspace holds data + combinations.csv at /workspace):
make run-vina  PROJECT=testvina  COMBINATIONS=/workspace/combinations.csv NO_DECOYS=1
make run-gnina PROJECT=testgnina COMBINATIONS=/workspace/combinations.csv NO_DECOYS=1
make run-guild PROJECT=testall   COMBINATIONS=/workspace/combinations.csv METHODS="diffdock karmadock boltz" NO_DECOYS=1

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.