CedLJH/rna-sbdd-v2

RNA-SBDD v2

A frozen benchmark for RNA structure-based drug design: 8,006 RNA pocket-ligand complexes derived from RCSB, with a sequence-identity-disjoint split, the evaluation artifacts, and the trained checkpoints the benchmark's numbers come from.

This repository exists because the cluster the work ran on was retired. It is a complete handoff — dataset, artifacts, weights, and the tooling to bring all of it up somewhere else.

Code: git@Ced3-han:Ced3-han/RNASBDD.git, branch iclr/rna-sbdd-integration.

The step-by-step operational guide is PORTABLE_TRAINING.md in that repository. It is 480 lines, it was written for exactly this move, and it has been executed end to end. This card covers what it does not: restoring this revision, remapping paths, and what to do next. Read §1–§3 here, then follow PORTABLE_TRAINING.md for environment and training detail.

Two places PORTABLE_TRAINING.md is stale — it predates the merged split:

• It pins HF_DATASET_REVISION=0fda7601… (2026-08-05), which contains only the dataset archive. Use c497759da15716c33449fa1082ac1202badeb50f or later to get the checkpoints, artifacts, and scripts.
• It says nothing about the merged split or the merged size prior. Where it describes training or sampling data, apply §2 and §4 below instead.

---

Contents

Path	Size	What it is
dataset/*.tar.zst	486 MB	Frozen release: 64,070 files, 8,006 samples — raw structures, processed LMDB, splits, full build provenance
dataset/SHA256SUMS, *.bundle.json	—	Archive digest and packaging record
benchmark_artifacts/	5.3 MB	Merged split, molecule-size priors, evaluation references, statistics
checkpoints/bound/	0.73 GB	The 20 checkpoints that specs and configs bind by SHA-256
checkpoints/merged-wave-terminal/	0.21 GB	6 checkpoints from a cancelled run — see §5
checkpoints/MANIFEST.json	—	Each checkpoint's digest, original path, and the specs that bind it
scripts/	—	Restore, path-remap, and config-localization tooling

The archive expands to ~13 GB, of which 8.7 GB is manifests/candidate_manifest.jsonl — every candidate considered before filtering. Needed to re-derive the freeze, not to use the dataset.

A duplicate copy of the archive sits at the repository root from an earlier upload. dataset/ is canonical.

---

The split

• train 7,269 · train-contained monitor 713 (a subset of train) · test 713 · 24 excluded · 8,006 total
• MMseqs2 sequence identity: min-seq-id 0.90, coverage 0.80, cov-mode 1
• The 713 test complexes are frozen and were never changed.
• There is no held-out validation. The former validation IDs survive only as a monitor living inside train, and it is never called validation. Selecting a checkpoint on it is selecting on training loss.

---

1. Restore

pip install -U huggingface_hub
hf auth login

git clone git@Ced3-han:Ced3-han/RNASBDD.git
cd RNASBDD && git checkout iclr/rna-sbdd-integration

./scripts/restore_from_huggingface.sh /your/data/root

That downloads the repository, checks SHA256SUMS, extracts the release, places artifacts and checkpoints, and verifies all 26 checkpoint digests against MANIFEST.json. It exits nonzero on any mismatch.

If your token is not in the default location, export HF_HOME first — the CLI reports "Not logged in" for a misplaced token, which reads like a credential failure rather than a path one.

---

2. Bind the code to the new paths

The repository hardcodes ~1,450 absolute paths under two retired cluster roots. Nothing runs until they are remapped. This step does not exist in PORTABLE_TRAINING.md.

python scripts/remap_paths.py --repo . \
    --data-root        /your/data/root \
    --env-prefix       /your/conda/envs/rnasbdd \
    --crossdocked-root /your/data/root/sbdd_data

Dry run until you add --apply; it prints whatever the mapping missed and exits nonzero rather than leaving you half-bound. Optional: --annapurna-env, --rnamigos2-env, --crossdocked-pocket10, --tools-root. scripts/build_deltaai_pyg_wheels.sh is skipped on purpose — it guards a build with a cluster-specific shell glob needing a hand edit.

Digests are deliberately not rewritten. Moving a file does not change its content, so every dataset, split, freeze, and checkpoint digest must still match afterwards. A failure here means a corrupted transfer, not a stale pin.

To bind only configs rather than the whole repository, scripts/localize_rnasbdd_configs.py produces an independent configs copy and validates the data root as it goes (covers the four baselines).

---

3. Environment

Follow PORTABLE_TRAINING.md §"Create the Python 3.10 environment" — it has tested paths for both x86_64 and ARM64. The pinned environment was built on linux-aarch64 (GH200); on x86_64 the conda YAML and the Torch pin carry over but install prebuilt torch-scatter / torch-cluster wheels instead of building from requirements.deltaai-pyg-sources.txt.

Pinned: python 3.10.20 · torch 2.5.0 / cu124 · torch-geometric 2.6.1 · torch-scatter 2.1.2 · torch-cluster 1.6.3 · numpy 1.26.4 · rdkit 2024.09.6 · gemmi 0.7.5. Torch is deliberately absent from environment.deltaai.yml: the generic ARM64 PyPI wheel is CPU-only.

Always export PYTHONNOUSERSITE=1. Verify CUDA inside a job allocation — a login-node probe is inconclusive.

PYTHONNOUSERSITE=1 python -m unittest discover -s tests -q   # expect 733 OK, 2 skipped

A green suite is the real signal that §1–§3 worked: it checks that the digest-pinned paths resolve and the native extensions execute.

---

4. Train, generate, evaluate

Single-GPU torchrun is the only supported form; the launcher rejects every other argv shape.

torchrun --standalone --nproc_per_node=1 train_ddp.py \
    --config configs/rna/train/targetdiff_merged.yml \
    --logdir /your/runs/targetdiff-merged-scratch \
    --tag merged-rna-v2-scratch-seed2022 \
    --batch_size 8 --num_workers 0 --tranche_seconds 36000

Weights-only finetune adds --resume <ckpt> --finetune, which resets optimizer moments, scheduler, and the iteration counter so the arm stays comparable to a scratch arm. Merged configs: configs/rna/train/*_merged.yml (targetdiff, diffsbdd, graphbp, pocket2mol, biopolymer_context_targetdiff, and the _finetune variant).

--num_workers 0 is not a performance setting: worker-local RNG is not checkpointed, so anything above 0 breaks exact resumability across tranches.

Batch size is a launch flag, not a config value, and this matters. TargetDiff runs pass --batch_size 8; every proposed-variant run inherited batch_size: 4 from configs/common/targetdiff_train.yml. Anything comparing those families at equal iteration count compares unequal data. Pass --batch_size explicitly for every arm you intend to compare.

Generation follows PORTABLE_TRAINING.md §"Sampling, reconstruction, and explicit evaluator base paths`, with one substitution — use the merged size prior:

--molecule-size-prior         .../rna-sbdd-v2-merged-train-molecule-size-prior/rna_sbdd_v2_merged_train_molecule_size_prior.npy
--molecule-size-prior-sha256  5d33e13f79146f5648b6b6785b8f3cc8c2a68d3086d66dd33c0ab021940182a3
--molecule-size-prior-sidecar .../rna_sbdd_v2_merged_train_molecule_size_prior.manifest.json
--molecule-size-prior-sidecar-sha256 cbfa8695ca4424b4b569c09dfe2e98d5ad4493bdf3f8e836c1da1ba8bdeb2e58

The size prior belongs to the benchmark, not to a model. Every arm being compared must bind the same one. The pre-merge rna-sbdd-v2-train-molecule-size-prior yields numbers not comparable with the merged one: the average size shift is only −0.30 heavy atoms, but it is bin-dependent and changes sign (4 of 10 pocket-size bins move by more than one atom, +1.13 to −1.41), so it does not cancel within a run.

Two evaluator facts that cost real time to rediscover:

• RNAmigos2's mixed head is a rank normalized within the submitted batch. Best-of-K is 1.0 by construction and any delta against a separately scored reference is meaningless. Score the K generated molecules and the native reference in one batch.
• AnnapuRNA is not concurrency-safe on a node. Naive parallelism fabricates a large data-quality failure rate; node-lock it. Its score also correlates with ligand size at r = −0.923, so it is not a size-free binding measure.

---

5. Checkpoints

checkpoints/MANIFEST.json maps each file to its SHA-256, original path, and binding specs.

bound/ — the 20 checkpoints referenced by run specs and configs, across six model families (TargetDiff, DiffSBDD, Pocket2Mol, GraphBP, CrossTargetDiff, BiopolymerContextTargetDiff) in CrossDocked source-pretraining and RNA scratch arms. These are the weights behind the reported numbers and cannot be regenerated. One — biopolymer-context-targetdiff-crossdocked-source-pretraining-e2/.../167000.pt — is the pre-registered transfer initialization; re-pretraining it later gives a different checkpoint and voids the pre-registration.

merged-wave-terminal/ — the last checkpoint from each of six runs cancelled at 6.4 hours of a 10-hour budget. Diagnostic evidence, not protocol results. The pre-registered rule is "freeze the budget in advance, take the terminal checkpoint, select nothing"; none reached the frozen budget, so picking among them is exactly the monitor-selection the rule forbids. Do not quote numbers from them without saying this.

---

6. What to do next

a. Decide what the merged wave becomes. As it stands nothing can produce a published number under the merged benchmark — every older formal run used the superseded 6,513 split, and the six merged arms were cancelled early. Two honest routes:

• Rerun the six to the frozen 10h latch (~72 GPU-h) for the pre-registered result unchanged.
• Or re-register the budget at the wall clock all six actually reached. They ran 6:21–6:27, so a ~6.3h matched budget preserves the rule's structure exactly — matched wall clock, terminal checkpoint, no monitor selection — and every arm's last checkpoint qualifies. Cost: none had converged at 6.3h, so the claim weakens to "compared at an equal sub-convergence budget", which must be stated wherever those numbers appear.

b. Run the batch-matched arms. run_specs/*-bs8-e2.toml are committed and contract-verified but never launched. Until they run, proposed-model vs TargetDiff is equal-compute but not equal-data and cannot settle the architecture question. Related: the convergence table reading "the proposed model needs ~2x the steps" (105,000 vs 56,000) is almost entirely the batch flag — in molecules those are 420,000 and 448,000, so it converges marginally earlier. Per-run statements are unaffected; each run's budget still exceeds its own convergence point.

c. DiffSBDD never converged inside the latch. It needs a longer budget or a stated caveat wherever its numbers appear.

d. Two implemented models have never been trained. rna_chemistry_targetdiff (per-nucleotide gates plus an atom-role residual) and masked_atom_targetdiff (MDLM/LLaDA-style absorbing-state discrete diffusion, with working fragment inpainting) are CPU-verified and covered by tests. Supporting evidence: the existing RNA residue-identity gate is zero-initialized, which starts the model blind to which nucleotide it is, and both merged arms switched it on monotonically without plateauing (|tanh g| 0 → 0.30 scratch, 0 → 0.27 finetune, with the residual weight growing alongside rather than shrinking to compensate). That says the model wants RNA-specific residue capacity; it does not establish that finer parameterization pays. Only training these tests that.

e. Generation and evaluation (~450 GPU-h) has not run under the merged prior. Every published number needs it, and old pilots cannot be reused because they bind the pre-merge prior.

f. Open human decisions: accepting scientific validity of the merged split and the interaction family; whether to extend source pretraining (its best validation sits at 95.4% of budget, so its transfer numbers read as a lower bound); and whether to adopt masked diffusion, which costs ~30 GPU-h of source re-pretraining and voids the current pre-registered selection.

---

Reproducibility notes

• Checkpoint selection never touches the frozen 713 test split.
• The train-contained monitor is not a held-out validation and is never described as one.
• Formal runs record branch, HEAD SHA, and a clean-tree check at job start; a mismatch aborts with exit 70.
• Repo-source digests (driver scripts, environment specs) are advisory and drift with legitimate edits. Frozen-input digests (dataset, split, freeze, checkpoint, indices, prior) stay fatal on mismatch.

Citation

Not yet published. If you use this before it is, please open a discussion.