GuoJicz518/RIDE-RL-pools

RIDE-RL RNA Inverse Folding Pools

Curated RNA target pools for reinforcement-learning post-training of RNA inverse-folding models, and for evaluating them. 527 training targets and 153 held-out test targets, drawn from RNA3DB and verified to be single, independently-foldable chains.

Companion code: https://github.com/Gabrile166/RIDE-RL

Why these pools exist

Two problems made the obvious choices unusable.

Composites do not fold on their own. Selecting chains by resolution, date and length alone yields many that are one piece of a larger assembly: one strand of a duplex, a segment threaded through a protein, a fragment of a ribosome. The PDB entry records the conformation the chain holds inside that assembly. Fold its sequence alone and you get the conformation it adopts alone, which is a different thing. In a batch selected this way, the native sequences themselves folded back to a median GDT-TS of 0.053. When the native sequence cannot recover its own reference, no designed sequence can score on that target either, so every model collapses to noise and the comparison measures nothing.

A pool must be learnable to be informative. Targets that are trivially easy or hopelessly hard both carry little signal.

Both pools are therefore filtered on a foldability criterion: the native sequence, folded by RhoFold+ and compared against the deposited structure, must reach GDT-TS >= 0.30. Median native ceiling is 0.64 (train) and 0.60 (test) -- see native_ceiling_* below.

Splits

split	targets	length range	median length	median native ceiling (GDT-TS)
train	527	27-258 nt	75	0.640
test	153	15-186 nt	65	0.601

Disjointness is verified, not assumed: the two splits share zero sequences and zero target ids.

Fields

field	description
target_id	stable identifier, rna3db_<pdb>_<chain>_<hash>
sequence	native RNA sequence (ACGU)
length	sequence length in nucleotides
pdb_id, auth_chain_id	provenance in the PDB
source_split	originating RNA3DB split
rna3db_component, rna3db_cluster_id	RNA3DB structural clustering, for diversity accounting
conformer_index	distinguishes multiple conformers of one chain
structure_hash	hash of the reference coordinates
native_ceiling_gdt_ts	GDT-TS of the native sequence folded back by RhoFold+
native_ceiling_tm_score	same, TM-score
native_ceiling_rmsd	same, RMSD in angstrom
ref_c4p_coords	reference C4' coordinates, flattened, length x 3
ref_backbone_coords	reference backbone coordinates, flattened, length x 3 x 3
mask_coords	per-residue validity mask, length length

Reading the coordinates

Both coordinate fields are flattened for portability. Restore them with:

import json
import numpy as np

rec = json.loads(open("test.jsonl").readline())
n = rec["length"]
c4p = np.array(rec["ref_c4p_coords"]).reshape(n, 3)
backbone = np.array(rec["ref_backbone_coords"]).reshape(n, 3, 3)
mask = np.array(rec["mask_coords"], dtype=bool)

The native ceiling, and how to use it

native_ceiling_gdt_ts is the score obtained by folding the native sequence and comparing it to the deposited structure. It is the practical upper bound for that target under this evaluation protocol: a designed sequence is measured against the same reference through the same folding model, so it inherits the same error floor.

Report designs relative to it. Absolute scores conflate design quality with how well the folding model handles the target. Stratifying by ceiling also matters: in our own study, a reward that appeared to improve on the pool average turned out to be improving only on low-ceiling targets, where the reference is least trustworthy, while regressing on the high-ceiling ones. Averages hide that.

Provenance

Derived from RNA3DB. Reference coordinates come from RCSB PDB entries. Native ceilings were computed with RhoFold+, structural comparison with US-align. Construction scripts are under scripts/dataset/ in the companion repository; the pipeline and every filter threshold are documented in docs/DATASET_PIPELINE.md.

Limitations

• Foldability screening uses RhoFold+, so the pools carry that model's inductive biases. A target excluded for a low native ceiling is not necessarily a poor structure; it may be one RhoFold+ handles badly.
• RNA3DB component labels are coarse. component_1 dominates by label, though sequence-level diversity is high (median pairwise identity 0.27).
• Multiple conformers of the same chain appear as separate records, distinguished by conformer_index. Group by pdb_id if that is not wanted.
• No experimental secondary structures. Work needing them must derive them, e.g. by folding the native sequence, which is a prediction and not ground truth.

Citation

Manuscript in preparation. Please cite the repository until it appears.