DRfold2¶

Single-sequence ab initio RNA 3D structure prediction with a composite RNA language model + denoised end-to-end learning. Tier-2 structure model alongside RhoFold; DRfold2 wins on novel-fold cases and matches AlphaFold3 on several benchmarks at ~10⁵× lower cost than fragment-assembly approaches like FarFar2/SimRNA.

Paper: Li et al., PLOS Biology 2026 — "Ab initio RNA structure prediction with composite language model and denoised end-to-end learning"
GitHub: leeyang/DRfold2
Source archive (citation): Zenodo 18401691 (DRfold2-V1.02 source code, 836 kB — not the weights)
Weights: 1.3 GB model_hub.tar.gz baked into the image at build time from upstream's zhanggroup.org plain HTTPS endpoint; no Zenodo/HF mirror exists yet
License: MIT (declared in upstream README body; no LICENSE file on the repo — flagged for redistribution review). Bundled Arena refinement utility (cloned from pylelab/Arena) has no declared license on its upstream — see "Limitations" below.
Device: GPU only. Auto-skipped under -profile cpu with a warning.
- rnazoo-drfold2:latest — single image; ~5 GB compressed
- No CPU image (the 4-model ensemble + IPA + Arena refinement is not feasible on CPU)

What it does¶

DRfold2 predicts atomic 3D structure from a single RNA sequence — no MSA, no template, no co-evolution from sister sequences. The pipeline is:

Composite RNA language model (RCLM) approximates full-likelihood co-evolutionary signal from unsupervised single-sequence input.
4-model deep-learning ensemble (cfg_95, cfg_96, cfg_97, cfg_99) generates candidate end-to-end structures and pairwise geometries.
Selection.py picks the best candidate; Optimization.py refines geometry with IPA (invariant point attention).
Arena (a vendored C-compiled chain refinement utility from pylelab/Arena) does coarse-to-atomic backbone reconstruction.
Optional clustering pass (--cluster) re-runs Selection+Optimization+Arena on alternative ensemble members to emit model_2.pdb, model_3.pdb, … alongside the primary model_1.pdb.

The final output is one PDB per input sequence.

Input format¶

Standard FASTA of RNA sequences (A/C/G/U; T accepted and converted internally). The wrapper iterates per record — each sequence runs as a separate DRfold2 invocation in a fresh per-sequence work dir.

Example smoke fixture (tests/data/drfold2_test.fa, 24-nt UUCG hairpin):

>uucg_hairpin
GAGCGGCAUCUUCGGAUGCCGCUC

Output format¶

A directory containing:

<safe_label>.pdb: one PDB file per input sequence (the top-selected DRfold2 model — equivalent to upstream's relax/model_1.pdb)
With --drfold2_cluster: additional <safe_label>_alt2.pdb, <safe_label>_alt3.pdb, … from the clustering pass

Per-sequence intermediate dirs (rets_dir/, folds/, relax/) are removed by default — pass --drfold2_keep_intermediate to retain them (can be many GB for longer sequences).

Run with Docker¶

# GPU only (no CPU image)
docker run --rm --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all \
  -u $(id -u):$(id -g) \
  -e HOME=/tmp -e USER=$(whoami) \
  -v /path/to/input.fa:/data/input.fa \
  -v /path/to/output:/out \
  ghcr.io/ericmalekos/rnazoo-drfold2:latest \
  drfold2_predict.py -i /data/input.fa -o /out

Add --cluster for alternative-conformation models, --keep-intermediate to preserve working dirs.

Run with Nextflow¶

# GPU only — DRfold2 auto-skips under -profile cpu
nextflow run main.nf -profile docker,gpu --drfold2_input /path/to/input.fa

Results land in results/drfold2/drfold2_out/.

Parameters¶

Parameter	Default	Description
`--drfold2_cluster`	`false`	Run additional clustering pass (emits `<label>_alt2.pdb`, …)
`--drfold2_keep_intermediate`	`false`	Retain per-sequence working dirs

Reading the output¶

PDB files are standard format — open in PyMOL, ChimeraX, or Bio.PDB:

from Bio.PDB import PDBParser
parser = PDBParser(QUIET=True)
struct = parser.get_structure("uucg_hairpin", "drfold2_out/uucg_hairpin.pdb")
for atom in struct.get_atoms():
    print(atom.element, atom.coord)

Comparison with the other structure models in the zoo¶

Model	Output	Method	MSA / template	Speed (per sequence)	License
SPOT-RNA	2D bpseq/CT (incl pseudoknots)	5-model TF ensemble	None	~ 1 s	MPL-2.0
RhoFold	3D PDB	E2E + AmberRelax (single-seq mode)	None	~ 5 min on CPU, < 1 min on GPU	Apache-2.0
DRfold2	3D PDB	4-model ensemble + IPA + Arena	None	~ 10-30 min for 30 nt; 40-120 min for typical RNAs	MIT

DRfold2's headline: it beats RhoFold/RoseTTAFoldNA/DeepFoldRNA on a 41-target TM-score benchmark (0.350 vs 0.292 next-best) and edges AlphaFold3 on novel folds. Top-L contact precision 49.0% vs RNA-FM 23.6% / RiNALMo 24.0%. Pseudoknot recovery 56% vs AF3 79%. Caveat: only 24% of benchmark targets clear TM≥0.45 (the structural-match threshold), so absolute accuracy on out-of-distribution RNAs is still poor in absolute terms — DRfold2 is a relative state-of-the-art, not a solved-problem tool.

For most use cases, start with RhoFold (faster, simpler, smaller image) and reach for DRfold2 when RhoFold's output is unsatisfactory or for sequences where co-evolution-free 3D prediction matters (novel folds, designed RNAs).

Limitations¶

GPU-only. The 4-model ensemble + IPA + Arena pipeline is impractical on CPU. Auto-skipped under -profile cpu with a warning.
Slow. ~10-30 min for short (24 nt) sequences on a consumer GPU (GTX 1650 Ti class), 40-120 min for typical 50-200 nt RNAs. Optional refinement (via OpenMM, not exposed in this wrapper) adds another ~180 min.
Memory. Large ensemble + IPA may require >4 GB VRAM for sequences > 50 nt. Consumer GPUs may OOM on long inputs.
Bundled Arena license. Arena (cloned from pylelab/Arena at build time) has no declared license on its upstream repo. The DRfold2 image as built can be used locally, but the project should not push it to a public registry until Arena's license is clarified or replaced. See docs/DRfold2-issues/arena-license.md for the redistribution decision log if you need to ship this image externally.
MIT license declared in README only. DRfold2's repo has no LICENSE file; the MIT terms appear in the README body. GitHub's API reports it as NO_LICENSE. The README declaration is enforceable but unconventional — flag before commercial use.
Inference-only. No fine-tuning support exposed.