// Experimental — provider-preview capability

Ligand-conditioned co-folding

OpenFold3-preview lets us extend the platform’s apo pocket assessment with ligand-conditioned holo prediction. Below are five canonical targets — three RNA, two protein — where experimental holo structures exist and can serve as ground truth. Metrics are honest: apo baseline vs. co-folded prediction vs. experimental holo, using consistent measurement definitions per target, full replicate distributions, and no hidden failures.

// 1 of 5 — The easy case

Risdiplam / SMN2 splice-site duplex

A short, mostly Watson-Crick-paired RNA duplex — the bulged 5’-splice-site element risdiplam binds to redirect SMN2 exon 7 splicing. Sets up the rest of the page: not every RNA target needs co-folding to get the fold right.

Ligand class

Splicing modifier (approved drug — the SMA therapy risdiplam)

Apo pocket-detection (this platform’s standard pipeline)

Groove binder — cleft-based detection isn’t in scope yet; public binding site used directly.

What that means here

Our cavity-based pocket detector (fpocket) is tuned for clefts, not shallow grooves — groove-binder detection isn't in scope yet. For this cross-validation we use the public, well-studied binding site from the published structure (8R62) directly, rather than our own detection.

// Apo vs. co-folded vs. experimental ground truth

Binding-site residues (all three structures)

Ligand — element coloring

Slate: apo prediction (no ligand information). Teal: co-folded prediction (sequence + ligand SMILES). Gold: experimental holo structure. Magenta marks the binding-site residues on every structure that shows them, so alignment or mismatch is visible directly. All three are pre-aligned on the binding-site region — toggling doesn’t require re-orienting.

// Metrics

Metric	Apo	Co-folded	Improvement
Binding-site RMSD	0.34 Å (±0.11)	0.22 Å (±0.06)	+0.12 Å
Global (whole-molecule) RMSD	2.05 Å (±0.18)	1.82 Å (±0.09)	+0.23 Å

Replicate distribution (9 apo / 9 co-folded runs)

Apo binding-site RMSDmin 0.23 · median 0.28 · max 0.52 Å (n=9)

Co-folded binding-site RMSDmin 0.15 · median 0.19 · max 0.36 Å (n=9)

Ligand-pose RMSD (co-folded)min 1.60 · median 1.72 · max 2.00 Å (n=9)

Default view shows the representative replicate closest to each arm’s median binding-site RMSD — apo: risdiplam_apo_seed_958682846_sample_2_model.cif (0.28 Å), co-folded: risdiplam_cofold_seed_1181241943_sample_2_model.cif (0.19 Å). Full distribution above.

// Ligand identity

Name: Risdiplam
Formula / MW: C22H23N7O · 401.2 Da
SMILES: CC1=CC(=NN2C1=NC(=C2)C)C3=CC(=O)N4C=C(C=CC4=N3)N5CCNC6(C5)CC6

// Ground truth reference

PDB 8R62 · 2024 · Solution NMR

"The diversity of splicing modifiers acting on A-1 bulged 5'-splice sites reveals rules for rational drug design." (2024).

// Query sequence source

Both 11-nt strands extracted directly from 8R62's own resolved chains A and B (chain A's 2 pseudouridine residues substituted with standard U for the OpenFold3 query — noted, not hidden). Chain B offset +11 applied uniformly for a single unambiguous residue-numbering space.

// Interpretation

Apo prediction already lands close to the true structure — a short, mostly Watson-Crick-paired duplex is not a hard fold to get right blind. Co-folding's improvement is real and consistent across every replicate (+0.12 Å binding-site RMSD), but small, because there isn't much conformational gap left to close. This is the expected shape of result when the apo baseline is already strong, not a limitation of co-folding itself — the same pattern shows up on FKBP12 below.

// 2 of 5 — The clean canonical win

HCV IRES IIa

A 38-nt hairpin domain of the hepatitis C virus internal ribosome entry site — a real, structured antiviral RNA target with a benzimidazole-class small-molecule binder. Shows what co-folding adds when there is genuine ligand-induced conformational change to recover.

Ligand class

Benzimidazole-class HCV IRES IIa binder

Apo pocket-detection (this platform’s standard pipeline)

Rank 1, 58% binding-site overlap — strict recovery.

What that means here

Apo pocket-detection lands the real site at rank 1, strict recovery, before any co-folding.

// Apo vs. co-folded vs. experimental ground truth

Binding-site residues (all three structures)

Ligand — element coloring

Slate: apo prediction (no ligand information). Teal: co-folded prediction (sequence + ligand SMILES). Gold: experimental holo structure. Magenta marks the binding-site residues on every structure that shows them, so alignment or mismatch is visible directly. All three are pre-aligned on the binding-site region — toggling doesn’t require re-orienting.

// Metrics

Metric	Apo	Co-folded	Improvement
Binding-site RMSD	5.48 Å (±0.17)	0.36 Å (±0.06)	+5.13 Å
Global (whole-molecule) RMSD	6.09 Å (±0.37)	0.70 Å (±0.14)	+5.39 Å

Replicate distribution (9 apo / 9 co-folded runs)

Apo binding-site RMSDmin 5.21 · median 5.40 · max 5.84 Å (n=9)

Co-folded binding-site RMSDmin 0.30 · median 0.34 · max 0.47 Å (n=9)

Ligand-pose RMSD (co-folded)min 0.50 · median 0.74 · max 1.24 Å (n=9)

Default view shows the representative replicate closest to each arm’s median binding-site RMSD — apo: hcv_iresiia_apo_seed_2746317213_sample_2_model.cif (5.40 Å), co-folded: hcv_iresiia_cofold_seed_2746317213_sample_2_model.cif (0.34 Å). Full distribution above.

// Ligand identity

Name: ISH (benzimidazole-class)
Formula / MW: C17H27N5O · 317.2 Da
SMILES: CN(C)CCCn1c(N)nc2ccc3O[C@@H](CN(C)C)Cc3c12

// Ground truth reference

PDB 2KTZ · 2010 · Solution NMR

Paulsen, Seth, Swayze, Griffey, Skalicky, Cheatham, Davis. "Inhibitor-induced structural change in the HCV IRES domain IIa RNA." PNAS 107 (2010).

// Query sequence source

Extracted directly from 2KTZ's own resolved chain A (1-38, no gaps) — used as both the query-sequence source and the holo reference, so registration is exact by construction.

// Interpretation

Clean, decisive recovery of both the fold and the ligand pose. Apo prediction already finds the right region (rank-1 pocket, strict recovery) but misses the binding-site geometry by ~5.5 Å; co-folding brings that to sub-Å accuracy (0.36 Å), with the ligand itself landing within 0.75 Å of its real experimental position. This is the target where ligand-conditioned co-folding does exactly what it's meant to do — resolve a real, otherwise-unrecoverable conformational gap on a simple hairpin fold.

// 3 of 5 — The honest partial case

FMN riboswitch

A 56-nt three-way-junction FMN riboswitch aptamer — a harder RNA tertiary fold than a simple hairpin, bound to a synthetic flavin-core cofactor analog. The trust-building beat: co-folding helps, it does not solve this one.

Ligand class

Flavin-core (isoalloxazine) cofactor analog

Apo pocket-detection (this platform’s standard pipeline)

Rank 1, 22% binding-site overlap — below near-recovery threshold (rank 2: 33%, near).

What that means here

Apo pocket-detection's rank-1 cluster misses the real site (22% overlap); the real site is closer to rank-2 (33%, near-recovery). Consistent with this being a harder three-way-junction tertiary fold.

// Apo vs. co-folded vs. experimental ground truth

Binding-site residues (all three structures)

Ligand — element coloring

Slate: apo prediction (no ligand information). Teal: co-folded prediction (sequence + ligand SMILES). Gold: experimental holo structure. Magenta marks the binding-site residues on every structure that shows them, so alignment or mismatch is visible directly. All three are pre-aligned on the binding-site region — toggling doesn’t require re-orienting.

// Metrics

Metric	Apo	Co-folded	Improvement
Binding-site RMSD	6.85 Å (±0.53)	4.80 Å (±0.40)	+2.06 Å
Global (whole-molecule) RMSD	17.79 Å (±0.58)	12.44 Å (±0.41)	+5.36 Å

Replicate distribution (9 apo / 9 co-folded runs)

Apo binding-site RMSDmin 6.22 · median 6.73 · max 7.85 Å (n=9)

Co-folded binding-site RMSDmin 4.23 · median 4.75 · max 5.53 Å (n=9)

Ligand-pose RMSD (co-folded)min 5.97 · median 7.84 · max 10.94 Å (n=9)

Default view shows the representative replicate closest to each arm’s median binding-site RMSD — apo: fmn_riboswitch_apo_seed_958682846_sample_3_model.cif (6.73 Å), co-folded: fmn_riboswitch_cofold_seed_958682846_sample_1_model.cif (4.75 Å). Full distribution above.

// Ligand identity

Name: GZG (flavin-core cofactor analog)
Formula / MW: C25H27N5O4 · 461.2 Da
SMILES: Cc1cc2N=C3C(=O)NC(=O)N=C3N(CCN(CCCC(O)=O)Cc4ccccc4)c2cc1C

// Ground truth reference

PDB 6DN2 · 2018 · X-ray crystallography

"Crystal structure of the FMN riboswitch bound to BRX1354 split RNA." (2018). Closest real match to “recently published” — no post-2023 FMN/RFN-element riboswitch-ligand structure exists in the PDB (checked directly).

// Query sequence source

Extracted directly from 6DN2's own resolved chain Y (56 nt, numbered 57-112 in the deposited file; offset -56 applied and verified before computing anything).

// Interpretation

Co-folding moves the prediction in the right direction — both the overall fold and the pocket region improve (+2.06 Å binding-site RMSD) — but neither arm reaches a usable absolute accuracy on this three-way-junction tertiary fold, and the ligand pose itself isn't reliably placed (8.1 Å mean). A harder structure-prediction problem than a simple hairpin, and co-folding alone doesn't close the gap. The honest read: this is a case for downstream refinement (e.g. short MD) on top of the co-folded output, not a final answer by itself.

// 4 of 5 — Protein, decisive

ABL1 kinase + imatinib (DFG-out)

The ABL1 kinase domain bound to imatinib (Gleevec), the classic DFG-out inhibitor complex. Shows the capability isn’t RNA-only — and isolates exactly what co-folding adds on a protein target where the pocket already exists in apo, but its conformational state does not.

Ligand class

ATP-competitive kinase inhibitor (DFG-out, approved drug)

Apo pocket-detection (this platform’s standard pipeline)

Rank 1, 76% binding-site overlap — strict recovery.

What that means here

Apo pocket-detection already finds the ATP-binding cleft at rank 1 — the cleft exists regardless of DFG-in/DFG-out state. Co-folding's value here is resolving the DFG-out conformational state itself, not finding the pocket.

// Apo vs. co-folded vs. experimental ground truth

Binding-site residues (all three structures)

Ligand — element coloring

Slate: apo prediction (no ligand information). Teal: co-folded prediction (sequence + ligand SMILES). Gold: experimental holo structure. Magenta marks the binding-site residues on every structure that shows them, so alignment or mismatch is visible directly. All three are pre-aligned on the binding-site region — toggling doesn’t require re-orienting.

// Metrics

Metric	Apo	Co-folded	Improvement
Binding-site RMSD	2.21 Å (±0.03)	0.72 Å (±0.61)	+1.49 Å
Global (whole-molecule) RMSD	4.97 Å (±0.10)	1.16 Å (±0.21)	+3.81 Å

Replicate distribution (9 apo / 9 co-folded runs)

Apo binding-site RMSDmin 2.16 · median 2.21 · max 2.25 Å (n=9)

Co-folded binding-site RMSDmin 0.36 · median 0.40 · max 1.88 Å (n=9)

Ligand-pose RMSD (co-folded)min 0.26 · median 0.37 · max 0.75 Å (n=9)

Default view shows the representative replicate closest to each arm’s median binding-site RMSD — apo: abl1_apo_seed_958682846_sample_3_model.cif (2.21 Å), co-folded: abl1_cofold_seed_1181241943_sample_3_model.cif (0.40 Å). Full distribution above.

// Ligand identity

Name: Imatinib (Gleevec / STI-571)
Formula / MW: C29H31N7O · 493.3 Da
SMILES: CN1CCN(CC1)Cc2ccc(cc2)C(=O)Nc3ccc(C)c(Nc4nccc(n4)c5cccnc5)c3

// Ground truth reference

PDB 1IEP · 2002 · X-ray crystallography

Nagar, Bornmann, Pellicena, Schindler, Veach, Miller, Clarkson, Kuriyan. "Crystal structures of the kinase domain of c-Abl in complex with the small molecule inhibitors PD173955 and imatinib (STI-571)." Cancer Res 62 (2002).

// Query sequence source

Extracted directly from 1IEP's own resolved chain A (274 residues, numbered 225-498 in the deposited file) rather than sliced from UniProt — UniProt numbering disagrees with 1IEP's construct at the N-terminus (likely an expression-tag artifact), confirmed by direct comparison before running anything.

// Interpretation

Apo prediction already finds the ATP-binding cleft — that part of the pocket doesn't depend on ligand information. What apo prediction can't do is resolve the DFG-out conformational state that imatinib specifically selects for; co-folding recovers that almost completely (binding-site RMSD 2.21 Å → 0.72 Å mean, global RMSD 4.97 Å → 1.16 Å). One flexible P-loop residue doesn't converge well in 2 of 9 replicates — even those two still clearly beat the apo baseline, and the DFG motif itself is unaffected in either case.

// 5 of 5 — Protein, smaller effect

FKBP12 + FK506 (rapamycin-class)

FKBP12 bound to FK506 (tacrolimus), a rigid, well-packed immunophilin pocket with comparatively little room for ligand-induced movement. Rounds out the honest picture: co-folding’s effect size tracks how much conformational change there actually is to recover.

Ligand class

Immunophilin ligand (rapamycin/FK506-class, approved drug)

Apo pocket-detection (this platform’s standard pipeline)

Rank 3, 43% binding-site overlap — near recovery.

What that means here

Apo pocket-detection finds the real site at rank 3 (near-recovery), not top-2 — consistent with the platform's documented “top-3, never the top” pattern on this exact target.

// Apo vs. co-folded vs. experimental ground truth

Binding-site residues (all three structures)

Ligand — element coloring

Slate: apo prediction (no ligand information). Teal: co-folded prediction (sequence + ligand SMILES). Gold: experimental holo structure. Magenta marks the binding-site residues on every structure that shows them, so alignment or mismatch is visible directly. All three are pre-aligned on the binding-site region — toggling doesn’t require re-orienting.

// Metrics

Metric	Apo	Co-folded	Improvement
Binding-site RMSD	0.40 Å (±0.04)	0.30 Å (±0.04)	+0.09 Å
Global (whole-molecule) RMSD	0.46 Å (±0.03)	0.36 Å (±0.01)	+0.09 Å

Replicate distribution (9 apo / 9 co-folded runs)

Apo binding-site RMSDmin 0.32 · median 0.40 · max 0.44 Å (n=9)

Co-folded binding-site RMSDmin 0.25 · median 0.32 · max 0.40 Å (n=9)

Ligand-pose RMSD (co-folded)min 0.53 · median 0.61 · max 0.66 Å (n=9)

Default view shows the representative replicate closest to each arm’s median binding-site RMSD — apo: fkbp12_apo_seed_958682846_sample_1_model.cif (0.40 Å), co-folded: fkbp12_cofold_seed_958682846_sample_3_model.cif (0.32 Å). Full distribution above.

// Ligand identity

Name: FK506 (tacrolimus)
Formula / MW: C44H69NO12 · 803.5 Da
SMILES: CO[C@@H]1C[C@@H](CC[C@H]1O)/C=C(C)/[C@H]2OC(=O)[C@@H]3CCCCN3C(=O)C(=O)[C@]4(O)O[C@H]([C@H](C[C@@H](C)C\C(=C\[C@@H](CC=C)C(=O)C[C@H](O)[C@H]2C)C)OC)[C@H](C[C@H]4C)OC

// Ground truth reference

PDB 1FKJ · 1995 · X-ray crystallography

Wilson, Yamashita, Sintchak, Rotstein, Murcko, Boger, Thomson, Fitzgibbon, Black, Navia. "Comparative X-ray structures of the major binding protein for the immunosuppressant FK506 (tacrolimus) in unliganded form and in complex with FK506 and rapamycin." Acta Crystallogr D 51 (1995).

// Query sequence source

UniProt P62942 canonical sequence (108 aa incl. initiator Met). 1FKJ's crystallized construct omits the initiator Met, so predicted residue N corresponds to 1FKJ residue N−1 — confirmed by direct residue-identity comparison at both chain termini after an initial off-by-one was caught independently (see /findings).

// Interpretation

FKBP12's pocket is small, rigid, and well-packed — there's comparatively little room for a ligand to induce conformational change in the first place. Co-folding still helps, consistently across all 9 replicates (+0.09 Å binding-site RMSD), but the effect size is an order of magnitude smaller than ABL1's. Read together, the two protein targets show the same pattern as the RNA ones: co-folding's benefit tracks how much real conformational change there is to recover, not a fixed capability level independent of the target.

This is a scoping-gate result, not a claim that ligand-induced conformational change is solved on either RNA or protein. See the platform’s core, benchmarked capability on /demos.

Get in touch