Three molecules, three methods, one disagreement that matters
A real run, start to finish. HIV-1 TAR, the Tat peptide that binds it, and a small molecule — folded together by three independent methods. Two of them put the compound against the same face of the RNA. The third put it somewhere else entirely, with not one contact residue in common. Nothing below is a mock-up.
01
Describe the complex
One row per molecule. The composition is the form — there is no separate workflow to pick for an RNA–protein pair versus an RNA–protein–ligand one.
RNA29 nt
GGCAGAUCUGAGCCUGGGAGCUCUCUGCC
HIV-1 TAR
Protein11 residues
YGRKKRRQRRR
Tat peptide
Small moleculeSMILES
NC(=O)[C@@H](N)CCCNC(=N)N
argininamide
Run analysis
02
It runs on our GPUs
queued → running → done · 124 s
Three methods run one after another on a single card, which is why a three-molecule complex takes a couple of minutes rather than seconds. Nothing to install, and no model to choose.
The compute sits on EU infrastructure and your sequences stay on it. That is the default, not a paid add-on — zero retention and a contractual data path are what the paid tier adds on top.
03
Read the result
Checks first. One of the three methods placed two heavy atoms 0.47 Å apart at the RNA–peptide interface — closer than two atoms can be — and that is stated before any of the numbers built on those coordinates.
RNA structure analysis
GGCAGAUCUGAGCCUGGGAGCUCUCUGCC
29 nt · RNA · 3 methods · 124 s
Before you read this
This does not tell you whether the compound binds. Co-folding places a molecule against a structure; it does not discriminate binders from non-binders. Tested directly on known binders and decoys, pose convergence, iptm and cross-method contact agreement each failed to separate them. What follows is where each method put the molecule, and where the methods disagree.
Key observation
14/29 compared residues have the same contact state with the protein (chain B) (heavy atoms within 5 A) across all 3 engines.
Structural checks
Measured from the coordinates, independently of any model's own confidence.
Contact with the partner, by position
in contactnot in contact
Check
Boltz-2
Chai-1
OpenFold3
base pairing
22/29 residues in canonical Watson-Crick pairs (76% paired); 7 unpaired.
22/29 residues in canonical Watson-Crick pairs (76% paired); 7 unpaired.
22/29 residues in canonical Watson-Crick pairs (76% paired); 7 unpaired.
g tetrads
0 G-tetrad(s) detected.
0 G-tetrad(s) detected.
0 G-tetrad(s) detected.
composition
29 residues across 1 chain(s); 12 annotated base pairs.
29 residues across 1 chain(s); 13 annotated base pairs.
29 residues across 1 chain(s); 13 annotated base pairs.
Interfaces
3 molecule pairs, each measured separately on every method's structure. Listed in the order the molecules were entered.
HIV-1 TAR ↔ Tat peptide⚠ 1 of 3 with steric clashes
16–25 residues in contact across methods · Boltz-2 1.55 Å closest · Chai-1 0.47 Å — a clash, not a contact · OpenFold3 1.95 Å closest
HIV-1 TAR ↔ argininamideall 3 methods find an interface
3–10 residues in contact across methods · Boltz-2 2.08 Å closest · Chai-1 2.17 Å closest · OpenFold3 1.71 Å closest
Tat peptide ↔ argininamideall 3 methods find an interface
1–2 residues in contact across methods · Boltz-2 3.31 Å closest · Chai-1 2.55 Å closest · OpenFold3 3.92 Å closest
Method agreement
The same contact state with the protein (chain B) (heavy atoms within 5 A) compared position by position. Agreement between methods is not evidence that they are right — read it with the training-set overlap below.
methods differ heresame state in every method
14/29
positions with the same contact state
5
regions differ
11–15
largest difference
Jump to a difference
Structures
One viewer, switchable between methods. The camera is shared, so switching changes the structure and nothing else.
Loading viewer…
The camera is shared between methods on purpose — switching changes the structure and nothing else. Drag to rotate, scroll to zoom.
Training-set overlap
Whether this target was already in the PDB when these models were trained.
close relatives deposited that yearpublished training cutoff
Earliest close relative: 1ARJ_N — 100% identity, deposited 1995-08-30. That date, not the closest match, is what decides training overlap.
No training cutoff is published for boltz2, chai1, openfold3, so exposure cannot be confirmed from the engines’ own documentation. This relative has been in the PDB for 31 years, which predates the public release of every method run here — read the date as when the structure became available for training, not as confirmed exposure.
Methods run
OpenFold3
v0.5.0
9 structures
27 conformers
31.9 s
Boltz-2
v2.2.1
3 structures
9 conformers
27.5 s
Chai-1
v0.6.1
3 structures
9 conformers
58.5 s
The methods did not produce equal numbers of structures. Each is asked for the same number of diffusion samples; OpenFold3 also runs several model seeds and returns one structure per seed and sample, so it yields more. Each method still contributes one vote per position in the agreement figure — its own majority across its structures — so a method that sampled more has a steadier vote here than one that sampled less.
Scope of this analysis
This analysis does not determine a detected pocket is biologically relevant; any compound binds, or how tightly; ranking of the engines, or a single consensus structure; overall quality, druggability or confidence score.
// Coordinates
Every structure on this page as a PDB file — one per method, ready for PyMOL or ChimeraX. Methods are kept in separate files on purpose: they are independent answers, not conformers of one.
04
Take it further
A result is an object, not a dead end. Each analysis below runs on what this one produced, so the coordinates you were shown are the coordinates it uses.
Analyse the interfaces
Every pair in detail — which residues touch on both sides, the clashes, and how far the methods agree about each one separately.
Find pockets, then dock
Cavities on these coordinates, and a ligand placed in one of them. Docking is protein-only, for a reason we measured rather than assumed — so on this complex it is the pockets that are offered.