An induced-fit survey that qualifies our own assumption

01 · Headline result

Across 52 apo/holo pairs where both global and binding-site displacement could be measured, binding-site RMSD exceeded global RMSD in 75% of pairs. Global backbone RMSD had a median of 1.71 Å (mean 2.39 Å, up to 14.7 Å); binding-site heavy-atom RMSD had a very similar median of 1.66 Å. The binding site moved less than the rest of the molecule in only 13 of 52 pairs (25%).

This is the opposite of what a “locally pre-organized pocket” intuition would predict. It does not contradict our benchmark result on the seven locked cleft-binder targets — that result is about a narrower, pre-organized fold class, and stands on its own (see the methodology page). It does mean that intuition does not generalize to apo/holo pairs broadly: RNA binding sites are, on average, exactly the part of the molecule most likely to move on ligand binding.

02 · The three counts

Multi-ligand sequences: 104

104 unique RNA sequences in the PDB are bound to 2 or more chemically distinct ligands, deduplicated from 2,485 raw (entry, chain) records covering 1,344 unique sequences overall. 50 of the 104 have exactly two distinct ligands — the cleanest comparison cases.

Composition caveat, load-bearing: the richest individual cases are SAR fragment series around one scaffold — 13 cobalamin analogs sharing a corrin core, 8 benzothiazolium dye variants, an aminoglycoside family sharing the 2-deoxystreptamine ring — not independent tests of unrelated chemotypes. Trimmed to genuinely orthogonal chemistry, the usable count is closer to a few dozen sequences than 104.

Apo/holo pairs: 78

78 sequences have both a ligand-free (apo) and at least one ligand-bound (holo) deposited structure. Geometry was computed for 58 of these; the global-vs-binding-site RMSD comparison specifically was measurable for 52. 18 of the 78 failed for identifiable reasons — 9 to a registration issue on very short (<30 nt) fragments, 4 to stale RCSB download links, 2 to chain-registration edge cases, a few others — reported here as failures rather than silently dropped.

Induced-fit distribution

See the headline result above. This is the count that answers the question the survey was built to ask.

03 · Ligand filtering methodology

315 ligand codes were included and 74 excluded, out of 389 distinct codes observed across the candidate structures. The filter was iterated twice after inspecting its own output — not assumed correct on the first pass:

• A first-pass name-match pattern (\bPHOSPHATE\b / \bSULFATE\b) wrongly excluded real metabolites — thiamine phosphate, PAPS, a cyclic-dinucleotide analog — because the true buffer/ion cases (PO4, SO4) were already caught by a separate ion pattern. The redundant, incorrect pattern was removed.
• Three common lab reagents (Tris, TAM, Hepes/MES) slipped past name-pattern matching because RCSB stores their systematic IUPAC name rather than the trade name a naive filter would catch. Explicit code-level exclusions were added.
• PEG-phosphate crystallization fragments (e.g. ligand codes S9L, PDI) were confirmed and excluded by manual inspection.

Deduplication is by sequence, not by PDB entry. Multiple deposited structures of the same RNA sequence — different crystal forms, different bound ligands, different resolutions — are grouped as one biological case, not counted separately. This is what makes the 104 and 78 counts meaningful as counts of distinct RNA targets rather than counts of PDB depositions.

A hard 400 nt length cap was applied consistently with the survey’s original scope, dropping 26 records above it — including two 2,880 nt “sequences” that were a chain-concatenation artifact for a single entry, not real biological data.

04 · Family distribution

The multi-ligand set is not one riboswitch class wearing different hats. Purine/preQ1 riboswitches (16 sequences) and aminoglycoside-model aptamers (15) are the two largest single buckets, but SAM, TPP, FMN, NAD, cobalamin, and cyclic-dinucleotide riboswitches, three aptamer subtypes, ribozymes, viral elements and tRNA all appear — 13 further named categories, plus 13 that did not classify cleanly against the rule-based family map.

About 21% of the whole ligand-bound universe (69 of 336 candidate sequences) did not classify against that family map. That is acceptable coverage for a survey; it is not an exhaustive taxonomy.

05 · What this means for pocket prediction

Every apo-structure pocket prediction method — ours included — starts from a structure that has not seen the ligand. If the binding site is, on average, the part of the RNA molecule most likely to move on binding, then any prediction made from the apo state alone is working against a real physical limit, not just a modeling gap that better methods can close.

This is the honest implication we want scientists evaluating this pipeline to have in front of them. It is also why the scope boundaries stated on the main page exist: we validate on pre-organized structural elements — riboswitches, aptamers, and structured viral elements — specifically because those are the cases where the binding site is comparatively stable across the apo/holo transition, not because we expect apo-structure prediction to generalize to every RNA target with a binding site.

This finding does not invalidate the locked seven-target cleft-binder benchmark (see methodology), which measures recovery on exactly that pre-organized fold class. It does mean the benchmark result should not be read as evidence that apo-structure prediction works broadly across RNA — it works where the physics allows it to, and we try to say so explicitly rather than let a narrow result imply a general one.

06 · Reproducibility

Validation query: the FMN apo (6WJR) / holo (3F2Q) pair correctly lands in the same deduplication group as 6 further FMN-bound entries. The HCV IRES domain IIa apo structure (1P5M) does not pair with any holo structure in this survey — confirmed not a query bug; 1P5M’s exact sequence appears nowhere else in the PDB, and the nearest bound structure is a genuinely different construct topology (continuous 55 nt vs. a 2-strand 16+20 nt split). Full per-pair RMSD and ANM descriptive-overlap tables are available on request.