// Worked examples
See the pipeline output
Each worked example is the same pipeline a paying customer runs — predicted 3D structure, five-frame conformational ensemble, RNA-applicable pocket ranking, and a downloadable bundle with full per-cluster metadata. Numbers are locked benchmark cells; reproducer is on the methodology page.
TPP riboswitch (2GDI)
Strict@1, 71% overlap. The canonical single-seq win.
Single-sequence arm. The pre-pilot screen does not flag 2GDI for MSA mode (min_id 0.761). The pipeline recovers the TPP binding-site cluster at rank 1 with 71% residue overlap — well above the 50% strict threshold. fpocket druggability score on the rank-1 cluster is 0.00; the geometric ranker recovers it on persistence and binding-residue stability.
Thi-box TPP (2HOJ)
Strict@1 at 53% even with backbone RMSD 14 Å — local geometry preserved.
Worst backbone fit in the v0.2 benchmark (RMSD 13.95 Å) yet the rank-1 cluster strictly recovers the binding site at 53% overlap. The TPP-binding pocket geometry is preserved within the prediction even as the global fold drifts. The companion piece to 5C45 (poor fold, NEAR) and 2GIS (good fold, NEAR).
THF riboswitch (4LVV)
MSA recovers strict@1 where single-seq gave neither (19% → 50%).
The pre-pilot screen flags 4LVV for MSA mode (min_id 0.762, diverse-tail homologs present). Single-seq arm: rank-1 cluster at 19% overlap — neither strict nor near. MSA arm: rank-1 cluster at 50% overlap — strict@1. The customer-facing message: on this kind of target, the opt-in MSA path is what produces recovery.
SAM-I riboswitch (2GIS)
Best-quality structure in the benchmark (pLDDT 0.82, RMSD 1.3 Å); near@1 at 38%.
Pre-pilot screen excludes MSA (min_id 0.780, no diverse tail). The single-seq prediction is the best in the benchmark by both structure metrics — pLDDT 0.822 and backbone RMSD 1.33 Å — yet the rank-1 cluster lands at 38% overlap: near@1, not strict. Geometry quality is necessary but not sufficient.
FMN riboswitch (5C45)
Near@1 at 40% overlap. Smallest target (54 nt); backbone RMSD 10 Å.
Smallest target in the v0.2 benchmark. Pre-pilot screen excludes MSA (min_id 0.808, no diverse tail). Backbone RMSD 10.2 Å against the deposited co-crystal — by global-fold standards, poor. The rank-1 cluster nonetheless picks up 40% of the FMN binding-site residues: above near@1, below strict@1. We ship this case because it makes the global-vs-local quality gap explicit.
B12 riboswitch (4GXY)
Longest target (161 nt). MSA produces near@1 where single-seq is neither (18% → 35%).
Longest target in the v0.2 benchmark and the only one that triggers the screen on both criteria (min_id 0.705 AND 29.5% homologs in the diverse 70–80% identity band). Single-seq lands at 18% overlap (neither); MSA lifts the rank-1 cluster to 35% (near@1). Pairs with 4LVV to define the empirical range over which MSA mode adds value.
// Full benchmark
Seven cleft-binder targets with strict@1, near@1 and overlap percentages. Including the one out-of-scope neither case.
See benchmark →// Methodology
Full pipeline, third-party attribution and licences, strict@K / near@K definitions, scope and limitations.
Read methodology →// Pilot pricing
Per-target Explorer pilot or multi-target Discovery campaign. EU jurisdiction; zero-retention option on Enterprise.
See pricing →