// Fragment discovery
Structural analysis for fragment discovery
Fragment campaigns share a workflow shape: compute the pocket, soak fragments, characterize weak binders by SPR, ITC, NMR, or X-ray, build up to leads. This page names what this platform surfaces for that workflow specifically, what it deliberately does not do, and the honest limitations that apply to fragment work in particular — not a general capability pitch.
// What this surfaces
Decision support before the first soak
Ranked pockets on an ensemble
Candidate pockets computed across a five-frame conformational ensemble, not a single static structure, ranked by cross-frame persistence. A top-3 shortlist to prioritize before campaign design — never a single confident top-1 answer.
Per-residue interaction fingerprints
Where a self-matching ligand-bound structure exists for the family, PLIP-derived contact fingerprints (hydrogen bonds, salt bridges, pi-stacking) from representative complexes, mapped onto the predicted pocket residues.
Structural neighbor & ligand precedent
Family classification, known ligand precedent from curated PDB structures, and Foldseek structure-based neighbors (protein only) as prioritization signal — what has been fragment-screened before, on related folds.
Honest confidence flags
Six rules-based checks (no learned calibration) — low structure confidence within a pocket, poor numbering-crosswalk coverage, disordered-region overlap — surfaced explicitly rather than folded into a single opaque score.
// Scope boundaries
What this explicitly does not do
Named as scope boundaries, not weaknesses to apologize for — this platform surfaces structural evidence; it does not replace the experimental and computational tools fragment teams already run.
- —Fragment library screening — this platform ranks pockets on a target, it does not screen a library against them.
- —Affinity prediction — no binding-affinity or Kd estimate is produced for any fragment or compound.
- —Molecular docking — no ligand poses are generated or scored against the predicted pockets.
- —Ligand design — no molecule generation, elaboration, or SAR guidance.
// Honest limitations
What to discount, specifically for fragment work
The apo-structure ceiling. Predicted structures are apo (ligand-free). A 78-pair survey found the binding site is the part of a molecule that moves most on ligand binding — a real limit on how much any apo-structure pocket prediction, including this one, can be trusted for an induced-fit site. Read the full findings.
Report the top-3, never the top. On both fragment-canonical targets run for this write-up, the correct pocket was present and well-resolved — but on HSP90AA1 it ranked #2, not #1, at 95% real contact-residue overlap. The same pattern already documented for FKBP12 and ADRB2 in the protein benchmark. Treat every report as a prioritized shortlist, not a single confident answer.
Foldseek is protein-only. Structure-based structural-neighbor search does not apply to RNA targets — the RNA evidence layer stops at family classification, ligand precedent, conservation, and PLIP interaction fingerprints.
// Worked examples
Two fragment-canonical targets, run and independently verified
HSP90AA1 (N-terminal domain)
Rank 2 · 18/19732 aa full-length — one of the most fragment-screened ATP pockets in structural biology.
Independently checked against PDB 1YET (geldanamycin-bound). The correct pocket matches 95% of real contact residues, including the well-characterized Asn51 and Phe138 — but ranks #2, not #1. Naive whole-domain RMSD is a poor 12.8 Å, almost entirely driven by the N-domain “lid” that closes on ligand binding; the core domain matches at 0.70 Å. The load-bearing example for this write-up precisely because the limitation is visible, not hidden.
Open full report →BRD4 (bromodomain 1)
Rank 1 · 12/141362 aa full-length — the acetyl-lysine pocket behind the BET-inhibitor chemical series.
Independently checked against PDB 3MXF (JQ1-bound). Clean rank-1 recovery at 86% real contact-residue overlap, including the signature Asn140 hydrogen bond — CA RMSD 0.89 Å on the folded BD1 domain itself. Full-length BRD4 carries large disordered linker regions the ranker had to correctly ignore; the 3D viewer colors by per-residue confidence so those regions read as low-confidence, not broken.
Open full report →Have a target from your own fragment campaign? See the engagement models on the main page, or reach out directly.
Get in touch