Semi-de-novo Triage

Funnel a wet-lab NGS pool down to a ranked shortlist. The lab supplies the diversity; the platform ranks and resolves epitope / specificity.

Validation reportWalkthrough
Workflow

The wet lab supplies the diversity; the platform triages ~10⁶ sequences to a ~10² shortlist.

01Original plan
  1. 1
    Library enrichment (phage-display panning + functional sort) → antibody pool
  2. 2
    NGS readout → ~10⁶ sequences (with abundance)
  3. 3
    Dock each clone against the target and score
  4. 4
    Filter by structural domain / sequence
  5. 5
    Classify against different antigens (epitope / specificity)
  6. 6
    Composite evaluation → best ranking → list usable at the bench
02As implemented
  1. 1
    Pool sequences1.2M
    the whole enriched library, streamed
  2. 2
    Sequence filter4,000
    length / liability gates, CDR3 clustering (cheap, full 10⁶)
  3. 3
    Fold + affinity400
    dock survivors only; score pose (iptm + ΔG→Kd)
  4. 4
    Epitope × specificity
    matrix vs the antigen panel
  5. 5
    Composite rank + anchor check
    known positives validate, never train
  6. 6
    Shortlist → expression100
    median predicted Kd ~0.7 nM
03Where the two differ, and why
  • Docking all 10⁶ is infeasible → two tiers: a cheap sequence filter first, fold only the survivors.
  • Two near-independent signals (interface confidence + predicted ΔG) ranked by consensus — single-metric top-10s shared nothing.
  • Abundance is a weak tiebreak, never a gate — rare-but-strong binders still surface.
  • Validated on the DLL3 known positives (21/21 recovered) before trusting it on unlabeled pools.
  • Predicted, not measured — wet-lab KD is the truth; specificity needs counter-screen constructs.
New run

The set of candidate sequences you want to triage — upload your own FASTA/CSV. Pools you upload appear in this list.

No sequence pool is loaded for this target yet. Choose a FASTA/CSV file below, then click Ingest & select.
Recent runs

No runs yet.