Validation Report

Two real runs: a labeled target (DLL3) used to validate the funnel on known binders, and an unlabeled target (BCMA) run in production to a ranked shortlist with real structure-based scoring. Every score is a model prediction; wet-lab binding is the only truth.

Generated 2026-07-05

DLL3 — recovery gate (validate before trusting)

A funnel that runs is not a funnel that works.

263,958

Pool (unique)

Known positives

0 / 21

Recall (first pass)

21 / 21

Recall (after fixes)

Three real causes found and fixed:

A constant-region tag on ~half the pool clones (absent from the reference binders) split the same molecule into two clusters.
CDR3 extraction failed on ~36% of real sequences, collapsing clustering.
The abundance-ranked budget cut discarded low-abundance real binders.

BCMA — production run

1,200,612

Pool

4,000

Stage 1 · sequence filter

400

Stage 2 · folded + scored

100

Shortlist

400

Folds

0.69 nM

Median predicted Kd

58 / 100

Sub-1 nM

97 / 100

Under 10 nM

Each folded complex gives two near-independent signals — interface confidence and a contact-based predicted Kd. Ranking on interface confidence alone versus the consensus produced top-10 lists that shared zero candidates, so the funnel ranks on the consensus: a candidate must be strong on both.

Rank	Combined score	Predicted Kd (nM)	Abundance
1	0.960	0.08	0.0414
2	0.976	0.07	0.0069
3	0.935	0.13	0.0005
4	0.880	0.12	0.0003
5	0.882	0.16	0.015
6	0.885	0.25	0.0003
7	0.883	0.19	0.0357
8	0.911	0.03	0

Many top hits are low-abundance — the funnel surfaces rare-but-strong binders that pure NGS-count ranking would bury.

Honest limits

Predicted, not measured. The shortlist is where to spend expression capacity, not a validated affinity table.

The Kd enrichment is partly circular — the ranking weighs predicted binding energy. The independent interface signal hedges this, but it is not external validation.

Specificity needs counter-screens. BCMA now includes TACI and BAFF-R counter-select constructs; specificity calls are only meaningful when the real folding backend scores both target and counter-screen constructs.