Semi-de-novo Walkthrough
A real run, end to end. The conceptual page explains the five stages; this one drives them on two real pools — first validating the funnel on a labeled target, then producing a shortlist on an unlabeled one — so you can see exactly what goes in, how much each stage cuts, and how to read the output.
For a DLL3 pool the wet lab had already confirmed a set of binders. We spike those known positives into the ~264,000-clone pool, run the whole funnel, and ask a simple question: do they come back out near the top? On the first pass, 0 of 21 known binders survived — the cheap Stage-1 filter was silently discarding them.
Three real causes, all found by the run: (1) roughly half the pool clones carried a constant-region tag the reference binders lacked, so the same molecule split into two clusters; (2) the CDR3 finder failed on about a third of real sequences; (3) the abundance-ranked budget cut dropped low-abundance real binders. After fixing all three, 21 of 21 came back. Only then is the ranking worth reading.
Takeaway: on a labeled target, always run the recovery gate first. If known binders don't come back, an upstream stage is dropping them — and that is surfaced, not hidden.
A BCMA pool of ~1.2M unique clones streams through Stage 1 (pure sequence, no structure) down to 4,000. Only then does the expensive tier run: the top survivors are folded as antibody–antigen complexes on the GPU folding backend and scored from the predicted pose. 400 folds, not 1.2M — that is the whole point of the funnel.
Each folded complex yields two independent signals: interface confidence (how sure the model is the two chains form the predicted interface) and a contact-based binding-energy estimate converted to a predicted Kd. On this shortlist the two are essentially uncorrelated — and ranking on one alone versus the consensus produced top-10 lists that shared zero candidates.
Neither signal is ground truth, so the robust choice is the consensus: a candidate has to score well on both to rise. The delivered shortlist has a median predicted Kd around 0.7 nM with most candidates under 10 nM — but see the limits below before reading that as measured affinity.
Combined score: the consensus rank across interface confidence and predicted binding energy — the primary sort.
Predicted Kd: a calibrated affinity estimate from the pose, in nM. A prediction, not a measurement.
Abundance: read fraction in the pool — a weak prior. Many top hits are low-abundance: the funnel surfaces rare-but-strong binders that pure NGS-count ranking would bury.
Epitope bin / specificity: which target construct the candidate engages, and whether it clears the counter-screens. Downloadable as CSV + FASTA for expression.
Predicted, not measured. Every score is a model estimate; wet-lab binding is the only truth. The shortlist is where to spend expression capacity, not a validated affinity table.
Specificity needs counter-screens. A single-antigen target with no paralog counter-constructs cannot be specificity-gated — the specificity column is only meaningful when the panel includes off-target constructs to score against.
The affinity enrichment is partly circular. If the ranking weighs predicted binding energy, the shortlist will of course look strong on predicted binding energy. The independent interface signal hedges this, but it is not external validation.
Cost scales with candidates × constructs. Folding is the expensive step; the funnel bounds it with a per-stage cardinality budget and folds only the specificity-critical constructs in bulk.