Wet-lab selection · pre-final

Fc variant final selection

Two disjoint design sets, one per RLXF reward arm, to test FcCALM as a standalone optimization signal against the coverage-gated ensemble reward. Selected from the 1.34 M-variant annotated pool; doubled quotas for a final trim by the team.

3,980

designs selected

two disjoint arms · doubled quotas

1,990 / 1,990

gate / fccalm

coverage-gated vs FcCALM-only

100%

in variable region

all mutations within positions 17–153

cross-arm overlap

every sequence is globally unique

313

observed (real binders)

166 memorized + 147 corpus fill

3,667

novel designs

2,413 full-match + 781 at 7/8

708

dual-verified

full-match by primary (ws4-ohe) + orthogonal ESM-C

5.1%

OOD (in-region)

rarely-varied positions inside 17–153

What this is

Each target profile's quota is doubled and filled per arm from two buckets: an observed bucket of real-label-verified binders ranked by predicted expression, and a generate bucket of novel designs (verifier full-match first, then 7/8) diversified across four mutation-position clusters with an in-zone preference. Every candidate is hard-filtered to wild-type length and mutations inside 17–153. The two arms are then made globally disjoint by alternating shared sequences, so the pair validates more unique sequences.

Use the per-profile explorer below to sort and filter every sequence and export your picks.

Composition

Where the 3,980 designs come from

Novel designs dominate; the observed bucket contributes 313 experimentally-grounded binders (real labels matching the profile).

Full-match (novel) · 2,4137/8 (novel) · 781Memorized binders · 166Corpus-fill binders · 147

Expression

Predicted expression by arm

ESM-C predicted expression (log2-enrichment; higher is better). gate vs fccalm.

gate fccalm

Per-profile breakdown (per arm: gate / fccalm)

Profile	Target	Difficulty	Selected	Observed	Full-match	7/8	OOD	Mean expr
11111111	Everything	easy	20/20	4/4	16/16	0/0	0/0	3.25/3.70
00000000	Nothing	easy	20/20	4/4	16/16	0/0	0/0	2.30/2.43
01000000	FcRn only	easy	20/20	4/4	16/16	0/0	0/0	2.72/2.58
10000000	R1 only	easy	20/20	4/4	16/16	0/0	0/0	2.07/2.21
11000000	R1; FcRn	easy	30/30	6/6	24/24	0/0	0/0	2.90/2.20
11111110	All except R3B	easy	40/40	8/8	32/32	0/0	1/0	3.45/2.88
10111111	All except FcRn	easy	40/40	8/8	32/32	0/0	0/0	2.69/2.90
01111111	All except R1	medium	200/200	7/7	193/193	0/0	0/0	2.55/2.12
01111110	no R1; no 3B	medium	100/100	15/15	85/85	0/0	0/0	3.05/2.85
10111110	no FcRn; no 3B	medium	100/100	8/8	92/92	0/0	0/0	2.47/2.07
-----110	3A selective	medium	100/100	20/20	80/80	0/0	16/25	2.81/2.49
11001111	All except R2A	hard	200/200	1/0	3/15	196/185	3/6	1.12/1.40
11110111	All except R2B	hard	200/200	1/1	199/199	0/0	0/18	2.21/1.58
11111001	All except R3A	hard	7/7	7/7	0/0	0/0	0/2	0.93/0.67
00111111	no R1; no FcRn	hard	200/200	0/0	200/200	0/0	0/0	3.10/2.76
10110111	no 2B; no FcRn	hard	200/200	0/0	0/0	200/200	0/0	3.21/3.84
--110---	2A selective	hard	100/63	20/20	80/43	0/0	2/11	2.34/2.04
--001---	2B selective	hard	100/100	20/20	80/80	0/0	2/0	2.36/2.14
-----001	3B selective	hard	50/100	20/20	30/80	0/0	28/21	1.17/2.39

Bit order R1·FcRn6·R2AH·R2AR·R2B·R3AF·R3AV·R3B. Each cell is gate / fccalm.

Per-profile explorer — every sequence, sortable & filterable

Click a column header to sort (click again to reverse). Filter row: text columns match substrings; numeric columns accept >=3, <2, 1..5 (a bare number means ≥); arm, bucket, source and the Y/N columns are dropdowns. Filters combine (AND). achieved = verifier's binarized call; real = training label (observed rows). Per-receptor columns (R1…3B) are binding probabilities. Full annotations: runs/final_selection/selected_all.csv.

Method

How each design was chosen

Hard constraints first: wild-type length and all mutations inside 17–153 (drops 23% of the pool; the assay cannot screen out-of-region mutations).
Observed bucket (up to 2×from_observed): real-label matches, memorized first then corpus fills, ranked by GPU-scored ESM-C expression (obs avg 3.3 vs 2.2 novel).
Generate bucket (to 2×total): full-match (tier 1) then 7/8 (tier 2); each tier split into 4 mutation-position clusters, round-robin. Within a cluster the gate arm orders in-zone then expression; the fccalm arm orders in-zone, then FcCALM agreement, then expression (its own reward) — lifting the fccalm arm's FcCALM full-match rate to 0.47 vs 0.20 for gate.
Binding oracle: the full-match gate is the FEV-tuned ws4-ohe verifier (FEV exact-match 0.86, vs 0.37 for the old one-hot); every design is also cross-checked with an orthogonal ESM-C verifier — conf = dual means both agree (highest confidence).
One-hot fallback (conf = fb): three profiles (11111001, --110---, -----001) can't fill on ws4-ohe — but ws4-ohe is blind there (it mis-calls even their real binders), so its rejection isn't informative. Their quota is topped up from the previous one-hot selection, flagged verifier-unvetted. Treat fb as the lowest-confidence tier.
Disjoint + alternation: a sequence wanted by both arms goes to alternating arms (1st→gate, 2nd→fccalm, …); the loser backfills.
Nothing else is filtered — OOD (5.1%), liabilities, expression ride along for the team's final trim (that is what this explorer is for).

Generated by scripts/generate/select_final.py + build_selection_dashboard.py from runs/annotated_pool/UNIFIED. Controls (IgG1 / N297Q / GAALIE / 299A-IYG / STR) added at plate design. Detail: runs/final_selection/README.md, docs/methodology_final_selection.md.