Fc Engineering — pipeline dashboard

Results for every stage of the Fc-variant design pipeline, data ingestion → final wet-lab selection.

1Data canonical corpora · per-source inputs · cross-source audit

Training variants (use these to train)

fc_retraining_v2
merged
dms + paper_refiltered + costain (1.47M rows). Priority dms > paper_refiltered > costain. dms is unfiltered (cleanest); paper_refiltered & costain get bind+non ≥ 2 per receptor. +48,939 costain-only rows over retraining_v1 for rare-profile coverage on R1/R2AR/R2B/R3AF.
rows
1,474,696
mean labels/row
1.61
all 8 labelled
5,066
edit dist (mean)
5.26 (p99 15)
N-glyco motif
14.2%
diversity
0.177 bits
20260528-104945_data_fc_retraining_v2_seed0 · 2026-05-28T11:05:07+00:00
fc_retraining_v1
merged
paper_refiltered + dms merged with bind+non ≥ 2 per receptor; dms wins on disagreement. The previous FcGPT retraining corpus; superseded by retraining_v2.
rows
1,425,757
mean labels/row
1.37
all 8 labelled
4,181
edit dist (mean)
5.28 (p99 15)
N-glyco motif
14.4%
diversity
0.177 bits
20260517-202905_data_fc_retraining_v1_seed0 · 2026-05-17T20:36:25+00:00
fc_paper_reproduction
merged
paper_refiltered alone, no extra count filter. Used to reproduce the original FcGPT-paper classifier and pretraining baselines.
rows
1,108,275
mean labels/row
1.67
all 8 labelled
4,458
edit dist (mean)
6.12 (p99 15)
N-glyco motif
17.6%
diversity
0.182 bits
20260517-202417_data_fc_paper_reproduction_seed0 · 2026-05-17T20:27:37+00:00

Per-source datasets (inputs to the merges; raw shape of each library)

paper_refiltered
per_source
Canonical paper-source (combi_umi). Library-author guidance: replaces `paper`. Bind/non read counts are trustworthy.
rows
1,108,275
mean labels/row
1.67
all 8 labelled
4,458
edit dist (mean)
6.12 (p99 15)
N-glyco motif
17.6%
diversity
0.181 bits
20260515_data_paper_refiltered_seed0 · 2026-05-15T12:48:42+00:00
dms
per_source
Deep mutational scan. Densest per-row labels of any training source; sequences cluster tight around WT (low-edit-distance).
rows
324,092
mean labels/row
2.36
all 8 labelled
2,240
edit dist (mean)
2.33 (p99 4)
N-glyco motif
3.2%
diversity
0.111 bits
20260515_data_dms_seed0 · 2026-05-15T10:49:17+00:00
costain
per_source
Co-stain assay (Edward Irvine, May 2026). 69,202 sequences, four receptors only (R1, R2AR, R2B, R3AF). Designed to fill rare profiles (R2AR-selective, broad activating) under-covered by combi_umi/dms. 70.7% of rows are sequences not in any other source.
rows
69,202
mean labels/row
2.53
all 8 labelled
0
edit dist (mean)
4.67 (p99 9)
N-glyco motif
8.6%
diversity
0.164 bits
20260528-082727_data_costain_seed0 · 2026-05-28T08:27:35+00:00
first_experimental_validation
per_source
Wet-lab observed labels from the FcGPT paper's experimental panel. Held-out benchmark, never merged into training.
rows
1,974
mean labels/row
4.46
all 8 labelled
658
edit dist (mean)
8.49 (p99 13)
N-glyco motif
0.0%
diversity
0.225 bits
20260515_data_first_experimental_validation_seed0 · 2026-05-15T10:46:10+00:00
paper
per_source
Original paper dataset. AUDIT ONLY — count_R* is PCR-bias-uncorrected per library-author guidance. Not part of training.
⚠️ audit only · counts uncorrected for PCR bias · not used for training
rows
3,100,154
mean labels/row
1.70
all 8 labelled
21,556
edit dist (mean)
6.21 (p99 16)
N-glyco motif
15.2%
diversity
0.193 bits
20260515-132223_data_paper_seed0 · 2026-05-15T13:37:24+00:00

Cross-source overlap audit

Cross-source overlap audit
audit
Pairwise sequence overlap and per-receptor label-agreement between every source, recomputed under several quality filters. The basis for the dms-wins-on-disagreement merge policy.

2Verifier / classifier per-receptor binding models · K-fold leaderboard

Classifier K-fold leaderboard
classifier
5-fold cross-validated test metrics for paper / retraining_v1 / retraining_v2 baseline / retraining_v2 sweep. The sweep wins the is_varied subgroup decisively (+4.6 pp MCC, +19.4 pp exact match over paper) with tight error bars.

3FcGPT pretrain conditional ProGen2 base checkpoints

FcGPT base pretrain comparison
pretrain
Three FcGPT-10M base checkpoints (paper_reproduction / retraining_v1 / retraining_v2), one per corpus. Likelihood + per-source loss, generation-panel aggregate, and per-group profile-conditioning agreement. retraining_v2 is the recommended RLXF base.

4RLXF specialists one GRPO+LoRA adapter per profile, per arm

RLXF specialist results
rlxf
63 per-prompt GRPO specialists (21 FEV/paper profiles × 3 bases). Target full-profile-match jumps ~0.10 → 0.89-0.94 vs base at ~0.97 validity; 49/63 reach ≥0.90. retraining_v2 yields the strongest specialists. Sparse 2-receptor partials flagged as laggards.

5Generation & annotation sample specialists → the 1.34 M-design annotated pool

Candidate pipeline results
candidates
1,575 final candidates (525/base, 25/profile) — valid, novel, on-profile, high-expression. Click any profile for the deep view: per-receptor binding vs target, metric comparison across bases, edit-distance + expression distributions, and a mutational heatmap.
Fc design set — results dashboard
designs
The 1,341,341-variant fully-annotated Fc design set: binding fidelity, expression, and developability across two reward strategies (coverage-gated ensemble vs FcCALM). Per-receptor target-vs-achieved, a gate-vs-FcCALM comparison, and a per-profile explorer with example sequences. Built for the wet-lab team and PI.

6Final wet-lab selection two disjoint arms · verifier tiers · order set

Fc variant final selection — wet-lab sets
selection
The pre-final wet-lab pick list: two disjoint design sets (1,747 + 1,759) across the coverage-gated ensemble and FcCALM-only reward arms, doubled quotas for a final trim. Observed real-binder bucket + novel tiered/cluster-diversified bucket, all mutations confined to the variable region (17-153). Per-profile breakdown with per-arm stats, plus an interactive per-sequence table (sort, filter, export) with primary + orthogonal verifier confidence tiers, so the wet-lab team can slice and decide.

RResearch & write-ups workstream reports & verdicts

WS7 · FcCALM
paper
Can a contrastive dual-encoder (FcCALM, after CALM/MRC) generalize where a one-hot verifier can't? A two-axis result across 13 runs: the receptor encoder does generalize to an unseen FcγR (leave-one-out AUROC 0.72–0.84, ≈no penalty ex-FcRn — the gospel's headline claim, validated), but the model does not rank binding at unseen Fc positions — confirmed even by a fully-faithful 1-to-1 build (per-receptor co-crystal pooling, hard-neg families 2–3, cross-attention, 5× data — all chance OOD, best-in-zone). Verdicts on MRC & CALM, fidelity-to-spec accounting, corrected gospel PDB errors.