1,341,341 designed Fc variants, fully annotated
Two AI reward strategies — a coverage-gated binding ensemble (gate) and FcCALM alone (fccalm) — each drove 19 profile-specialist generators. This is what they produced, how faithfully they hit each target binding profile, and how the two compare on binding, expression, and developability.
The one-paragraph verdict
The coverage-gated ensemble reward is the stronger binding signal —
higher full-match (13% vs 10%), higher agreement (0.72 vs 0.69), lower OOD, and a visibly
crisper match to the target receptor pattern. FcCALM alone is genuinely
competitive: essentially tied on expression and developability, and it even wins binding on a
few hard profiles. Both collapse on the hardest selective targets (10110111,
-----001) — which is exactly why we over-generated those. The set is delivered
fully annotated and unfiltered: pick your thresholds per column.
Did the specialists hit the target?
Target vs. achieved binding, receptor by receptor
What we see
Each row is a target profile; each cell is the mean predicted bind rate the fixed verifier assigns that receptor. The middle panel (gate) tracks the target pattern more crisply than the right panel (fccalm) — brighter where the target wants binding, darker where it doesn't. Both arms nail the easy anchors (11111111, 00000000) and both blur on the hard selective profiles at the bottom, exactly where we over-generated.
Which arm designs the better binder
What we see
Each dot is one profile; points below the diagonal favour gate, above favour fccalm. Gate wins binding fidelity across most profiles (full-match and agreement mostly below the line), while expression is essentially tied (dots hug the diagonal). The coverage-gated ensemble reward is the stronger binding signal; FcCALM keeps up on developability and even wins binding on a few specific hard profiles.
How on-profile each specialist landed
What we see
Full-match rate = fraction the verifier calls fully on-profile. Easy anchors are near-perfect; hard profiles drop to single digits — expected, and the reason we generated 500k there. The per-profile winner flips: gate dominates 00111111 (51% vs 17%) and 11000000 (71% vs 21%); fccalm wins -----001 (14% vs 0.4%). 10110111 (no 2B; no FcRn) is the hardest — near-zero full-match for both, though agreement stays ~0.7–0.8.
Yield and diversity after dedup
What we see
Specialists re-emit the same sequences heavily, so most raw samples are duplicates. The three hardest profiles (11001111, 11111001, 10110111) were topped up with a higher-temperature (1.1) re-generation to expand their unique pool — the gen_temp column marks each design's source. After cross-temperature dedup the set holds 1,341,341 unique variants from ~12M sampled. Easy profiles collapse to a few thousand uniques; the topped-up hard profiles now hold tens of thousands each.
Per-profile scorecard
| profile | meaning | tier | unique | full-match | agree | expr | novel | novel picks | OOD |
|---|---|---|---|---|---|---|---|---|---|
| g · f | g · f | g · f | g · f | g · f | g · f | g · f | |||
| 11111111 | Everything | easy | 3,591·3,852 | 93.9%·88.6% | 0.99·0.98 | 1.56·1.68 | 98%·98% | 3,323·3,351 | 2%·1% |
| 00000000 | Nothing | easy | 11,709·9,277 | 99.9%·100.0% | 1.00·1.00 | -0.21·-0.18 | 100%·97% | 4,484·3,417 | 0%·0% |
| 01000000 | FcRn only | easy | 12,077·7,197 | 100.0%·99.8% | 1.00·1.00 | 0.78·1.36 | 100%·54% | 11,680·3,826 | 0%·0% |
| 10000000 | R1 only | easy | 9,134·5,058 | 100.0%·99.2% | 1.00·1.00 | 0.99·1.19 | 96%·32% | 8,691·1,604 | 0%·1% |
| 11000000 | R1; FcRn | easy | 7,579·1,536 | 70.8%·20.6% | 0.96·0.86 | 0.33·0.76 | 100%·92% | 3,922·238 | 1%·5% |
| 11111110 | except R3B | easy | 3,430·4,084 | 87.1%·58.5% | 0.98·0.93 | 2.12·1.67 | 97%·97% | 2,932·2,363 | 2%·1% |
| 10111111 | except FcRn | easy | 2,975·3,847 | 95.2%·93.2% | 0.99·0.99 | 1.52·1.31 | 40%·68% | 1,125·2,307 | 1%·1% |
| 01111111 | except R1 | medium | 3,192·6,710 | 30.1%·9.4% | 0.87·0.80 | 1.91·1.65 | 89%·91% | 943·581 | 10%·4% |
| 01111110 | no R1; no 3B | medium | 3,671·5,390 | 33.1%·14.8% | 0.88·0.73 | 1.90·1.84 | 67%·75% | 901·580 | 7%·4% |
| 10111110 | no FcRn; no 3B | medium | 10,587·13,603 | 18.7%·47.5% | 0.89·0.91 | 1.46·1.17 | 60%·95% | 1,421·6,337 | 3%·2% |
| -----110 | 3A selective | medium | 23,164·16,279 | 74.4%·48.3% | 0.89·0.71 | 1.92·1.57 | 96%·94% | 12,962·6,348 | 94%·78% |
| 11001111 | except R2A | hard | 15,661·10,670 | 0.0%·0.1% | 0.71·0.65 | 1.30·1.14 | 100%·99% | 12·64 | 11%·18% |
| 11110111 | except R2B | hard | 11,225·10,018 | 8.0%·1.5% | 0.78·0.73 | 1.68·1.37 | 98%·97% | 891·150 | 22%·18% |
| 11111001 | except R3A | hard | 15,105·9,928 | 0.2%·0.5% | 0.74·0.66 | 1.05·0.78 | 99%·99% | 350·261 | 7%·18% |
| 00111111 | no R1; no FcRn | hard | 24,276·18,218 | 50.7%·17.3% | 0.92·0.78 | 1.74·1.32 | 100%·98% | 12,244·2,979 | 6%·8% |
| 10110111 | no 2B; no FcRn | hard | 17,382·28,455 | 0.0%·0.2% | 0.67·0.83 | 1.39·1.62 | 99%·98% | 71·759 | 10%·3% |
| --110--- | 2A selective | hard | 37,113·47,070 | 15.8%·1.5% | 0.59·0.40 | 1.36·0.85 | 98%·98% | 5,063·584 | 38%·34% |
| --001--- | 2B selective | hard | 122,461·100,014 | 4.6%·4.6% | 0.62·0.62 | 0.40·0.36 | 98%·98% | 2,170·2,273 | 48%·69% |
| -----001 | 3B selective | hard | 65,175·105,480 | 0.4%·13.6% | 0.66·0.62 | 0.43·0.64 | 99%·98% | 84·8,971 | 57%·56% |
Each cell is gate · fccalm; the bold value is the better arm (lower is better for dup and OOD).
Developability
Predicted expression
What we see
Expression is the ESM-C predictor's output on the log2-enrichment scale (>0 = enriched = expresses). The vast majority of designs express well (medians ~1.0–2.0), with the two arms comparable. The null binder 00000000 sits lowest — sequences engineered to bind nothing also express a touch worse — but still mostly positive.
Sequence liabilities
What we see
Liability counts are near-identical across arms — the reward never optimized against them, so they reflect the base library. Oxidation-prone residues (Met/Trp) dominate the count; unpaired-cysteine risk is essentially zero (disulfides intact). Downstream filtering on these columns is cheap and lossless — nothing here is pre-removed.
Instability, hydrophobicity, charge, naturalness
What we see
Both arms produce tightly-overlapping biophysical distributions — instability index ~46, GRAVY ~−0.6 (hydrophilic), pI ~6.7. FcGPT naturalness (rightmost) is also matched, confirming both reward signals stay in-distribution rather than drifting to exotic sequence space.
What kind of sequences we generated
Where the designs mutate
What we see
Mutations concentrate almost entirely on the competence-zone positions — the residues the screening library actually varied. Designs rarely stray outside it, which is what keeps the binding verifier trustworthy for most of the set.
Amino-acid usage at the varied positions
What we see
At the most-varied positions a handful of substitutions dominate (bright cells; · marks the wild-type residue). The designs explore a focused, biologically-plausible sequence space rather than random noise.
Mutation load
What we see
Most designs carry ~3–8 mutations from wild-type (median 6), consistent across difficulty tiers — realistic engineering distances, not heavily-mutated outliers.
Out-of-distribution rate
What we see
OOD = a design mutates outside the competence zone, where binding predictions are least reliable. Anchor profiles are almost fully in-zone; the selective (--) profiles are heavily OOD (up to 94% for -----110) — they only constrain 3 receptors, so the specialist roams the other positions freely. Filter on is_ood for OOD-sensitive picks.
Novelty — are we designing or memorizing?
How novel are the designs?
What we see
97.5% of all designs are novel — exact copies of a training sequence are rare (2.5% overall). But a few easy anchors leak: fccalm memorizes heavily on 10000000 (R1-only, only 32% novel) and 01000000 (FcRn-only, 54%), and both arms partially on some medium profiles. The hard/selective profiles are almost entirely novel (>95%) — there's simply no training sequence to copy. The bottom panel is the one that matters: the actionable-novel pool — novel and on-profile and expressing — totals 120,262 designs (gate 73.3k, fccalm 47k), swinging by orders of magnitude per profile: gate owns 00111111 (12k) and -----110 (13k); fccalm owns -----001 (9k vs gate's 84). This is the pool a wet-lab list is drawn from.
Novel vs. memorized designs
What we see
Novel and memorized designs are biophysically similar — overlapping expression — so novelty here isn't bought at the cost of quality. Novel designs carry slightly more mutations, as expected (memorized ones are exact library hits, clustered at the library's typical edit distances). Only 37 designs across the whole set match an already-assayed FEV sequence, so almost everything is genuinely untested.
Do the annotations agree with reality?
What we see
The 3.8% of designs that are exact training-library sequences come with real measured binding labels — an independent check on both the verifier we annotate with and the prompt (24,396 recovered sequences, 68,391 measured receptor calls). The verifier matches the measured label 96% of the time (mean MCC 0.91) across all receptors — validating the annotation the whole 806k set is built on. The prompted profile matches truth less often (86%), worst on R3B (66%): the specialist sometimes retrieves a sequence whose real R3B/R2 state differs from what was asked. Decisively, when the verifier and the prompt disagree, the measured label sides with the verifier 85% of the time (vs 15% for the prompt). Practical takeaway: trust classifier_full_match / pred_* over the profile a sequence was generated for. Caveat: these are in-distribution library sequences, so this is best-case in-zone accuracy — reliability drops where is_ood.
How the signals relate
How the annotations relate
What we see
Expression and FcGPT naturalness correlate positively (both favour natural sequence). Coverage and OOD are mirror images by construction. Full-match and agreement track together, and edit distance / liabilities sit largely independent — so filtering on developability doesn't silently throw away binders.
Explore each profile
Pick a target profile to see, for that specialist specifically: how each arm hit the target receptor-by-receptor, the expression and mutation-load distributions, which positions it varies and to what, and concrete example sequences.