← Back to Home

RFantibody vs Biohub ESMFold2: a neutral-oracle face-off

De novo binder design head-to-head — RFantibody VHHs versus Biohub ESMFold2 minibinders & scFvs against challenging cancer antigens

June 2026 — Preprint v1.0

📄 Read the Full Preprint (PDF)    🧬 The RFantibody Designs

What this is

Two de novo binder design paradigms, on the same antigens. RFantibody (RFdiffusion → ProteinMPNN → RoseTTAFold2) — from our earlier dogfooding run, one lead VHH per target. And Biohub's ESMFold2 protocol, which designs binders by gradient descent over the binder sequence against an ESMFold2 structure objective (predicted inter-/intra-chain contacts), regularized by a down-weighted ESMC-6B language-model term. ESMFold2 is the design model; the 6-billion-parameter ESMC-6B is an auxiliary sequence prior.

We ran the Biohub protocol on 14 antigens, generating minibinders and scFvs, then scored every design with ESMFold2 interface confidence (ipTM). Biohub wins on all 10 shared targets — but we show why that is largely an artifact of who is grading the exam, and settle it with an independent oracle.

By the Numbers

2
Design platforms
14
Antigens
224
Biohub designs
38
Neutral-scored complexes
10/10
Targets Biohub leads (both oracles)
H200
GPU (Modal); ESMC-6B + ESMFold2
Boltz-2
Neutral oracle (AF3-class)

The headline — and the catch

Scored by ESMFold2, Biohub's best design beats the RFantibody VHH lead on every shared target, often by 0.6–0.9 ipTM units (median best ipTM 0.85 for Biohub vs 0.13 for RFantibody).

Per-target ESMFold2 ipTM: RFantibody VHH vs Biohub minibinder vs scFv
Figure 1. Best ESMFold2 ipTM per target. The grey baseline is RFantibody; blue and orange are Biohub minibinders and scFvs. Biohub leads on all 10 shared targets — but read Figure 2 before calling it quality.

The catch: evaluation circularity

ESMFold2 ipTM is exactly what the Biohub protocol optimizes. A high score is guaranteed by construction — the precise mirror of our earlier finding that the same RFantibody VHHs sail through RoseTTAFold2's own filters (pLDDT ≈ 0.90) yet score near zero on ESMFold2. Each platform looks excellent under the oracle that built it.

Pooled ESMFold2 ipTM distributions showing the circularity
Figure 2. Pooled ESMFold2 ipTM: RFantibody leads (n=10) vs all Biohub designs over the same 10 targets. Stark — but ESMFold2 is Biohub's home oracle, so the gap is expected, not proof of superiority.

The decisive test: a neutral oracle

To remove home-oracle bias, we re-scored the best design per platform per target with Boltz-2, an open AlphaFold3-class model that neither platform optimized against (RFantibody used RoseTTAFold2; Biohub used ESMFold2). Genuine DeepMind AF3 cannot batch ~200 jobs and its weights are access-gated; Boltz-2 is the runnable, equally independent substitute.

Home-oracle vs neutral-oracle ipTM dumbbell
Figure 3. Median best-per-target ipTM under each platform's design-region oracle (ESMFold2, open markers) versus the neutral Boltz-2 oracle (filled). RFantibody recovers; both Biohub modalities fall.

Under a neutral judge, the gap halves — but doesn't vanish

RFantibody designs recover sharply (median ipTM 0.13 → 0.37): ESMFold2 had been unusually harsh on them. Biohub designs fall (0.85 → 0.74 minibinder, 0.69 scFv), directly quantifying home-oracle inflation. Yet Biohub's best design still out-scores the RFantibody lead on 10/10 shared targets — by smaller, more variable margins (on EGFRvIII the RFantibody VHH itself reaches 0.79). The advantage is real, but modest.

Neutral Boltz-2 scores (10 shared targets)

TargetRFab VHHBiohub MBBiohub scFvBiohub best
b7h30.4960.7930.6010.793
cd470.1520.4530.5070.507
ceacam50.2530.3200.4720.472
egfr (cetuximab)0.2260.2700.7820.782
egfrviii0.7880.8940.6680.894
epha20.2730.7360.3990.736
gpc20.1750.7850.8610.861
her2 dom. IV0.4640.7420.8350.835
msln (N-term)0.5320.9070.7990.907
msln (C-term)*0.6820.9040.9220.922
median0.3690.7390.685

*MSLN C-terminal antigen is a 15-residue peptide; ipTM there is unreliable for both platforms.

What We Learned

Cross-oracle disagreement is the whole story

Confidence metrics from a design platform's own structure model are not evidence of binding — they are evidence the design satisfied its own objective. RFantibody looks confident under RoseTTAFold2; Biohub looks confident under ESMFold2. Only a model neither side optimized against (here, Boltz-2) gives a comparison worth anything. It should become standard practice in this field.

Minibinders generally beat scFvs

On most targets, free-form 60–200 aa minibinders reach higher ipTM than CDR design inside a fixed VH–VL framework — consistent with an easier optimization geometry. Useful when format flexibility is acceptable.

Read against the sampling asymmetry — and no wet lab yet

The Biohub number is the best of 8 designs per modality; the RFantibody number is a single lead. That alone favours Biohub. There is no experimental binding data for any design — ipTM is a confidence proxy, not affinity. The honest gold-standard experiments remain a symmetric best-of-N comparison and, ultimately, SPR/BLI. A data-quality aside: two PDB IDs in our internal target list were wrong (6W0B is a K⁺ channel, not HDAC8; 6R8H is triosephosphate isomerase, not PHD2), so those antigens were sourced from UniProt.

Engineering notes

Running the Biohub protocol reproducibly took two non-obvious fixes. The ESMC-6B + ESMFold2 critic ensemble occupies ~50–75 GB; at batch size 8 even an H200 (141 GB) ran out of memory, and a CUDA OOM on a warm, shared container corrupted its CUDA context, cascading into cuDNN/CUB failures on later jobs. Batch size 2 with automatic fallback eliminated all failures. On real Modal pricing (H200 $4.54/h, B200 $6.25/h) the dominant cost lever is amortizing the one-time model load across a batch — not the GPU tier; B200 is roughly cost-neutral for this small-batch, latency-bound optimization.

Open & reproducible

Full Preprint (PDF)  |  The RFantibody Designs  |  Antibody Research Hub

All designs (sequences, per-critic metrics, loss trajectories, predicted structures) and the neutral Boltz-2 scores are released for the computational protein design community.