ESMFold2 design
Choose your target, pick a small de novo binder or a paired heavy + light scFv (a single-chain antibody fragment), and get back designs ranked by a 0-to-1 interface confidence score in one model pass.
What it is for
You want a paired heavy + light scFv — a single-chain antibody fragment — with all six binding loops designed against your target at once, which no other tool here does. It also builds small de novo binders by a different route to RFdiffusion, worth a run when a target has gone quiet. For single-domain nanobodies use RFantibody.
Designs a binder by running a structure predictor backwards: it starts from a soft, blurred sequence and nudges it one gradient step at a time until the fold network believes the result binds your target. The same machinery does two jobs — small de novo binders, and a paired heavy + light scFv (a single-chain antibody fragment) with all six binding loops designed together, which no other tool here does. Designs from this method have been taken to the bench against PDGFRB, EGFR, PD-L1, CD45 and CTLA4, reaching nanomolar affinity. ESMFold2 design, Candido et al., bioRxiv 2026, built on the ESMC protein language model.
When it fits:
- You want a paired heavy + light scFv with all six binding loops designed jointly against your target. No other tool here does this.
- Your target has gone quiet under RFdiffusion and you want a method that searches differently — a different prior often surfaces different binders.
- You want to compare designed loops across three humanised frameworks that have been to the bench (trastuzumab, atezolizumab, ocankitug).
- You want to calibrate your own wet-lab setup against one of the five targets from the paper before spending on a novel one.
Inputs
You will need:
- Target sequence: pick one of five paper-validated presets or paste a single chain (30 to 800 aa).
- Pick minibinder mode or scFv mode. For scFv, pick a framework (trastuzumab, atezolizumab, or ocankitug).
- No PDB required. The gradient loop is sequence-only.
Each run uses a preset that sets the scale and scope:
- De novo minibinder (60 to 200 aa)
- Free 60 to 200 aa scaffold generated with an isoelectric point filter (pI < 6) baked in. No framework constraints. Equivalent goal to RFdiffusion plus ProteinMPNN but in one gradient pass through ESMFold2.
- scFv with framework-locked CDRs
- All six CDRs designed jointly on a locked humanized framework. Three frameworks available: trastuzumab, atezolizumab, ocankitug. The only catalog tool that designs paired heavy + light scFv CDRs end-to-end.
Parameters you set on the form:
- Preset
- De novo minibinder generates a free 60 to 200 aa scaffold with an isoelectric-point filter (pI < 6). scFv designs all six CDRs on a locked humanized framework. Same model, different binder factory.
- Target
- Pick one of five paper-validated presets (CD45, CTLA4, EGFR, PD-L1, or PDGFR, with sequences from UniProt cropped to the relevant ectodomain) or paste your own protein sequence (30 to 800 aa, canonical amino acids only).
- Binder framework
- scFv mode only. Locks the framework backbone and sequence; the gradient loop only mutates the six CDR regions. Trastuzumab (anti-HER2 IgG1, humanized), atezolizumab (anti-PD-L1 IgG1, humanized), or ocankitug (humanized IgG1). All three are clinically validated frameworks.
- Starting seed
- Integer seed for the soft-sequence initialization. Different seeds yield different designs. When Seeds to run is greater than 1 this is the first seed in the sweep; the orchestrator runs
[seed, seed + n)in parallel. - Seeds to run
- Number of parallel seeds to sweep (1 to 64). Each seed gets its own H100 worker, all run in parallel, so a 16-seed sweep finishes in about the same wall-clock as one seed. Results from every seed merge into one globally-ranked table. Use this when you need to build a candidate library against a target. GPU time scales with seeds, not with batch size: each seed is a separate container, and a seed's designs share one.
- Batch size
- Designs produced per gradient run (1 to 6). They share one H100 container, so a higher batch adds no container — but you pay for the GPU time your run uses, and a bigger batch runs longer: measured wall-clock grew a little faster than in proportion to the batch. (It used to claim a batch of 6 was free of both; two production runs and the GPU-seconds bill disproved that.) Default 3. Single-design runs often return
dropafter the iPTM and pI gates. Bump to 6 for first-pass exploration when you can wait for the longer run; drop to 1 only when you already know the target gives clean hits.
How to read the results
Per-design table with designed sequence, iPTM, distogram iPTM proxy (or CDR distogram iPTM proxy for scFvs), final loss, isoelectric point, source seed, and predicted complex PDB. Strict-pass classification surfaces designs worth ordering (minibinder: iptm ≥ 0.75 AND pI < 6; scfv: cdr_distogram_iptm_proxy ≥ 0.5 AND iptm ≥ 0.75). Sweep mode (Seeds to run > 1) merges every seed's designs into one globally-ranked table.
Where a tool reports them, the scores mean:
- ipTM
- Predicted confidence in the contact between two chains, on a 0 to 1 scale. Higher is better: 0.65 to 0.75 depending on the tool. Individual tools set their own pass bar a little either side of that — this guide's own results summary above states this tool's. On a multi-chain target the number may cover the target's own chain–chain interface as well as the binder's, so read the per-tool note on the results table before comparing designs on it.
- pLDDT
- Per-residue confidence in the predicted fold. Higher means the model is more sure of that part of the structure.
- i_pAE and pAE
- Predicted alignment error, at the interface (i_pAE) or across the whole structure (pAE). Lower is better.
References
Candido et al., bioRxiv 2026