ESMFold2 Binder Design

ESMFold2 Binder Design

Design minibinders and antibody scFvs against a target protein using Biohub ESM release 3.4.1.post1. Each trajectory runs the published 150-step optimization and scores candidates with the native models. Download every candidate, optimization trajectory, scoring result, and predicted complex.

Pricing

Runs use metered billing at 42 credits per minute. The reservation shown before submission estimates 20 minutes per candidate per trajectory, with a minimum reservation of 42 credits. Unused reserved credits are returned after settlement. A run stops when its credit budget is exhausted; the maximum runtime is 12 hours.

Inputs

Input	Formats	Requirements
Target sequence	Text, .txt, .fasta, .fa	One uppercase protein sequence using canonical amino acids. Omit when using a target preset.
Custom binder template	Text, .txt, .fasta, .fa	One template using uppercase amino acids for fixed positions and # for positions to design. Omit when using a binder preset.

Each uploaded file can be up to 10 MiB. Target presets are cd45, ctla4, egfr, pd-l1, and pdgfr. Binder presets are minibinder, trastuzumab_framework_vhvl, atezolizumab_framework_vhvl, and ocankitug_framework_vhvl. A custom input requires a name different from these preset names.

A custom binder template must include at least one # position to design. Hotspots must match the residue and one-based position in your supplied target or the selected preset sequence.

In batch mode, each target and template combination becomes a separate job. With a target preset, omit target inputs and supply the custom templates to batch.

Settings

Parameter	Type	Default	Description
Target name or preset	string	custom-target	Name for the supplied target, or an available target preset.
Binder template name or preset	string	minibinder	Available binder preset or a custom template name.
Design trajectories	integer	1	Independent optimization runs. Seeds increase by one between runs.
Candidates per trajectory	integer	1	Candidates optimized together. Larger batches need more GPU memory.
Seed	integer	0	Nonnegative random seed.
Additional scoring models	boolean	true	Runs the published 15 additional scoring checkpoints. Increases runtime.
Antibody scoring	enum	auto	Automatic recognition, antibody, or minibinder scoring. Must match a selected binder preset.
Target hotspot residues	JSON array	[]	Residue letter and one-based target position, for example ["L150"].
Epitope contact distance	number	12	Contact distance in angstroms.
Deterministic execution	boolean	false	Requests deterministic PyTorch algorithms where supported.

Outputs

Output	Contents
Selected designs	Published selection results and selection.parquet, including the minibinder pI filter and native ranking. Selection can be empty.
All scoring results	Native critic scores in JSON and TSV, including logits in the JSON download.
Structures	Native PDB files for scored and selected binder-target complexes.
Sequences	Binder FASTA and complete target-binder candidate sequences in JSON.
Run files	Optimization trajectories, original inputs, pinned model and source identities, settings, and complete run log.

Scores describe predicted structural confidence and model preferences. They do not establish experimental binding or affinity. All candidates remain downloadable even when none pass the published selection filter.

The Biohub binder design notebook describes the design and selection protocol.