Use Protein Hunter Online
Commercially Available Protein Hunter No-Code Web Server
Protein Hunter: Fine-Tuning-Free De Novo Protein and Binder Design with Iterative Structure Prediction
Protein Hunter is a fine-tuning-free framework for de novo protein design that starts from a blank, all-unknown sequence and iteratively refines it using an AlphaFold3-style structure predictor together with ProteinMPNN. The method is described in a bioRxiv preprint (DOI 10.1101/2025.10.10.681530) and covers unconditional design, protein binders, cyclic peptides, small-molecule binders, nucleic-acid binders, motif scaffolding and partial redesign.
Generative protein design usually relies on dedicated diffusion models that must be trained or fine-tuned for each task. Protein Hunter instead reuses structure-prediction models that are already trained, so the same pipeline can be pointed at very different design problems without any additional training.
A known weakness of single-pass hallucination is that long designs often fold poorly, and earlier cycling approaches such as AF2cycler were more constrained in how much of the sequence could change. Protein Hunter lets the sequence be fully replaced at each iteration, which helps the design move away from a poor starting backbone.
Property | Detail |
|---|---|
Starting point | All-X (unknown) sequence, with or without target information |
Structure predictor | AF3-style models (Chai-1, Boltz-2 and AF3 were tested) |
Sequence design | ProteinMPNN, or LigandMPNN for ligand tasks |
Training required | None (no fine-tuning) |
Reported runtime | About 10 s for 100 residues, 130 s for 900 residues |
How Protein Hunter Designs Proteins Through Iterative Hallucination and Redesign
Initialize: an all-X sequence, optionally with target protein, ligand or nucleic-acid context, is passed to the structure predictor.
Hallucinate a fold: diffusion-based hallucination produces an initial backbone.
Redesign the sequence: ProteinMPNN (or LigandMPNN) proposes a new sequence for that backbone.
Re-predict and repeat: the complex is predicted again, and the cycle runs for N rounds or until convergence, producing one sequence per cycle.
Extensions for Cyclic Peptides and Beta-Sheets
The authors add a cyclic positional encoding for macrocycles and a negative helix bias on the Pairformer pair features to enrich beta-sheet content, countering the helical bias of AF3-style models.
Protein Hunter Results: Binder Design, Long Proteins and Small-Molecule Targets
Task | Reported outcome |
|---|---|
Unconditional design | Lengths of about 100 to 900 residues tested with 20 sequences per length; Boltz-2 gave the highest in silico success rate |
Long proteins | Single-pass designs above about 300 residues often had pLDDT below 0.7; cycling kept pLDDT around 0.8 up to 700 residues |
Protein binders vs. RFdiffusion | Boltz-2 gave higher ipTM on 9 of 11 targets; Chai gave higher ipTM on 7 of 11 |
Small-molecule binders | Higher success rates than RFdiffusion-AA, BoltzDesign2 and single-pass mode on all four targets |
B-DNA target | ipTM rose from 0.28 to 0.73 with cycling |
Small-molecule success was defined as backbone RMSD of 1.5 A or less, ligand RMSD of 5 A or less, interface PAE of 1.5 or less and ipTM of at least 0.8. Further examples include a TNF-alpha trimer binder, a heme-binding protein, antibody CDR redesign and MDM2-targeting cyclic peptides of 14, 16 and 18 residues.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with Protein Hunter on Tamarind Bio
Protein binder design: generate candidate binders against a target protein, such as the TNF-alpha trimer example in the paper.
Small-molecule and heme binders: design pockets around a ligand using LigandMPNN-based redesign.
Cyclic peptides: propose macrocyclic binders, as shown for MDM2.
Motif scaffolding and partial redesign: build a protein around a functional motif or redesign part of an existing complex, such as antibody CDRs.
How to Use Protein Hunter on Tamarind Bio
Open the tool: log in to tamarind.bio and select Protein Hunter.
Choose a design task: unconditional design, or binder design against a protein, ligand or nucleic-acid target.
Provide the target: supply the target structure or sequence if the task needs one.
Set design parameters: choose the design length and the number of cycles, and the structure predictor if exposed.
Run the job: submit it and wait for the cycles to finish.
Download outputs: retrieve the designed structures and sequences.
Interpret confidence: rank designs by ipTM, pLDDT and interface PAE, the metrics the authors use.
Use downstream: validate top designs with an independent predictor before experimental work.
Things to Keep in Mind
Initial backbones often have high alanine content; cycling reduces this but does not eliminate it.
ProteinMPNN can fail on noisy or poorly folded backbones.
AF3-style models show a helical bias that limits fold diversity.
The authors list catalytic side-chain modeling and broader secondary-structure sampling as future work.
Reported success rates are in silico predictions, not experimental validation.
Source: Protein Hunter, bioRxiv preprint, DOI 10.1101/2025.10.10.681530 (v2), bioRxiv full text.