Use RFD4-Proteina Online

RFD4-Proteina: Biomolecular Design Through Latent Reasoning for Binders, Enzymes and More

RFD4-Proteina is a biomolecular design model that reasons in a shared chemical and conformational latent space covering standard and modified amino acids, sugars, DNA and RNA. The project is led by Nathaniel Corley, Kieran Didi, Maxwell W. Kazman and Rafael I. Brent (equal contributors), with roughly 50 further co-authors from institutions including the Institute for Protein Design at the University of Washington and NVIDIA. The corresponding authors are Kieran Didi and David Baker.

The model applies two kinds of reasoning on this latent space. Intrinsic reasoning uses recurrent refinement to allocate more computation to harder denoising steps, and extrinsic reasoning uses search guided by arbitrary scores such as energetics, geometric criteria or a structure predictor. The project page also says the model handles side-chain packing, fixed-backbone sequence design, conformational sampling and design refinement zero-shot, without additional training.

The paper is not yet published; the page lists a preprint as forthcoming. All results below are from the project page.

How RFD4-Proteina Works

• Shared latent space: an autoencoder represents amino acids, sugars and nucleotides in one space, and a decoder reconstructs atoms.

• Energy supervision: the latent space is supervised with force-field energies.

• Adaptive compute: recurrent refinement spends more computation on harder denoising steps, so a higher thinking setting improves quality at the cost of throughput.

• Guided search: designs can be steered by any score, for example energetics, geometry or a structure predictor.

RFD4-Proteina In Silico Results

Pass rates are clustered, meaning only distinct folds are counted, and depend on AF3 refolding criteria.

Protein-binding throughput is 58-fold higher than BindCraft. Raising thinking from low to high lifts the protein-binding pass rate from 7.0% to 9.8%, while throughput halves from 2,716 to 1,363 designs per GPU-hour. For small molecules, RFD4-Proteina passes 77 targets versus 64 for the next best method.

RFD4-Proteina Wet-Lab Validation

Binders

• Calmodulin: 12 of 18 designs bound with apparent KD below 1 µM and no binding without calcium; the 4.24 pM best value is a model-dependent fit and not precisely resolved.

• Peptide-MHC: binders to MART-1 (340 pM), SARS-CoV-1 (710 pM) and Phox2B (1.60 nM) showed over 100-fold on-target preference.

• Disordered peptides: binders for eight peptides, including the reported first de novo binders to YAP1 (112 pM) and Tau (527 pM) epitopes by SPR.

• TrkA: 158 of 188 unconstrained designs (84%) were sub-micromolar, with human-specific and human/mouse cross-reactive binders; a TrkB binder (30.2 nM) and 20 TrkC screening hits were also reported.

• Phosphopeptides (histone H3): binders for pS10, pT11 and pS28, with half-maximal binding of about 138 nM (pT11) and 1.03 µM (pS10) measured on yeast without avidity.

• Glycan shielding: 8 of 9 designs had over 90% oligomannose on occupied sites.

Enzymes

• Phosphotriesterases: 89 of 288 designs were active across 25 of 47 scaffold families, versus none of 192 RFD3 designs; the best reached kcat/KM of 3,600 ± 500 M-1 s-1, a 1.5 × 107-fold rate enhancement.

• Ligase for a lenacapavir synthesis step: 4 of 67 designs formed the amide above background; design 52 converted 79.0% of the (S)-amine versus 8.8% of the (R)-amine. This is presented as an early proof of concept.

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 Biomolecular Design with RFD4-Proteina on Tamarind Bio

• Protein binder design: generate binders to protein targets, evaluated on 24 targets in the project's benchmark.

• Small-molecule binders and enzyme scaffolds: design binding pockets and scaffold enzyme active sites.

• Peptide, phosphopeptide and MHC targets: design binders to disordered peptides and peptide-MHC complexes.

• Search-guided design: steer designs with energetics, geometry or a structure predictor score.

• No local setup: run the model without managing a CUDA 13 GPU environment.

How to Use RFD4-Proteina on Tamarind Bio

1. Log in and open the tool: sign in at tamarind.bio and select RFD4-Proteina.

2. Define your target: upload a structure or sequence for the protein, peptide, small molecule or nucleic acid you want to bind.

3. Choose the design task: pick binder design or enzyme scaffolding, and set any constraints the tool exposes.

4. Set the thinking level: higher thinking raises pass rates but halves throughput, so balance quality against the number of designs.

5. Run the job: submit and wait for the designs.

6. Review and filter: download the designs and refold them with a structure predictor to check them against your criteria before ordering.

Parameters shown on Tamarind may differ from the command-line options, so refer to the tool page for what is exposed.

Limitations

• The paper is not yet published, and structures shown are computational models, not experimental structures.

• In silico pass rates depend on AF3 refolding criteria.

• Some campaigns are small (9 glycan designs, 4 of 67 ligase designs, 3 DNA targets).

• Higher thinking settings halve throughput.

• Phosphopeptide affinities were measured without avidity, and natural readers of these marks report mid-micromolar affinities.

• Enzyme results are early proofs of concept.

Source: RFD4-Proteina project page (rfd4-proteina.org), Corley, Didi, Kazman, Brent and others; preprint forthcoming. Code: github.com/RosettaCommons/RFD4-Proteina; weights: huggingface.co/nvidia/RFD4-Proteina.