Tamarind Bio is excited to be working with NVIDIA Healthcare to bring access to RFD4-Proteina on day 0.
Wet-lab validation yielded 84% sub-micromolar binders against human TrkA (158/188 designs) and 31% active phosphotriesterases (89/288 designs). Other highlights include picomolar binders to disordered YAP1 and Tau epitopes, peptide-MHC binders with over 100-fold on-target specificity, and a designed enzyme that catalyzes a step in lenacapavir synthesis.
The authors also propose an in silico, structure-based success comparison, running and AlphaFold3 re-folding process, measuring the pas-rate of clusters of structures.
The model builds on the Institute for Protein Design’s RFdiffusion series and NVIDIA AIR’s Proteina-Complexa. RFD4 can design de novo miniproteins, VHHs, scFvs, TCRs, and macrocycles against a diverse set of target types, including binders to protein targets, small molecules, substrates, nucleic acids and more. Beyond binder design, it supports partial diffusion, motif scaffolding, and atom-level conditioning for enzyme design.
For protein binding, RFD4-Proteina single-shot achieves 9.8% clustered pass rate versus 5.9% for the next-best model BindCraft, with 58-fold higher design throughput. For an Atomic Motif Enzyme benchmark achieves 8.9% clustered pass rate versus 5.7% for AlphaProtein-novo codesign.
Notably, the model outperforms specialized models for tasks like Side-chain packing, inverse folding, and design refinement. In in silico metrics, RFD4 shows improved sequence tasks performance compared to LigandMPNN, along with improving failed designs in a process called design refinement.
Here, the authors take binder design candidates that failed a co-folding based filter and found that they could rescue 100/102 designs at six attempts per design using RFD4.
Fine-tuning and training code is also available, meaning the protocol can be specialized on your data beyond the public data it's been trained on.
RFD4-Proteina is available on Tamarind today, alongside hundreds of molecular design models. Tamarind also offers a streamlined workflow for fine-tuning the model on your own data.