Introducing the Tamarind Protein Design Agent

The Protein Design Agent is presented as a hosted workflow for de novo binder campaigns. It can help select a target region using literature and structural context, then coordinate design methods, candidate optimization, and quality checks.

The workflow combines several design approaches, including backbone generation with methods such as RFdiffusion, BindCraft, and BoltzGen, followed by sequence design. Antibody and nanobody formats are included alongside other binders.

A campaign proceeds through epitope selection, diverse backbone generation, inverse folding, inexpensive candidate filters, and structure-based scoring. The article describes evaluating candidates with multiple co-folding predictions and selecting a diverse set for each target.

The post emphasizes long-running GPU work and coordination across many candidate-generation steps. Tamarind's hosted environment manages that orchestration so the scientist can focus on the design objective and review the resulting candidates.

These computational stages prioritize candidates for further study; binding, specificity, and developability still require experimental assessment.

Here's how it works

A design campaign starts by choosing a target region with help from literature, structure context, and molecular tools. It then moves through generation, sequence design, filters, structure scoring, and selection.

A typical campaign

1. Select a biologically relevant epitope using target context and available evidence.
2. Generate diverse candidate backbones with several design methods.
3. Assign sequences to those backbones with inverse-folding models.
4. Apply low-cost filters for sequence liabilities and monomer foldability.
5. Compare candidates using multiple co-folding predictions and pose-consistency measures.
6. Choose a diverse set of candidates for each target for follow-up.

Infrastructure and safety

Long-running campaigns need compute budgets, orchestration, and clear stop conditions. Tamarind describes handling that operational work around its model tools. The article also calls out safeguards for binder-design use; computational candidates still need appropriate review and experimental validation.