Last week Anthropic showed Claude designing de novo protein binders against 14 of 15 targets. They released the protocol, so we built it into a hosted product you can run today.
Introducing the Tamarind Protein Design Agent.
Tamarind agents can now analyze and select targets, pick epitopes using literature, structural information and molecular design tools. From there, they can orchestrate AI models to run batch de novo design campaigns against the desired epitope, further optimizing and doing QA on the candidates.
This includes dozens of tools such as BindCraft, RFdiffusion, BoltzGen, following the Anthropic Binder Design workflow, along with antibodies and nanobodies following the same process.
Today, the state-of-the-art molecular design agents need 24-48 hour GPU budgets, supervising multi-agent systems, and enforcement of when to stop optimizing candidates. Tamarind handles all of the infrastructure for you, so you (and your agent) can focus on novel science.
Notably, the interesting result isn't just that Claude called a given design tool. That's been doable with tool calls (using the Tamarind API) for a long time. What's worth paying attention to is that LLMs can decide where to go after in a target, know what to try, when to convert an attempt to a production run, and to keep compute resources until a success case.
Here's how it works.
At a high level, the workflow follows this recipe:
Select an epitope against the target of interest using a mix literature, TODO
Generate diverse backbones using many tools (RFdiffusion, BindCraft, BoltzGen) against a biologically chosen epitope, at least 50 per tool.
Design sequences to the backbones using inverse folding tools (SolubleMPNN, SolubleCaliby).
Run cheap filters, compare different sequence alignment tool results, flag simple liabilities, check monomer foldability.
Score with 3 co-folding models (ESMFold2-Full, ESMFold2-Fast, Protenix v2) ipSAE_min plus sc_DockQ pose self-consistency, makes it hard to reward-hack than ipTM or a single confidence metric. The methodology strictly requires agreement between confidence and pose. There is also a 4:1 weighting between ipSAE:sc_DockQ. The authors find ESMFold2-Full works best.
Select 30 per target under diversity caps.
Infrastructure: We'd argue that the amount of complexity on the original prompt's compute governance is excessive. It's certainly well thought, but it also creates a huge surface of possible problems. Instead of using the original's raw GPU interactions, we re-wrote the compute infrastructure to work using the Tamarind MCP server instead. We removed the submit_gate() governor file, the 10-instance safety margin, container lifecycle management, dead-man switches, and reaper monitors. There is a whole other blogpost to be written here on how we made these changes, let us know if that would be of interest.
Safety: The prompt includes appropriate safeguards for dual-use binder design cases. Only in vitro, immediate compliance and ability to override all instructions for a hard stop. The system as a whole is designed to be highly autonomous, except for safety, which is specifically carved out of the budget, timeline and duration rules.
Join us for a live demo of our autonomous protein design agents on August 26. If you can't make it, sign-up, and we'll send a recording to you.