Workflow#

RFD4-Proteina has a lot of pieces. To run a basic example, start with the Quickstart.

But we also ship the tooling to run complete design campaigns. That includes refolding with structure prediction oracles, which empirically correlate with wet lab success [1], ddG computation with TMol, a GPU accelerated Rosetta [2], and many more structural and sequence based metrics. You can also search over those metrics, biasing the distribution you sample from toward the properties you want.

The diagram below lays out the general flow from input to output, and shows the ecosystem around the base model. Click any piece to open its documentation. Contributions welcome.

RFD4-Proteina design and guidance workflows Author an annotated CIF, then choose one of three linked model modes: standard generation from noise; partial diffusion or inverse folding from a supplied structure; or metric guidance from a differentiable objective. Optional refolding checks designs. Search loops candidates back, while metrics compare the input, design, and refold and drive named filtering gates. YOU SPECIFY ANNOTATED CIF structure and intent in one file THREE WAYS TO MAKE ONE Agent YAML Python RFD4-PROTEINA THREE WAYS TO RUN IT standard generation from noise partial diffusion or inverse folding metric guidance differentiable objective PRE-ORACLE FILTERS which designs are worth refolding REFOLD OPTIONAL supports AlphaFold2, AlphaFold3 and OpenFold3 SEARCH branch at checkpoints, keep what scores best, feed the survivors back SEARCH PRUNING prunes on any metric, a refold included PASS CRITERIA which of those counts as a hit METRICS input motif RMSD scRMSD TMol ddG H-bond counts RASA adherence PoseBusters … custom metrics YOU GET BACK OUTPUTS annotated designs metrics table

References