Glossary#

This glossary defines the conditioning and inference terms used throughout RFD4-Proteina. For structure parsing nomenclature such as entities, instances, chains, and polymer or non-polymer units, see the AtomWorks glossary.

Conditioning#

Conditioning gives the model information or constraints to preserve while it generates the rest of the sequence and structure. In an annotated CIF, one or more standard annotation columns carry each condition. See Annotations reference for the complete schema.

Indexed and unindexed residues#

An indexed residue has its position in the designed chain exposed to the model. Its C_IDX mask is on. Most residues are indexed by default.

An unindexed residue has its C_IDX mask off. The model knows the residue exists but is free to place it within the generated chain. This is useful for motif scaffolding when the motif’s absolute sequence location should not be fixed.

C_IDXGRP assigns residues to an index group. Residues in the same group retain known relative offsets even when their absolute placement is free. Atomized tokens cannot be unindexed.

RASA#

Relative accessible surface area (RASA) describes how exposed an atom or residue is relative to a reference maximum. C_RAS provides binned RASA conditioning:

0 is unspecified, so no RASA condition is applied
1 is buried, with RASA in [0, 0.02)
2 is intermediate, with RASA in [0.02, 0.1]
3 is exposed, with RASA above 0.1

The training transform computes solvent accessible surface area with the Shrake and Rupley algorithm, a 1.4 Å probe, element radii, and 100 sphere points. RASA conditioning steers surface exposure and the formation of buried ligand pockets.

Common condition aliases#

C_CRD: Fixes selected atom coordinates. Unmasked coordinates are generated.
C_SEQ: Reveals selected residue identities. Unmasked sequence positions are designed.
C_IDX: Reveals residue positions in a chain. Turning the mask off makes residues unindexed.
C_IDXGRP: Groups residues whose relative sequence offsets are known.
C_HOT: Marks hotspot atoms on the target interface that a designed binder should engage.
C_RAS: Requests buried, intermediate, or exposed surface placement.
C_HB: Requests donor, acceptor, or combined hydrogen bond roles for selected atoms.
C_SEC: Requests helix, sheet, or loop secondary structure at selected residues.
C_CENTROID: Supplies a target center of mass for a chain without fixing its individual atoms.
C_CRDGRP: Groups atoms with known relative coordinates but possibly unknown global placement. Group 0 is the absolutely conditioned frame, -1 is unconditioned, and every id >= 1 is one internal coordinate group.
C_INTCRD: Internal coordinate conditioning. It gives a group’s centered, randomly rotated coordinates, so the model knows the internal structure and designs the global placement.

Input formats#

Annotated CIF: The preferred input format. It stores coordinates, chemistry, and model conditions in a standard mmCIF file and can be edited with the Python API.

Outputs#

Provenance: Standard annotations written into an output CIF that record where each atom came from and how the model produced it, such as which input region it originated from or which motif region it was coordinate-conditioned against. The opposite direction from conditioning: conditioning tells the model what to build, provenance records what it actually built. See Output format reference for the full annotation schema.

Generation and search#

Flow matching: The generative process used by RFD4-Proteina to transform noise into a biomolecular sequence and structure over a fixed number of sampling steps.
Vanilla sampling: One uninterrupted generation trajectory. This is the fastest baseline.
Reward guided search: Generation that periodically branches trajectories, scores rolled out candidates with a reward model, and keeps the strongest ones.
Beam search: The reward guided strategy that ships with the model. search.checkpoints controls when candidates get scored, n_branch controls children per surviving beam, and beam_width controls how many candidates survive each pruning step.
Reward: A weighted sum of metric outputs named in each checkpoint’s rewards. Positive weights maximize a metric and negative weights minimize it. Optional clipping limits a metric’s influence past a threshold.
Oracle: An external evaluator used during generation or final evaluation. For example, the AF3 refolding oracle predicts a generated sequence and scores confidence or agreement with the design.
Metrics tag: A named group of metrics selected through configuration. Top level metrics_tags controls output evaluation; search instead lists individual metric providers in each checkpoint’s metrics field.