Describe your design to an agent#
Tell your coding agent what you want to build. It interviews you about the parts you left open, checks the structure for the usual traps, and hands back a model ready annotated CIF.
How you state the intent is up to you. A paragraph, lab note, source structure, existing annotated CIF, or condition-spec JSON/YAML can all provide evidence. The agent examines what you already have, asks about unresolved choices, and edits the annotated CIF directly with the Python API.
The conversion is not the product. The interview is. A structured input records fields whose material meaning still needs confirmation, while prose commonly leaves choices unstated. Either way the job is to work out what you meant and encode that, rather than treating an artifact as a substitute for intent.
What to expect#
Questions first. The agent opens the structure rather than reasoning about it, then restates each thing it plans to do as intent and asks. The pattern is “I read this as X, is that right?” Your answers determine the output, so take them seriously.
An intent plan. It separates what you stated, what the artifacts show, what it inferred, and what still needs a decision. This is ordinary language, not a generated JSON spec.
A built or edited CIF. It uses the Python AtomWorks API directly.
A report. Everything worth a second look, described in terms of your design rather than stack traces.
What it asks about#
The agent asks about the choices material to your design. It uses your artifacts to avoid generic questions and does not make you decide conditions that are irrelevant to the requested task.
Length. Is that range meant to be sampled per design, or did you mean one fixed length?
What to condition. Split in two, always. Do you want the model to know where an atom is, or what it is, or both? For a ligand, all atoms or a catalytically relevant subset?
What to index. Unindexed means the model chooses where in the chain a residue goes. Should your unindexed fragments float independently, or stay rigidly tied to each other?
Ligand pose, charges and bonds. Whether the pose is given or invented, what the formal charges should be, and whether the inferred bond graph is right.
Everything preprocessing changed. So you can catch it before it costs you a run.
Gotchas it looks for#
These are the ones that bite people. Most are silent.
What happens |
Why it matters |
|---|---|
Fixing one sidechain atom also reveals that residue’s identity |
The atom37 slots fingerprint the amino acid. Backbone only conditioning on |
A length written as a range gets resampled on every design |
|
The length counts only what gets designed |
180 plus a four residue motif gives you 184 residues, not 180 |
Leaving ligand coordinate conditioning unset conditions no pose |
The model invents the ligand’s pose and its conformer. It gets identity, bonds and charge only |
A |
The model embeds formal charge per atom, so unknown is a question, not a default |
PDB-derived ligands often arrive single bonded |
Aromaticity and bond order are lost, and the model reads both |
A metal with no bonds is normal, and still a decision |
The agent surfaces it and asks. It never picks coordination partners by distance, because a nearby atom might be a real bond or an incidental contact and only you know which |
Missing atoms require intent review |
Keep complete residues and express tip-atom conditioning with masks. Raw unresolved structures may need preparation; custom components are valid when their explicit atoms, bonds and charges are complete |
Token types can have different indexing capabilities |
The agent preserves your requested indexing, runs the checked-out validator, and reports any unsupported combination instead of silently changing it |
Covalent polymer-to-ligand bonds must remain present |
The annotated CIF preserves and validates the bond explicitly |
Check the result. An agent can misread intent. Before you run the model, inspect the annotated CIF and its validation report. Confirm that the conditioning says what you meant, including fixed atoms, sequence, ligands and unindexing.
What it will not do#
Guess its way past a real ambiguity. It asks instead.
Silently choose structural preparation. It tells you which preset it used and asks you to confirm material atom, component, and bond changes.
Silently drop covalent bonds, charges or PTMs.
Freeze a variable source template unintentionally. For intentionally materialized child CIFs, follow the expandable-segment guidance and workspace policy.
Claim a check passed that it did not run.
Create a JSON/YAML intermediate unless you explicitly ask for one.