User Guide

1. Go to New Job.
2. Enter a 4-character PDB code, or upload a structure file (.pdb, .cif, or .mmcif, up to 10 MB).
3. Once the structure loads, pick the chain you want to design against from the dropdown. Each option shows the chain's length and a short description (e.g. protein name) to help you pick the right one on a multi-chain structure.
4. If the structure is a complex, you'll see the 3D viewer load the whole structure - use it in the next step to help pick hotspots.

Target hotspots are the residues on your target's surface you want a binder to engage directly. You can select them three ways - all three stay in sync:

• Click residues in the 3D viewer. Clicking a residue on your selected chain adds it to your hotspot list; clicking again removes it. If the structure is a known complex, residues that already contact another chain are highlighted as a hint - a common starting point, not a guarantee of a good hotspot.
• Type or paste residue codes, e.g. A54, A62, A99. Each one is checked against the real structure as you enter it - an invalid or out-of-range residue is rejected with an explanation rather than silently accepted, so you don't waste a run on a typo.
• Pick from a searchable residue list, with optional filters for "surface-exposed only" and "known interface only."

You can leave this empty if you only want to design against the interacting fragment (see below) - but at least one of target hotspots or fragment hotspots is required.

Every job also needs a short interacting fragment - a peptide or protein region already known to engage your target (the natural partner, or a known ligand/fragment). Paste its sequence (raw amino acids or FASTA). You don't need a solved structure for this fragment - the sequence is enough; the pipeline works out its 3D binding geometry against your target automatically as the first step of the run.

If you want to design a binder that clamps the fragment itself (rather than, or in addition to, attacking the target), specify which positions on the fragment matter - using the fragment's own numbering: position 1 to its length, counting from the start of the sequence you pasted (not the residue's number in its source protein). Optionally prefix each with the expected amino acid, e.g. 4, W4. (For example, in the MDM2 / p53 case on the How It Works page, p53's key Phe19 is the 3rd residue of the pasted 13-mer fragment - so you'd enter it as 3 or F3, not 19.)

Reminder - how track selection actually works:

Compute cost scales with how many sides you select - designing against both roughly doubles it versus one side alone. The cost estimate in step 7 always reflects your actual selection.

Every designed peptide follows the same basic shape: two fixed cysteines forming the ring-closing disulfide bond, a variable stretch between them that forms the binding surface, and one residue outside the ring left as a chemical handle (useful later for attaching a dye, linker, or similar).

• Binding-surface residues - leave every position as "unknown" to let the pipeline fully machine-design the binding surface (the default and the right choice for most campaigns), or lock specific positions to a chosen amino acid if you already have a reason to fix something there. 5–20 positions.
• Terminal handle residue - the single residue at the far end, outside the ring. Defaults to lysine, whose side chain is a convenient attachment point for a dye, linker, or payload later on. Pick a different amino acid if you have a specific conjugation chemistry in mind, or choose X — let AI choose to leave the position to the design model. Choosing X gives up that guaranteed attachment point: the model picks whatever suits the structure, and most amino acids offer nothing convenient to attach to. Cysteine is never offered here - the two cysteines in the ring are reserved for the disulfide bond.
• Number of designs - how many candidate backbones to sample per side selected. This is the main lever on both result diversity and cost - more designs means a bigger, more thorough search, and a bigger bill.

Before you can submit, you'll see an estimated cost range against your current balance. If your balance can't cover it, top up before the Run button unlocks - this is a real, server-checked gate, not just a UI suggestion, so it will also catch you at submit time even if you somehow bypass the up-front warning.

Once submitted, you'll see live progress through each stage of the pipeline (roughly: establishing the interaction geometry → generating backbones → designing sequences → verifying folds → scoring and ranking). You can close the tab and come back later - the job keeps running and you'll see the current state when you return.

When a job completes, you get a ranked table of candidates. For each one:

• Composite score - the overall rank-by number, 0–1, higher is better.
• Structural confidence and interface confidence - how sure the model is that the candidate actually folds correctly and actually forms a real interface with the target (not just a peptide that folds fine on its own).
• Predicted binding strength - an estimated binding free energy; more negative means a predicted stronger binder.
• Which side it came from - target-binder or fragment-binder, if you ran both.
• Ring closure - whether the candidate's disulfide ring actually closed in the validated fold. Almost every candidate closes cleanly; a candidate flagged here didn't, and is worth inspecting before you commit it to synthesis. This is a check on the structure, not a change to its rank.

If a whole batch's predicted binding strengths are too close together to be meaningful, you'll see a clear note that the ranking for that batch is driven by structural and interface confidence instead - never a fabricated affinity ordering.

View any candidate's 3D structure directly in the browser, and download individual structures or the full ranked table as a spreadsheet.

From the results table, select one or more candidates (checkboxes) and choose:

• Relax selected - runs a physics-based energy minimization on just those candidates, settling them into a more realistic geometry.
• Cyclize selected - attempts to close a second ring between each candidate's two loose ends (in addition to its existing disulfide ring), for extra protease stability. Only candidates whose geometry actually supports this are processed - look for the Cyclizable badge on each row; a candidate without it isn't a worse binder, its ends just aren't positioned for this particular refinement. A cyclized result's reported binding strength is carried over from its original design (this step is validated by physics, not re-scored for affinity) - that's called out wherever the result is shown.

Both refinements are optional, run only on the candidates you pick (never your whole batch automatically), and are billed separately from the original job - you'll see a cost preview for each before confirming.

• How It Works - the science and design logic behind the pipeline
• AI Co-Scientist Team - the specialist panel that reviews and discusses your results