BindCraft2 Project Setup

BindCraft2 designs protein and peptide binders using AlphaFold2-guided optimization, ProteinMPNN sequence redesign, and structural validation. Ariax brings these steps together in a managed campaign, with a setup form, project controls, results tables, structure previews, and downloads. Read the BindCraft2 announcement for the background and what has changed from BindCraft.

BindCraft2 is free to use on Ariax Bio apart from compute costs. Academic and commercial customers do not need a separate BindCraft2 commercial license to run projects on Ariax.

Start with a small campaign#

1. Choose a design format in the sidebar or New Project, then select BindCraft2 where an engine choice is offered.
2. Upload a PDB or mmCIF target, or search the Protein Data Bank from the setup page. Select the chains you want the binder to recognize.
3. Keep the format's default design settings for your first run. Add hotspot residues if you have a specific binding site in mind.
4. The form starts with 10 final designs. Set Maximum Trajectories if you want to cap the number of attempts in a pilot; leaving it blank does not impose an attempt limit.
5. Review GPU preferences and the estimated hourly cost. Name the project, then select Start BindCraft2.

An accepted design is a candidate that passed the campaign's computational filters. One design attempt can fail before producing an accepted candidate, so requesting 10 final designs does not mean the engine will run only 10 attempts. An attempt limit bounds the search, not its total cost or duration.

Choose a format#

The 7–20 cyclic-peptide range is the starting choice for a new setup. Cloning preserves an explicitly recorded range.

Length ranges are inclusive. Set the minimum and maximum to the same value for a fixed length. Scaffolded formats derive their lengths from the scaffold and editable regions rather than the de novo length fields.

Homo-oligomer and Multidomain are specialized choices within BindCraft2 setup. Homo-oligomers use identical binder copies, with lengths specified per copy; multidomain designs place multiple domains in one chain. They are not separate engines. Helicon design remains a BoltzGen workflow.

For a custom scaffold, choose Custom scaffold, upload its PDB/mmCIF file, and specify Editable Residues. The field supports the engine's residue-selection syntax, including variable-length loop edits such as A52-57(5-7). Append ! to hold a span, for example A35-40!. It removes the DESIGN flag and keeps the original length. Any span that removes DESIGN, including -DESIGN, must retain that length; A35-40(5-7)! requests a resize and fails validation. A plain span marks paratope residues; * removes the contact flag for framework residues, and +CONTACT restores it. Keep bundled editable-region defaults unless you have a reason to change them. Review the resulting chain sequences carefully before preparing antibody constructs for synthesis.

Prepare targets and binding sites#

The first target must be a PDB or mmCIF structure. Standalone sequence-only projects are not yet supported in Ariax. Additional targets can use structures or FASTA files; for FASTA, select the intended record. A sequence target has no structure preview or structure-based hotspot picker.

Use the target selector to move between uploaded targets. Give each a recognizable name, select its chains, and review the structure preview. You can use Prep Inputs to inspect and trim a structure before uploading it. Keep enough structural context to represent the binding site, and check the proposed pose against the full biological assembly before selecting designs for experiments.

• Hotspots identify residues you want the binder to engage. Use input residue numbering, for example A54,A56,B12-16. Unprefixed numbers refer to the first selected chain. You can also select residues in the structure viewer.
• Coldspots identify regions you want the binder to avoid. They are an optional targeting setting, not a guarantee of selectivity in an experiment.
• Objective: Target promotes binding to that input. Multiple positive targets ask the same binder to work across the supplied targets or states.
• Objective: Detarget discourages binding to a supplied off-target. Keep at least one positive target and choose off-targets that answer a meaningful biological question.
• Weight controls a target's relative contribution to the design objective. Leave the default unless deliberately balancing a multi-target experiment.

Additional target options include FASTA cropping and detargeting controls. Open these only when relevant; the form checks whether they are compatible with the selected inputs and objectives.

Defaults and optional settings#

BindCraft2 Design Settings contains the main decisions: format, scaffold or binder length, requested final designs, and optional attempt limit. The default presets set the optimization and filtering behavior for that format.

Optional Design Objectives exposes the properties supported by the active engine, including Forced targeting and Humanize. Coordinate initialization appears under Advanced Settings.

In v1.0.4, Hard target (hard_target) seeds re-prediction from the trajectory's folded pose. Large complex (large_complex) also seeds gradient stages from the available coordinates; a de novo binder starts at the origin. Ariax accepts the deprecated initial_guess and bigbang aliases for new submissions and collapses each alias with its canonical spelling. Recorded historical settings retain their original spellings. An older active engine displays the legacy choices advertised by its capabilities.

Induced fit and Fold switching compare free and bound conformations. These objectives have compatibility restrictions; some require one target, exclude detargeting, or cannot be combined with a scaffold. Let the setup form's validation guide the supported combinations.

Forced targeting reshapes every target structure around its hotspots, including off-targets. It therefore needs a structure file with hotspots for each target, and it cannot be combined with detargeting or FASTA targets. To use it, remove the off-target; to keep the off-target, turn forced targeting off.

These settings optimize computational proxies. For example, Humanize does not establish that a design will be non-immunogenic, and a protease-stability objective does not establish experimental stability. Use only the properties your experiment needs.

Advanced Settings contains optional optimization, redesign, validation, output, and metric-filter overrides. Blank fields retain the engine's defaults; zero is an explicit value, not another way to request the default. Changing a threshold can change which candidates are accepted. A campaign with easier filters should not be compared directly with a stricter campaign on accepted-design count alone.

Amino-acid preferences accepts sparse relative propensities for editable binder residues. Leave blank to inherit the selected native preset; 0 excludes, 1 explicitly removes the preset preference for that amino acid, values below 1 discourage, and values above 1 favor. For example, C: 0; W: 0.4 excludes mutable cysteine and discourages tryptophan. These weights are not percentages or count limits and do not change target or fixed scaffold residues. Removing an entry restores preset inheritance, which differs from entering 1. Overrides survive format changes and cloning; resuming requires the same scientific settings. If you exclude 19 amino acids, explicitly give the remaining amino acid a positive propensity because its preset may otherwise exclude it.

The setup page shows the active engine's version and source identity when the server supplies them. Optional controls appear when the active generation advertises support. Confidence floor (advanced.confidence_floor) accepts a finite fraction from 0 through 1; v1.0.4 uses 0.7 when omitted. Conformational comparisons weight a residue by the smaller confidence from the bound and unbound predictions above that floor. At 1, every residue receives zero weight, so these metrics have insufficient support and are unreported.

Mutation gate draws (advanced.mutate_gate_draws) accepts an integer from 1 through 8. Leaving it blank preserves the native default of 1. The engine evaluates each prediction draw at the mutation gate and keeps the draw with the lowest weighted loss. Scores belong to that selected draw. Increasing the count adds prediction work and can increase compute costs.

Trajectories before desperation controls when recovery starts after consecutive trajectories without an accepted design. Leave it blank for BindCraft2's 750-trajectory default, including for VHH. Entering 0 starts the first recovery rung immediately. The counter resets after an acceptance, but a zero threshold keeps that first rung active. The switch and threshold are separate from autotuning.

BindCraft2 can adapt its search as a campaign runs. The trajectory CSV's autotuned field records per-attempt changes, including changes that make a difficult search more permissive. Review those records when comparing campaigns or selecting candidates. The recorded campaign settings and individual trajectory records answer different questions: what you requested, and how a particular attempt ran.

Before launching, open Submission Summary to review the scientific settings and explicit overrides. After launch, Project Settings → Additional settings → Recorded settings provides the project's recorded configuration. Use Clone & Reuse to start a separate experiment with revised settings; editing GPU preferences does not change the scientific design task.

GPU preferences and costs#

Choose Performance or Cost to rank eligible GPU offers. Allowed GPUs limits the classes Ariax may use; choices without enough memory for the current campaign are excluded. Larger structures, longer binders, extra targets, and additional binder chains can increase memory requirements.

Turbo Mode allows a campaign to use 2, 4, or 8 GPUs when a matching provider allocation is available. BindCraft2 assigns native workers across the visible devices automatically. Turbo can shorten elapsed time, but the combined hourly cost is higher when more resources run. The displayed hourly estimate is not a fixed price for the completed campaign. Scientific rejection rates and provider availability also affect how long a run takes.

If no compatible resource can be allocated, the project pauses with an explanation. Review GPU preferences, available credit, and the displayed reason before restarting. Do not loosen scientific settings merely to clear a compute-selection problem.

See Billing & Credits for account funding and usage charges.

Monitor and control a project#

The project page uses the familiar Project Settings, Project Status, Resource Usage, and Results sections. Accepted-design progress measures progress toward the requested count; it is not an estimate of time remaining.

Under Project Settings, expand Campaign engine and source to read the version and source identity recorded for that campaign. Missing values appear as Unknown. The setup page's active engine metadata applies to new campaigns. A pause with runtime_generation_unavailable means continuation needs a proven compatible runtime for the recorded generation. Contact support to restore that path, or choose Clone & Reuse to submit a new campaign.

Project status is running, paused, aborted, failed, completed, or deleted. A latest job can separately be pending while waiting to start. A Campaign phase, such as planning or designing, describes what the engine is doing; it is not a separate project status.

• Pause asks the running campaign to stop and save its progress. Allow time for results to finish saving.
• Restart continues a paused campaign from compatible saved progress with its recorded scientific settings and source generation. Resolve the pause reason first, for example by adding credit or adjusting GPU preferences. Ariax checks compatibility before allocating compute; missing historical evidence can prevent continuation.
• Abort ends the campaign. Use Pause when you intend to resume it.
• Clone & Reuse opens a new setup form populated from the project and submits a new campaign against the active engine. Review its version and preserved settings before launch. A saved setting that the active generation cannot support stays visible for review and must be cleared to submit. Use Restart on the original project to continue its saved campaign.
• Delete is available when the project is eligible for deletion. Download anything you need before deleting a project.

Provider interruptions can trigger recovery onto replacement compute. Work completed after the most recent saved checkpoint may need to be repeated. Check the project message and Resource Usage for the current state; avoid submitting another copy of the campaign while recovery or final saving is in progress.

Read results and CSV files#

The Results tabs show three stages of the design process:

The page reads these CSVs into paginated tables. Expand a design's name to inspect its sequences and recorded measurements. Where several structures are available, choose the relevant target complex or binder structure, then select View to open the structure viewer or download the file. Select Refresh Results to request the latest published results.

Each tab has Download CSV for the underlying table, including columns beyond the compact on-screen summary. A table can be empty before its stage has produced output. A completed search can also have no accepted designs if it reached its attempt limit without a candidate passing the filters.

Useful measurements include:

Expand Conformation and backbone beneath a candidate's name to read the reported measurements. Missing conformational values indicate unavailable states or insufficient support. Historical campaigns use the definitions and explicit filter thresholds recorded by their engine. A new preset leaves deliberately supplied legacy scientific settings intact. Interpret conformational RMSD against the campaign's filters because a conformational objective can require a minimum movement.

These v1.0.4 definitions follow the pinned metric implementations, confidence weighting, and campaign configuration. The induced-fit preset sets the 5 Å optimization target.

For multi-target campaigns, CSV metric cells can contain semicolon-separated values aligned with the targets column. Preserve empty entries when parsing the CSV, and do not average positive-target and detarget scores together. Use a CSV reader or spreadsheet import rather than splitting lines on commas: quoted fields can themselves contain commas.

Rank is not a probability of experimental success. Inspect the interface, target accessibility, glycans, membrane context, and intended assay geometry before choosing sequences for synthesis.

Files, archives, and logs#

Project Files lets you browse Inputs and Results, download individual files, and preview supported structures. Preserve the CSVs and campaign_metadata.json alongside the structures so the shortlist remains connected to its settings and measurements.

Download Results Archive collects the published project results into one archive. Inputs and logs can be downloaded separately. The download dialog supports a browser download or a command/link for tools such as cURL, wget, and aria2c. If a signed link expires, reopen the dialog to get a fresh link. An archive may not be available until the campaign has saved and published its results.

The Logs panel displays the available campaign log and offers Download Logs. It is not limited to a 200-line preview. If the job has not started, there may be no log yet; use Load Logs or Refresh Logs to check again.

Troubleshooting#

• The target will not load: check the file format, selected chains, and residue numbering. The first target must have a structure; a FASTA-only project cannot be launched yet.
• No designs have been accepted: check Trajectories and Refolded Candidates to see whether the search is still progressing or repeatedly failing a particular filter. Review the target and intended site before relaxing thresholds.
• The project is paused: read the reason, address credit or GPU eligibility issues as applicable, then use Restart. Wait for saving to finish if controls are temporarily disabled.
• The submission response was lost: open the dashboard and check whether the project was created before attempting another submission.
• A download is unavailable: refresh results or reopen the archive dialog. A result cannot be downloaded until it has been published by the running job.

For help, contact info@ariax.bio with the project ID and a description of what you expected. For other design approaches, see Project Types.

Saved structures#

New campaigns enable save_design_trajectory, save_failed_trajectories, save_failed_refolds, and save_binder_monomers in protocol_config.advanced. Choose these in the optional Saved structures section, or set individual JSON values to false. Accepted and ranked structures are always saved. Trajectory saving keeps a terminal prediction per predicted state, including early exits when a prediction exists. Binder monomer saving controls refolded monomers; terminal trajectory output may also contain monomers. Frames, animations, and relaxation remain separate settings.

The Trajectories tab initially shows only records with published structures, with filtered and overall counts. Clear Only show trajectories with structures to see every record. This does not change attempt/progress counts or full native CSV downloads. A saved structure does not imply scientific acceptance, and early exits may have no scored metrics. Historical unsaved structures cannot be recovered by changing this display filter.

API callers can opt in with view=diagnostics&structures_only=true. Filtering happens before pagination; restart pagination after changing the filter or publication. Omitted/false preserves the full table. Older outputs without a manifest use bounded directory discovery; if discovery exceeds its limit, retrieve the full CSV and artifacts instead.

Whole-chain hotspots and coldspots#

Enter A to select all protein residues in selected chain A, or combine chains and numbered spans, for example A,B12-16. The structure viewer highlights the actual residues, including chains with gaps in their original numbering. Clicking residues refines the selection into compact numbered spans.

The saved configuration retains your shorthand. Ariax validates the input and records the chain's author residue bounds, then expands the shorthand for native BindCraft2. This works through the browser and CLI/API without changing structure bytes. Unknown or unselected chains are rejected. Native restrictions on negative numbering and insertion codes still apply. Whole-chain shorthand applies to structured-target hotspots/coldspots, not FASTA targets or scaffold edits.

Multi-GPU logs#

The project Logs panel keeps the combined Campaign log and offers each published GPU 0, GPU 1, and subsequent GPU log in its source selector. Refresh and Download Logs apply to the selected source. Individual logs appear after synchronization; a single-GPU campaign may only have the Campaign log. Log access uses the project's existing permissions and sanitization. Agents can discover the same worker logs through the existing job logs API.