8. PXDesign Project Setup

Looking for BindCraft or BoltzGen? See 6. BindCraft Project Setup or 7. BoltzGen Project Setup.

8.1 Overview

PXDesign is a diffusion-based engine for de novo miniprotein binder design, developed by ByteDance's AI for Science team (the creators of Protenix). It pairs a Diffusion Transformer (DiT) backbone generator with a dual-filter ranking system using both Protenix and AF2-IG confidence models, delivering experimentally validated hit rates of 17-82% across diverse protein targets.

PXDesign is fully open-source under the Apache 2.0 license. Read our announcement for experimental benchmarks and comparisons with other design engines.

8.2 Inputs

• Target structure: Upload a .pdb or .cif file, or search RCSB by PDB ID or keyword directly from the setup page.
• Chain selection: After uploading, select which chains to target. Unselected chains are ignored.
• Hotspot residues: Optionally specify residues to focus binding. Enter comma-separated residue numbers (e.g., 45,46,47) or ranges (e.g., 10-20). Leave empty to let PXDesign choose binding sites automatically.
• Crop ranges: Optionally restrict the target to a specific residue range per chain (e.g., 50-150). Useful for large structures where you want to focus on a particular domain.
• Bind Anywhere: Toggle per chain to target all residues on that chain for binding.

Sequence handling

PXDesign requires the full amino acid sequence for each target chain. Sequences are automatically extracted from PDB SEQRES records or mmCIF _entity_poly_seq tables when available. If the structure has gaps or missing sequence data, you will be prompted to enter the sequence manually.

Hotspot and crop numbering

When specifying hotspot residues or crop ranges on the setup page, use the residue numbering from your input file (PDB auth_seq_id numbering). PXDesign automatically maps these to canonical mmCIF indices internally during processing.

8.3 Design Settings

Binder Length Range

Configure the minimum and maximum length of designed binders:

• Default range: 65-150 amino acids
• Minimum: Must be greater than 0
• Both values must be positive integers, and min must be less than or equal to max.

Number of Designs

Set the number of diffusion designs to generate. The default is 100. For production campaigns, PXDesign authors recommend generating 10,000+ designs to maximize the chances of finding high-affinity binders after dual-filter ranking.

8.4 GPU Selection & Turbo Mode

• Mode: Choose Performance (fastest GPUs) or Cost (cheapest GPUs).
• Primary GPUs: Select at least one of H100, A100, or L40/L40S (required for availability).
• Additional GPUs: H200 and RTX 6000 Ada are available as optional selections.
• Turbo Mode (2x/4x/8x): Parallelize your campaign across multiple GPUs for faster completion. Pricing is transparent: base GPU rate + $1/hr per extra GPU. Turbo is especially useful for large campaigns (1,000+ designs).

8.5 Validate & Start

Click Validate Settings to check your configuration. Validation ensures:

• Target structure is properly parsed with valid chains
• At least one chain is selected
• Sequences are available for all selected chains
• Hotspot and crop syntax is valid with no out-of-range residues
• Binder length range is valid
• GPU selection requirements are met

Once validated, click Start PXDesign to launch. Your job appears on the dashboard immediately.

8.6 Pipeline Stages & Monitoring

PXDesign runs a multi-stage Extended pipeline:

1. Diffusion Generation — The DiT backbone generator creates candidate binder backbones conditioned on your target structure and hotspot constraints.
2. AF2 Evaluation — Each candidate is evaluated by AlphaFold2-based confidence scoring (AF2-IG). Designs passing confidence thresholds are retained.
3. Protenix Evaluation — Retained candidates are independently evaluated by the Protenix confidence model, capturing binders that AF2-IG alone might miss.
4. Final Ranking — Designs are ranked by combined confidence scores from both evaluators, producing a final set of high-confidence candidates.

Progress is visible on the project status page, with per-stage indicators showing which pipeline step is active and percent complete. Ariax runs PXDesign via a cloud-optimized fork with full resume capability, so if a multi-hour job is interrupted, it picks up where it left off.

8.7 Outputs & Results

Results stream to your project page in real time:

• Summary Table — Ranked designs with confidence scores from both AF2 and Protenix evaluators, plus combined metrics.
• Structure Files — Predicted structures available for each design:
  • Diffusion outputs (original backbone predictions)
  • AF2 predictions (AlphaFold2 evaluation structures)
  • Protenix predictions (Protenix evaluation structures)
• Interactive Viewer — Select any design to visualize its 3D structure directly in the browser.
• Archive Downloads — Per-tab .zip archives for bulk download of diffusion outputs, AF2 predictions, Protenix predictions, and final results. For large archives, command-line download snippets (cURL/wget/aria2c) are provided.

8.8 Helpful Links

• Announcement: PXDesign: A New Diffusion-Based Engine for Miniprotein Design
• PXDesign GitHub: https://github.com/bytedance/PXDesign
• Ariax cloud-optimized fork: https://github.com/cytokineking/PXDesign
• Preprint: PXDesign: Fast, Modular, and Accurate De Novo Design of Protein Binders