PXDesign and Protenix

PXDesign is a model suite for de novo protein-binder design. It combines a diffusion generator (PXDesign-d) with Protenix and AlphaFold2-IG (AF2-IG) confidence models to score and rank generated binders. Across seven benchmark targets, PXDesign delivers 17–82% nanomolar hits on six.

On Nuvolos, pxdesign is pre-installed in a dedicated conda environment along with all required model weights for Protenix and AlphaFold2. No additional installation or weight downloads are needed.

Quick Start

This guide walks you through a complete binder-design task using the bundled PDL1 demo input.

Step 1: Open a Terminal and activate the environment

Launch a Terminal application on Nuvolos and activate the pxdesign conda environment:

conda activate pxdesign

All subsequent commands assume this environment is active.

Step 2: Prepare the input YAML

Create a YAML file that describes your design task. You can copy the template below, or use the pre-built demo file at /files/example/PDL1_quick_start.yaml.

target:
  file: "/files/example/5o45.cif"       # Path to the target structure (CIF or PDB)
  chains:
    A:
      crop: ["1-116"]                    # Region of chain A to keep
      hotspots: [40, 99, 107]            # Interface residues that guide diffusion
      msa: "/files/example/msa/PDL1/0"  # Pre-computed MSA directory (recommended)

binder_length: 80                        # Length of the binder to design (residues)

Tip: Before running a full job, validate your YAML with:

pxdesign check-input --yaml /files/example/PDL1_quick_start.yaml

Step 3: Run the pipeline

Execute the full pipeline in extended mode (Protenix + AF2-IG evaluation). The command below runs 10 samples as a quick test:

pxdesign pipeline \
  --preset extended \
  -i /files/example/PDL1_quick_start.yaml \
  -o /files/example/test_run \
  --N_sample 10 \
  --dtype bf16 \
  --use_fast_ln True \
  --use_deepspeed_evo_attention True

Flag	Recommended value	Notes
--preset	extended	Full pipeline with AF2-IG + Protenix evaluation
--N_sample	10 (test) / 10 000+ (production)	More samples → higher chance of finding nanomolar binders
--dtype	bf16	Use fp32 on older GPUs (e.g., V100)
--use_fast_ln	True	Optimized LayerNorm kernels; generally safe to keep enabled
--use_deepspeed_evo_attention	True	DeepSpeed Evo attention for Protenix; requires CUTLASS

First-run notice: The very first execution performs one-time kernel compilation. Subsequent runs are significantly faster.

Step 4: Check the results

Results are written to <out_dir>/design_outputs/<task_name>/. Open summary.csv to see all ranked binder candidates:

/files/example/test_run/design_outputs/PDL1_quick_start/
├── summary.csv                  ← Master ranked list with all scores
├── server_xx_mode.png           ← Diagnostic plot showing task difficulty
├── task_info.json               ← Run metadata
├── orig_designed/               ← Raw backbone structures from the diffusion model
├── passing-AF2-IG-easy/         ← Designs passing the AF2-IG-easy filter
└── passing-Protenix-basic/      ← Designs passing the Protenix-basic filter

The summary.csv contains boolean *-success columns for each filter. Use these to quickly select candidates:

Filter	Criteria
AF2-IG-easy-success	ipAE < 10.85, ipTM > 0.5, pLDDT > 0.8
AF2-IG-success	ipAE < 7.0, pLDDT > 0.9
Protenix-basic-success	binder ipTM > 0.8, binder pTM > 0.8
Protenix-success	binder ipTM > 0.85, binder pTM > 0.88

Example files

The /files/example/ directory contains everything needed to run the PDL1 demo out of the box:

Path	Description
/files/example/PDL1_quick_start.yaml	Ready-to-use input configuration
/files/example/5o45.cif	PDL1 target structure (mmCIF)
/files/example/msa/PDL1/0	Pre-computed MSA for chain A

Further reference

For detailed documentation on input preparation, running modes, output interpretation, and wet-lab validation workflows, refer to the PXDesign README on GitHub.