### Install Protenix from GitHub

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Installs the latest development version directly from the repository.

```bash
pip3 install git+https://github.com/bytedance/Protenix.git
```

--------------------------------

### Clone ColabFold, Prepare Databases, and Install

Source: https://github.com/lucidrains/protenix/blob/main/scripts/msa/step2-get_msa.ipynb

Clones the ColabFold repository, checks out a specific commit, and prepares the necessary databases using the provided setup script. It also includes instructions for adding taxonomy information and installing ColabFold via pip.

```bash
git clone https://github.com/sokrypton/ColabFold
cd ColabFold 
git checkout 9bd39fd8be88ca71c039be9a8000bb0db92f80a2 
# download and prepare the database
# Note: Uniref only version uniref30_2103 has contained taxonomy information
# you can download it from https://wwwuser.gwdg.de/~compbiol/colabfold/uniref30_2103_taxonomy.tar.gz
bash setup_databases.sh 
# If you want to add Taxonomy ID using uniref30_2103 database
# refer to https://github.com/sokrypton/ColabFold/issues/216
# Add the following line after Line 93 in colabfold/mmseqs/search.py
# run_mmseqs(mmseqs, ["convertalis", base.joinpath("qdb"), dbbase.joinpath(f"{uniref_db}{dbSuffix1}"),
#                     base.joinpath("res_exp_realign_filter"), base.joinpath("uniref_tax.m8"), "--format-output",
#                     "query,target,taxid,taxname,taxlineage,fident,alnlen,mismatch,gapopen,qstart,qend,tstart,tend,evalue,bits,cigar",
#                     "--db-load-mode", str(db_load_mode), "--threads", str(threads)])

pip3 install .
pip install colabfold[alphafold]==1.5.5
```

--------------------------------

### Install Protenix

Source: https://context7.com/lucidrains/protenix/llms.txt

Commands to install Protenix from PyPI, GitHub, or local source.

```bash
# Install from PyPI (stable version)
pip install protenix

# Install from GitHub (latest development version)
pip install git+https://github.com/bytedance/Protenix.git

# Install from local source (for development)
git clone https://github.com/bytedance/Protenix.git
cd Protenix
pip install -e .
```

--------------------------------

### Install Protenix from Local Source

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Clones the repository and installs it in editable mode for development.

```bash
git clone https://github.com/bytedance/Protenix.git
cd Protenix
pip3 install -e .
```

--------------------------------

### Install and Activate Conda Environment for ColabFold

Source: https://github.com/lucidrains/protenix/blob/main/scripts/msa/step2-get_msa.ipynb

Installs miniforge and creates a Python 3.10 environment named 'colabfold_search', then activates it. This is the initial setup step for the ColabFold search process.

```bash
# Install miniforge
# https://github.com/conda-forge/miniforge#download
conda create -n colabfold_search python=3.10
conda activate colabfold_search
cd ~ && mkdir workdir && cd workdir
```

--------------------------------

### Install Protenix via PyPI

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Installs the stable version of Protenix using pip.

```bash
pip3 install protenix
```

--------------------------------

### Install Protenix via pip

Source: https://github.com/lucidrains/protenix/blob/main/README.md

Standard installation command for the Protenix package.

```bash
pip install protenix
```

--------------------------------

### Install Pre-commit Hooks

Source: https://github.com/lucidrains/protenix/blob/main/README.md

Commands to install and initialize pre-commit hooks for maintaining code quality.

```bash
pip install pre-commit
pre-commit install
```

--------------------------------

### Install and Verify Protenix

Source: https://github.com/lucidrains/protenix/blob/main/docs/docker_installation.md

Inside the Docker container, install Protenix in editable mode using pip and verify the installation by checking its help message.

```bash
cd /app
pip install -e .

# Verify the installation by checking the help message
protenix --help
```

--------------------------------

### Clone, Build, and Install MMseqs2

Source: https://github.com/lucidrains/protenix/blob/main/scripts/msa/step2-get_msa.ipynb

Clones the MMseqs2 repository, checks out a specific commit version as recommended by the ColabFold readme, builds it with CMake for release, and installs it. The MMseqs2 binary directory is then added to the PATH.

```bash
git clone https://github.com/soedinglab/MMseqs2.git
cd MMseqs2
git checkout 71dd32ec43e3ac4dabf111bbc4b124f1c66a85f1 # As in colabfold readme
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=RELEASE -DCMAKE_INSTALL_PREFIX=. ..
make
make install 
export PATH=$(pwd)/bin/:$PATH
cd ../
```

--------------------------------

### Install Colabfold

Source: https://github.com/lucidrains/protenix/blob/main/docs/colabfold_compatible_msa.md

Install Colabfold with alphafold support using pip. This is a prerequisite for using `colabfold_search`.

```bash
pip install colabfold[alphafold]
```

--------------------------------

### Build MMseqs2 from Source

Source: https://github.com/lucidrains/protenix/blob/main/docs/colabfold_compatible_msa.md

Compile MMseqs2 from source to use with Colabfold. This process involves downloading the source, configuring the build, and installing it.

```bash
wget https://github.com/soedinglab/MMseqs2/archive/refs/tags/16-747c6.tar.gz
tar xzf 16-747c6.tar.gz 
cd MMseqs2-16-747c6/
mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=RELEASE -DCMAKE_INSTALL_PREFIX=. ..
make -j8
make install
```

--------------------------------

### Verify Docker GPU Setup

Source: https://github.com/lucidrains/protenix/blob/main/docs/docker_installation.md

Run this command to ensure Docker is correctly configured with GPU support. It uses a specific NVIDIA CUDA image.

```bash
docker run --rm --gpus all nvidia/cuda:12.6.3-base-ubuntu22.04 nvidia-smi
```

--------------------------------

### Run CLI Prediction

Source: https://context7.com/lucidrains/protenix/llms.txt

Examples of using the protenix pred command for various inference scenarios.

```bash
# Standard inference with the default v1.0.0 model
protenix pred \
    -i examples/input.json \
    -o ./output \
    -s 101 \
    -n protenix_base_default_v1.0.0 \
    --use_default_params true

# Inference with template support (v1.0.0+ only)
protenix pred \
    -i examples/examples_with_template/example_9fm7.json \
    -o ./output \
    -n protenix_base_default_v1.0.0 \
    --use_template true \
    --use_seeds_in_json true

# Inference with RNA MSA support (v1.0.0+ only)
protenix pred \
    -i examples/examples_with_rna_msa/example_9gmw_2.json \
    -o ./output \
    -n protenix_base_default_v1.0.0 \
    --use_rna_msa true

# Lightweight Mini model inference (faster, lower resource)
protenix pred \
    -i examples/input.json \
    -o ./output \
    -n protenix_mini_esm_v0.5.0 \
    --use_default_params true

# Custom inference parameters
protenix pred \
    -i examples/input.json \
    -o ./output \
    -s 101,102,103 \
    -c 10 \
    -p 200 \
    -e 5 \
    -d bf16 \
    -n protenix_base_default_v1.0.0 \
    --enable_cache true \
    --enable_fusion true
```

--------------------------------

### Run Protenix Docker Container

Source: https://github.com/lucidrains/protenix/blob/main/docs/docker_installation.md

Start the Protenix Docker container with GPU access and interactive mode. It mounts the current directory and shared memory for persistent storage and performance.

```bash
docker run --gpus all -it \
    -v "$(pwd)":/app \
    -v /dev/shm:/dev/shm \
    ai4s-share-public-cn-beijing.cr.volces.com/release/protenix:1.0.0.4 \
    /bin/bash
```

--------------------------------

### MSA Header Format Examples

Source: https://github.com/lucidrains/protenix/blob/main/docs/msa_template_pipeline.md

Required header formats for UniRef and UniProt sequences to enable correct taxonomic pairing.

```text
>UniRef100_{hit_name}_{taxonomy_id}/
```

```text
>tr|ACCESSION|ID_SPECIES/START-END [optional description] OS=Species Name OX=TaxonomyID ...
```

--------------------------------

### Define PDB Subset File

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Example content for a subset text file containing PDB IDs, one per line.

```text
6hvq
5mqc
5zin
```

--------------------------------

### Constraint-Guided Prediction JSON

Source: https://context7.com/lucidrains/protenix/llms.txt

Demonstrates how to specify contact and pocket constraints within a JSON input for guided prediction.

```json
[
    {
        "name": "constraint_example",
        "sequences": [
            {
                "proteinChain": {
                    "sequence": "YMLDLQPETTDLYCGGGGSGGGGSGGGGSIQRTPKIQVYSRHPAENGKSNFLNCYVSGFHPSDIEVDLLKNGER...",
                    "count": 1
                }
            },
            {
                "proteinChain": {
                    "sequence": "EVKLVESGGGLVQPGGSLRLSCAASGSIFSINTMGWYRQTPGKQRDLVADISSGGSTKYGDSVKGRFTISRDNTKNTVYLQMNSLKPEDTAVYYCYGLSYSNDDYWGQGTQVTVSS",
                    "count": 1
                }
            }
        ],
        "constraint": {
            "contact": [
                {
                    "entity1": 1,
                    "copy1": 1,
                    "position1": 169,
                    "atom1": "CA",
                    "entity2": 2,
                    "copy2": 1,
                    "position2": 1,
                    "atom2": "C5",
                    "max_distance": 6,
                    "min_distance": 3
                }
            ],
            "pocket": {
                "binder_chain": {
                    "entity": 2,
                    "copy": 1
                },
                "contact_residues": [
                    {"entity": 1, "copy": 1, "position": 69},
                    {"entity": 1, "copy": 1, "position": 72},
                    {"entity": 1, "copy": 1, "position": 126}
                ],
                "max_distance": 8
            }
        }
    }
]
```

--------------------------------

### Protenix Prediction Commands

Source: https://context7.com/lucidrains/protenix/llms.txt

Examples of using the 'protenix pred' command for inference with different base model variants. Specify input JSON and model name.

```bash
protenix pred -i input.json -n protenix_base_default_v1.0.0    # Template + RNA MSA support
```

```bash
protenix pred -i input.json -n protenix_base_20250630_v1.0.0   # Latest training cutoff (2025-06-30)
```

```bash
protenix pred -i input.json -n protenix_base_default_v0.5.0    # Standard base model
```

```bash
protenix pred -i input.json -n protenix_base_constraint_v0.5.0 # With constraint support
```

--------------------------------

### Define Protein Chain Sequence

Source: https://github.com/lucidrains/protenix/blob/main/docs/infer_json_format.md

Example of a proteinChain entry including optional PTMs and MSA/template paths.

```json
{
  "proteinChain": {
    "sequence": "PREACHINGS", 
    "count": 1,
    "modifications": [
      {
        "ptmType": "CCD_HY3", 
        "ptmPosition": 1,
      },
      {
        "ptmType": "CCD_P1L",
        "ptmPosition": 5
      }
    ],
    "pairedMsaPath": "/path/to/pairing.a3m",
    "unpairedMsaPath": "/path/to/non_pairing.a3m",
    "templatesPath": "/path/to/hmmsearch.a3m"
  },
}
```

--------------------------------

### View Preprocessing Help

Source: https://github.com/lucidrains/protenix/blob/main/docs/prepare_training_data.md

Displays parameter documentation for the training data preparation script.

```bash
python3 scripts/prepare_training_data.py --help
```

--------------------------------

### Run Training Demo

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Executes the training script from scratch using the provided bash demo script.

```bash
bash train_demo.sh
```

--------------------------------

### Download and Prepare Training Data

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Set the PROTENIX_ROOT_DIR environment variable to your desired data location and then run this script to download and extract full training data components. Ensure sufficient disk space (1.5T).

```bash
# Set your data root directory
export PROTENIX_ROOT_DIR=/path/to/your/data_root
mkdir -p $PROTENIX_ROOT_DIR

# Download and extract data components using the provided script
# Use --inference_only (default) or --full for training/finetuning
bash scripts/database/download_protenix_data.sh --full
```

--------------------------------

### Download and Configure ColabFold Database

Source: https://github.com/lucidrains/protenix/blob/main/docs/colabfold_compatible_msa.md

Clone the ColabFold repository and set up the necessary databases. This step requires MMSEQS_NO_INDEX=1 environment variable to be set.

```bash
git clone https://github.com/sokrypton/ColabFold.git
cd ColabFold
# Configure database:
MMSEQS_NO_INDEX=1 ./setup_databases.sh <path/to/colabfold_db>
```

--------------------------------

### Model Training from Scratch

Source: https://context7.com/lucidrains/protenix/llms.txt

Train Protenix from scratch using the training script with configurable hyperparameters. Ensure PYTHONPATH and PROTENIX_ROOT_DIR are set.

```bash
export PYTHONPATH="${PYTHONPATH}:$(pwd)"
export PROTENIX_ROOT_DIR="/path/to/your/data_root"

# Initial training from scratch
python3 ./runner/train.py \
    --run_name protenix_train \
    --model_name "protenix_base_default_v1.0.0" \
    --seed 42 \
    --base_dir ./output \
    --dtype bf16 \
    --project protenix \
    --use_wandb false \
    --diffusion_batch_size 48 \
    --eval_interval 400 \
    --log_interval 50 \
    --checkpoint_interval 400 \
    --ema_decay 0.999 \
    --train_crop_size 384 \
    --max_steps 100000 \
    --warmup_steps 2000 \
    --lr 0.001 \
    --model.N_cycle 4 \
    --sample_diffusion.N_step 20 \
    --triangle_attention "cuequivariance" \
    --triangle_multiplicative "cuequivariance" \
    --data.train_sets weightedPDB_before2109_wopb_nometalc_0925 \
    --data.test_sets recentPDB_1536_sample384_0925,posebusters_0925
```

--------------------------------

### Build and run Protenix in Docker

Source: https://context7.com/lucidrains/protenix/llms.txt

Build the container image and run inference with GPU support.

```bash
# Build Docker image
docker build -t protenix:latest .

# Run inference in container
docker run --gpus all -v /path/to/data:/data -v /path/to/output:/output \
    protenix:latest protenix pred \
    -i /data/input.json \
    -o /output \
    -n protenix_base_default_v1.0.0
```

--------------------------------

### Run Fine-tuning Demo

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Executes the fine-tuning script to load pretrained weights and train on a specific subset.

```bash
bash finetune_demo.sh
```

--------------------------------

### Run Template Search

Source: https://github.com/lucidrains/protenix/blob/main/docs/msa_template_pipeline.md

Command to execute the template search pipeline, which builds HMM profiles and searches the database.

```bash
python3 runner/template_search.py
```

--------------------------------

### Full Input Preprocessing Pipeline

Source: https://context7.com/lucidrains/protenix/llms.txt

Runs a comprehensive preprocessing pipeline including protein MSA, template search, and RNA MSA search. Requires paths to various HMMER and RNA database binaries.

```bash
protenix prep \
    --input examples/input.json \
    --out_dir ./output \
    --msa_server_mode protenix \
    --hmmsearch_binary_path /usr/bin/hmmsearch \
    --hmmbuild_binary_path /usr/bin/hmmbuild \
    --seqres_database_path /path/to/pdb_seqres.fasta \
    --nhmmer_binary_path /usr/bin/nhmmer \
    --hmmalign_binary_path /usr/bin/hmmalign \
    --rfam_database_path /path/to/rfam.fasta \
    --rna_central_database_path /path/to/rnacentral.fasta \
    --ntrna_database_path /path/to/nt_rna.fasta
```

--------------------------------

### Generate MSA and template information via CLI

Source: https://github.com/lucidrains/protenix/blob/main/docs/infer_json_format.md

Automates the search for MSAs and templates to update input JSON files.

```bash
# Add template information to an existing JSON
protenix mt -i examples/example_without_msa.json -o ./output

# Add both template and RNA MSA information to an existing JSON
# Note: To force a fresh search, ensure any existing 'unpairedMsaPath' fields are removed from the JSON.
protenix prep -i examples/examples_with_rna_msa/example_9gmw_2.json -o ./output
```

--------------------------------

### Run Protenix inference with CLI

Source: https://context7.com/lucidrains/protenix/llms.txt

Execute inference using different model variants. Mini and Tiny models do not require MSA.

```bash
protenix pred -i input.json -n protenix_mini_esm_v0.5.0        # ESM embeddings (no MSA required)
protenix pred -i input.json -n protenix_mini_ism_v0.5.0        # ISM embeddings (no MSA required)
protenix pred -i input.json -n protenix_mini_default_v0.5.0    # Standard lightweight
protenix pred -i input.json -n protenix_tiny_default_v0.5.0
```

--------------------------------

### Distributed Training with torchrun

Source: https://context7.com/lucidrains/protenix/llms.txt

Scale training across multiple GPUs using PyTorch's distributed training. This command initiates multi-GPU training with specified parameters.

```bash
# Multi-GPU training with torchrun
torchrun --nproc_per_node=8 runner/train.py \
    --model_name "protenix_base_default_v1.0.0" \
    --dtype bf16 \
    --diffusion_batch_size 48 \
    --train_crop_size 384 \
    --triangle_attention "cuequivariance" \
    --triangle_multiplicative "cuequivariance"
```

--------------------------------

### Configure Fine-tuning Subset

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Specifies the path to a text file containing PDB IDs for fine-tuning.

```bash
--data.weightedPDB_before2109_wopb_nometalc_0925.base_info.pdb_list examples/finetune_subset.txt
```

--------------------------------

### Execute MSA Search with ColabFold

Source: https://github.com/lucidrains/protenix/blob/main/scripts/msa/step2-get_msa.ipynb

Sets the PYTHONPATH and creates log and result directories. It then executes the ColabFold search script using MMseqs2 with specified databases and parameters, including template usage and thread count.

```bash
export PYTHONPATH=$PYTHONPATH:ColabFold
mkdir -p logs
mkdir -p results
python3 ColabFold/colabfold/mmseqs/search.py --mmseqs=mmseqs {input_file} {db_dir} {output_dir} --db1 uniref30_2103_db --db2 pdb70_220313_db --db3 colabfold_envdb_202108_db --use-templates 1 --db-load-mode 2 --threads 32
# if you use scripts/msa/data/pdb_seqs/pdb_seq.fasta as {input_file}
# you will get the following results as in scripts/msa/data/mmcif_msa_initial
# 0.a3m  1.a3m  2.a3m  3.a3m  pdb70_220313_db.m8  uniref_tax.m8
```

--------------------------------

### Perform Input Preprocessing

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Generates MSA and template information required for optimal prediction accuracy.

```bash
# Full preprocessing (Protein MSA + Template + RNA MSA)
protenix prep --input examples/input.json --out_dir ./output

# Sequential Protein MSA and Template search
protenix mt --input examples/input.json --out_dir ./output

# Independent MSA search (supports JSON or Protein FASTA)
protenix msa --input examples/prot.fasta --out_dir ./output --msa_server_mode protenix
```

--------------------------------

### Clone Protenix Repository

Source: https://github.com/lucidrains/protenix/blob/main/docs/docker_installation.md

Clone the Protenix GitHub repository to your local machine and navigate into the project directory.

```bash
git clone https://github.com/bytedance/protenix.git 
cd ./protenix
```

--------------------------------

### Pull Protenix Docker Image

Source: https://github.com/lucidrains/protenix/blob/main/docs/docker_installation.md

Download the Protenix Docker image from the specified registry. This image includes all necessary dependencies.

```bash
docker pull ai4s-share-public-cn-beijing.cr.volces.com/release/protenix:1.0.0.4
```

--------------------------------

### Preprocess Training Data

Source: https://github.com/lucidrains/protenix/blob/main/docs/prepare_training_data.md

Executes the main preprocessing script to convert CIF files into training-ready formats.

```bash
python3 scripts/prepare_training_data.py -i [input_path] -o [output_csv] -b [output_dir] -c [cluster_txt] -n [num_cpu]
```

--------------------------------

### Generate Protenix-Compatible MSA with colabfold_msa.py

Source: https://github.com/lucidrains/protenix/blob/main/docs/colabfold_compatible_msa.md

Use this script to post-process `colabfold_search` results by adding pseudo taxonomy IDs to paired MSAs, making them compatible with Protenix's data pipeline. Ensure Colabfold databases and MMseqs2 are correctly configured.

```bash
python3 scripts/colabfold_msa.py examples/dimer.fasta <path/to/colabfold_db> dimer_colabfold_msa --db1 uniref30_2103_db --db3 colabfold_envdb_202108_db --mmseqs_path <path/to/mmseqs>
```

--------------------------------

### Run Distributed Training

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Utilizes PyTorch's torchrun to perform distributed training across multiple nodes.

```bash
torchrun --nproc_per_node=8 runner/train.py --model_name "protenix_base_default_v1.0.0" [OTHER_ARGS]
```

--------------------------------

### Direct Script Inference (Mini Model)

Source: https://context7.com/lucidrains/protenix/llms.txt

Run inference directly using the runner script with full parameter control for the mini model, optimized for fast inference. Ensure PYTHONPATH is set.

```bash
export PYTHONPATH="${PYTHONPATH}:$(pwd)"

# Mini model for fast inference
python3 runner/inference.py \
    --model_name protenix_mini_esm_v0.5.0 \
    --seeds 101 \
    --dump_dir ./output \
    --input_json_path ./examples/input.json \
    --model.N_cycle 4 \
    --sample_diffusion.N_sample 1 \
    --sample_diffusion.N_step 5 \
    --triangle_attention "cuequivariance" \
    --triangle_multiplicative "cuequivariance"
```

--------------------------------

### Model Fine-tuning

Source: https://context7.com/lucidrains/protenix/llms.txt

Fine-tune a pretrained model on a specific subset of PDB structures using the training script. Specify checkpoint paths and the list of PDB IDs for fine-tuning.

```bash
export PYTHONPATH="${PYTHONPATH}:$(pwd)"

# Fine-tune on specific PDB IDs
python3 ./runner/train.py \
    --run_name protenix_finetune \
    --model_name "protenix_base_default_v1.0.0" \
    --load_checkpoint_path /path/to/pretrained_model.pt \
    --load_ema_checkpoint_path /path/to/pretrained_ema.pt \
    --seed 42 \
    --base_dir ./output \
    --dtype bf16 \
    --diffusion_batch_size 32 \
    --train_crop_size 640 \
    --max_steps 10000 \
    --data.weightedPDB_before2109_wopb_nometalc_0925.base_info.pdb_list examples/finetune_subset.txt
```

--------------------------------

### Define ligand input

Source: https://github.com/lucidrains/protenix/blob/main/docs/infer_json_format.md

Supports ligands defined by CCD codes, SMILES strings, or paths to structure files (PDB, SDF, MOL, MOL2).

```json
{
    "ligand": {
        "ligand": "CCD_ATP",
        "count": 1
    }
},
{
    "ligand": {
        "ligand": "FILE_your_file_path/atp.sdf",
        "count": 1
    }
},
{
    "ligand": {
        "ligand": "Nc1ncnc2c1ncn2[C@@H]1O[C@H](CO[P@@](=O)(O)O[P@](=O)(O)OP(=O)(O)O)[C@@H](O)[C@H]1O",
        "count": 1
    }
}
```

--------------------------------

### Run Inference Demo Script

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Use this bash script for automated inference runs. Specify the model name, input JSON path, output directory, data type, and whether to use MSA features.

```bash
bash inference_demo.sh <model_name> <input_json> <output_dir> <dtype> <use_msa>
```

--------------------------------

### Direct Script Inference (Base Model)

Source: https://context7.com/lucidrains/protenix/llms.txt

Run inference directly using the runner script with full parameter control for the base model v1.0.0 with template support. Ensure PYTHONPATH is set.

```bash
export PYTHONPATH="${PYTHONPATH}:$(pwd)"

# Base model v1.0.0 with template support
python3 runner/inference.py \
    --model_name protenix_base_default_v1.0.0 \
    --seeds 101 \
    --dump_dir ./output \
    --input_json_path ./examples/input.json \
    --model.N_cycle 10 \
    --sample_diffusion.N_sample 5 \
    --sample_diffusion.N_step 200 \
    --triangle_attention "cuequivariance" \
    --triangle_multiplicative "cuequivariance" \
    --use_template true
```

--------------------------------

### Download Protein Clustering File

Source: https://github.com/lucidrains/protenix/blob/main/docs/prepare_training_data.md

Downloads the official RCSB PDB clustering results for sequence identity thresholds.

```bash
wget https://cdn.rcsb.org/resources/sequence/clusters/clusters-by-entity-40.txt
```

--------------------------------

### Configure Triangle Multiplicative

Source: https://github.com/lucidrains/protenix/blob/main/docs/kernels.md

Set the triangle multiplicative implementation in configs_base.py.

```python
triangle_multiplicative = "cuequivariance"  # or "torch"
```

--------------------------------

### Ligand Input Formats in JSON

Source: https://context7.com/lucidrains/protenix/llms.txt

Specifies how ligands can be provided in JSON format, supporting CCD codes, SMILES strings, and file paths.

```json
[
    {
        "name": "ligand_examples",
        "sequences": [
            {
                "proteinChain": {
                    "sequence": "MAEVIRSSAFWRSFPIFEEFDSETLCELS...",
                    "count": 1
                }
            },
            {
                "ligand": {
                    "ligand": "CCD_ATP",
                    "count": 1
                }
            },
            {
                "ligand": {
                    "ligand": "CCD_NAG_BMA_BGC",
                    "count": 1
                }
            },
            {
                "ligand": {
                    "ligand": "Nc1ncnc2c1ncn2[C@@H]1O[C@H](CO[P@@](=O)(O)O[P@](=O)(O)OP(=O)(O)O)[C@@H](O)[C@H]1O",
                    "count": 1
                }
            },
            {
                "ligand": {
                    "ligand": "FILE_/path/to/molecule.sdf",
                    "count": 1
                }
            }
        ]
    }
]
```

--------------------------------

### Convert PDB to JSON using Protenix CLI

Source: https://context7.com/lucidrains/protenix/llms.txt

Command to convert a PDB file to Protenix-compatible JSON format. Ensure the input path is correct.

```bash
protenix json --input ./examples/7pzb.pdb --out_dir ./output --altloc first
```

--------------------------------

### MSA Search from JSON Input

Source: https://context7.com/lucidrains/protenix/llms.txt

Performs an MSA search using a JSON file as input and updates the JSON with MSA paths. Requires specifying the MSA server mode.

```bash
protenix msa --input examples/input.json --out_dir ./output --msa_server_mode protenix
```

--------------------------------

### Run Model Inference

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Executes the prediction engine with various configuration options for models, seeds, and feature toggles.

```bash
# Standard inference (using default model and parameters)
protenix pred -i examples/input.json -o ./output -s 101 -n  protenix_base_default_v1.0.0 --use_template true

# Standard inference with seeds from JSON
protenix pred -i  examples/examples_with_template/example_mgyp004658859411.json --use_seeds_in_json true

# Standard inference with RNA MSA, only support protenix_base_default_v1.0.0
protenix pred -i  examples/examples_with_rna_msa/example_9gmw_2.json --use_rna_msa true -n protenix_base_default_v1.0.0

# Inference using Protenix-Mini (faster, lightweight)
protenix pred --input examples/input.json --model_name protenix_mini_default_v0.5.0

# Customized inference: Disable MSA and use shared variable caching
protenix pred --input examples/input.json --use_msa false --enable_cache true
```

--------------------------------

### Combined MSA and Template Search

Source: https://context7.com/lucidrains/protenix/llms.txt

Executes sequential MSA and template searches. Requires paths to HMMER binaries and the PDB sequence database.

```bash
protenix mt \
    --input examples/input.json \
    --out_dir ./output \
    --msa_server_mode protenix \
    --hmmsearch_binary_path /usr/bin/hmmsearch \
    --hmmbuild_binary_path /usr/bin/hmmbuild \
    --seqres_database_path /path/to/pdb_seqres.fasta
```

--------------------------------

### Execute MSA Search Notebook

Source: https://github.com/lucidrains/protenix/blob/main/docs/msa_template_pipeline.md

The provided Jupyter notebook template for performing the MSA search process.

```python
scripts/msa/step2-get_msa.ipynb
```

--------------------------------

### Cite Related Foundational Works

Source: https://github.com/lucidrains/protenix/blob/main/README.md

BibTeX entries for foundational projects that inspired Protenix.

```bibtex
@article{abramson2024accurate,
  title={Accurate structure prediction of biomolecular interactions with AlphaFold 3},
  author={Abramson, Josh and Adler, Jonas and Dunger, Jack and Evans, Richard and Green, Tim and Pritzel, Alexander and Ronneberger, Olaf and Willmore, Lindsay and Ballard, Andrew J and Bambrick, Joshua and others},
  journal={Nature},
  volume={630},
  number={8016},
  pages={493--500},
  year={2024},
  publisher={Nature Publishing Group UK London}
}
@article{ahdritz2024openfold,
  title={OpenFold: Retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization},
  author={Ahdritz, Gustaf and Bouatta, Nazim and Floristean, Christina and Kadyan, Sachin and Xia, Qinghui and Gerecke, William and O’Donnell, Timothy J and Berenberg, Daniel and Fisk, Ian and Zanichelli, Niccol{\`o} and others},
  journal={Nature Methods},
  volume={21},
  number={8},
  pages={1514--1524},
  year={2024},
  publisher={Nature Publishing Group US New York}
}
@article{mirdita2022colabfold,
  title={ColabFold: making protein folding accessible to all},
  author={Mirdita, Milot and Sch{"u}tze, Konstantin and Moriwaki, Yoshitaka and Heo, Lim and Ovchinnikov, Sergey and Steinegger, Martin},
  journal={Nature methods},
  volume={19},
  number={6},
  pages={679--682},
  year={2022},
  publisher={Nature Publishing Group US New York}
}
```

--------------------------------

### Configure DeepSpeed CUTLASS Path

Source: https://github.com/lucidrains/protenix/blob/main/docs/kernels.md

Set the CUTLASS path for the DS4Sci_EvoformerAttention kernel via Dockerfile or environment variables.

```bash
RUN git clone -b v3.5.1 https://github.com/NVIDIA/cutlass.git  /opt/cutlass
ENV CUTLASS_PATH=/opt/cutlass
```

```bash
git clone -b v3.5.1 https://github.com/NVIDIA/cutlass.git  /path/to/cutlass
export CUTLASS_PATH=/path/to/cutlass
```

--------------------------------

### MSA Search from FASTA Input

Source: https://context7.com/lucidrains/protenix/llms.txt

Performs an MSA search using a FASTA file as input. Requires specifying the MSA server mode.

```bash
protenix msa --input examples/prot.fasta --out_dir ./output --msa_server_mode colabfold
```

--------------------------------

### Run RNA MSA search

Source: https://github.com/lucidrains/protenix/blob/main/docs/msa_template_pipeline.md

Executes the RNA MSA search pipeline using the provided runner script.

```bash
python3 runner/rna_msa_search.py
```

--------------------------------

### Process MSA Data

Source: https://github.com/lucidrains/protenix/blob/main/docs/msa_template_pipeline.md

Commands to append taxonomy IDs and organize the raw A3M files into the final Protenix-compatible directory structure.

```bash
python3 scripts/msa/step3-uniref_add_taxid.py
```

```bash
python3 scripts/msa/step4-split_msa_to_uniref_and_others.py --input_msa_dir scripts/msa/data/mmcif_msa_with_taxid --output_msa_dir scripts/msa/data/mmcif_msa
```

--------------------------------

### Convert CIF to JSON with Assembly ID

Source: https://context7.com/lucidrains/protenix/llms.txt

Command to convert a CIF file to JSON, specifying an assembly ID. Requires downloading the CIF file first.

```bash
wget -P ./examples/ https://files.rcsb.org/download/7pzb.cif
protenix json --input ./examples/7pzb.cif --out_dir ./output --altloc first --assembly_id 1
```

--------------------------------

### Run Protenix structure prediction

Source: https://github.com/lucidrains/protenix/blob/main/README.md

Execute a structure prediction task using a JSON input file and a specified model version.

```bash
# Predict structure using a JSON input
protenix pred -i examples/input.json -o ./output -n protenix_base_default_v1.0.0
```

--------------------------------

### Configure Triangle Attention

Source: https://github.com/lucidrains/protenix/blob/main/docs/kernels.md

Set the triangle attention implementation in configs_base.py.

```python
triangle_attention = "cuequivariance"  # or "triattention"/"deepspeed"/"torch"
```

--------------------------------

### Convert Structural Files to JSON

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

Converts PDB or CIF files into the JSON format required by Protenix.

```bash
# Convert PDB/CIF to JSON
protenix json --input ./examples/7pzb.pdb --out_dir ./output --altloc first

# Advanced: Specify assembly ID for biological assemblies
wget -P ./examples/ https://files.rcsb.org/download/7pzb.cif 
protenix json --input ./examples/7pzb.cif --out_dir ./output --altloc first
```

--------------------------------

### Python Inference with InferenceRunner

Source: https://context7.com/lucidrains/protenix/llms.txt

Use InferenceRunner for programmatic inference with fine-grained control over model configuration. Ensure necessary imports are present.

```python
from runner.inference import InferenceRunner, infer_predict, download_inference_cache
from runner.batch_inference import get_default_runner
from configs.configs_base import configs as configs_base
from configs.configs_inference import inference_configs
from protenix.config.config import parse_configs

# Quick inference using the default runner
runner = get_default_runner(
    seeds=[101, 102, 103],
    n_cycle=10,
    n_step=200,
    n_sample=5,
    dtype="bf16",
    model_name="protenix_base_default_v1.0.0",
    use_msa=True,
    use_template=True,
    use_rna_msa=False,
    trimul_kernel="cuequivariance",
    triatt_kernel="cuequivariance",
    enable_cache=True,
    enable_fusion=True,
    enable_tf32=True
)

# Run inference on input JSON
configs = runner.configs
configs["input_json_path"] = "./examples/input.json"
configs["dump_dir"] = "./output"
infer_predict(runner, configs)
```

--------------------------------

### Define Protenix Inference Job Structure

Source: https://github.com/lucidrains/protenix/blob/main/docs/infer_json_format.md

The top-level structure of the input JSON file must be a list of dictionaries, even for a single job.

```json
[
  {
    "name": "Test Fold Job Number One",
    "sequences": [...],
    "covalent_bonds": [...]
  }
]
```

--------------------------------

### Convert Directory of PDB/CIF Files to JSON

Source: https://context7.com/lucidrains/protenix/llms.txt

Command to convert all PDB and CIF files within a specified directory to JSON format.

```bash
protenix json --input ./structures/ --out_dir ./output
```

--------------------------------

### Update Inference Configurations Based on Token Count

Source: https://github.com/lucidrains/protenix/blob/main/docs/training_inference_instructions.md

This function dynamically adjusts Automatic Mixed Precision (AMP) settings for `SampleDiffusion` and `ConfidenceHead` modules to manage memory usage and prevent OOM errors for large sequences. It prioritizes accuracy by disabling AMP for smaller token counts and enables it for larger ones.

```python
def update_inference_configs(configs: Any, n_token: int) -> Any:
    """
    Adjust inference configurations based on the number of tokens to manage memory usage and prevent OOM.
    
    Args:
        configs (Any): Original configurations.
        n_token (int): Number of tokens in the sample.

    Returns:
        Any: Updated configurations.
    """
    if n_token > 3840:
        # Enable AMP for both modules to save memory for extremely large sequences
        configs.skip_amp.confidence_head = False
        configs.skip_amp.sample_diffusion = False
    elif n_token > 2560:
        # Enable AMP only for ConfidenceHead
        configs.skip_amp.confidence_head = False
        configs.skip_amp.sample_diffusion = True
    else:
        # Default: Disable AMP for both (run in FP32) to prioritize accuracy
        configs.skip_amp.confidence_head = True
        configs.skip_amp.sample_diffusion = True

    return configs
```

--------------------------------

### Define Input JSON

Source: https://context7.com/lucidrains/protenix/llms.txt

Structure of the JSON input file for defining sequences, modifications, and covalent bonds.

```json
[
    {
        "name": "protein_dna_complex",
        "sequences": [
            {
                "proteinChain": {
                    "sequence": "MGSSHHHHHHSSGLVPRGSHMSGKIQHKAVVPAPSRIPLTLSEIEDLRRKGFNQTEIAELYGVTRQAVSWHKKTYGGRLTTRQIVQQNWPWDTRKPHDKSKAFQRLRDHGEYMRVGSFRTMSEDKKKRLLSWWKMLRDNDLVLEFDPSIEPYEGMAGGGFRYVPRDISDDDLLIRVNEHTQLTAEGELLWSWPDDIEELLSEP",
                    "count": 1,
                    "pairedMsaPath": "/path/to/pairing.a3m",
                    "unpairedMsaPath": "/path/to/non_pairing.a3m",
                    "templatesPath": "/path/to/hmmsearch.a3m",
                    "modifications": [
                        {"ptmType": "CCD_HY3", "ptmPosition": 1}
                    ]
                }
            },
            {
                "dnaSequence": {
                    "sequence": "TTTCGGTGGCTGTCAAGCGGG",
                    "count": 1
                }
            },
            {
                "rnaSequence": {
                    "sequence": "GUACGUAC",
                    "count": 1,
                    "unpairedMsaPath": "/path/to/rna_msa.a3m"
                }
            },
            {
                "ligand": {
                    "ligand": "CCD_ATP",
                    "count": 1
                }
            },
            {
                "ion": {
                    "ion": "MG",
                    "count": 2
                }
            }
        ],
        "covalent_bonds": [
            {
                "entity1": "2",
                "copy1": 1,
                "position1": "2",
                "atom1": "N6",
                "entity2": "3",
                "copy2": 1,
                "position2": "1",
                "atom2": "C1"
            }
        ]
    }
]
```

--------------------------------

### Disable Fast Layernorm

Source: https://github.com/lucidrains/protenix/blob/main/docs/kernels.md

Set the environment variable to use native PyTorch layernorm instead of the default fast_layernorm.

```bash
export LAYERNORM_TYPE=torch
```

--------------------------------

### Define Token Contact Constraint

Source: https://github.com/lucidrains/protenix/blob/main/docs/infer_json_format.md

Use this to define contact between tokens, specifying entities, copies, and positions. Minimum distance defaults to 0.

```json
"contact": [
    {
        "entity1": 1,
        "copy1": 1,
        "position1": 169,
        "entity2": 2,
        "copy2": 1,
        "position2": 1,
        "atom2": "C5",
        "max_distance": 6,
        "min_distance": 0
    }, // token-contact
    {
        "entity1": 1,
        "copy1": 1,
        "position1": 169,
        "atom1": "CA",
        "entity2": 2,
        "copy2": 1,
        "position2": 1,
        "max_distance": 6,
        "min_distance": 0
    }, // token-contact
    {
        "entity1": 1,
        "copy1": 1,
        "position1": 169,
        "entity2": 2,
        "copy2": 1,
        "position2": 1,
        "max_distance": 6,
        "min_distance": 0
    }, // token-contact
    {
        "entity1": 1,
        "copy1": 1,
        "position1": 169,
        "atom1": "CA",
        "entity2": 2,
        "copy2": 1,
        "position2": 1,
        "atom2": "C5",
        "max_distance": 6,
        "min_distance": 3
    }, // atom-contact
    ...
]
```

--------------------------------

### Define ion input

Source: https://github.com/lucidrains/protenix/blob/main/docs/infer_json_format.md

Represents ions using their CCD codes without the 'CCD_' prefix.

```json
{
    "ion": {
        "ion": "MG",
        "count": 2
    }
},
{
    "ion": {
        "ion": "NA",
        "count": 3
    }
}
```

--------------------------------

### Define Pocket Constraint

Source: https://github.com/lucidrains/protenix/blob/main/docs/infer_json_format.md

Use this to define a pocket constraint, specifying the binder chain, contact residues, and maximum distance.

```json
"pocket": {
  "binder_chain": 
    {
      "entity": 2,
      "copy": 1
    }, 
  "contact_residues": [
    {
      "entity": 1,
      "copy": 1,
      "position": 126
    },
    ...
  ], 
  "max_distance": 6
}
```

--------------------------------

### Update CCD Cache

Source: https://github.com/lucidrains/protenix/blob/main/docs/prepare_training_data.md

Generates or updates the Chemical Component Dictionary cache for recent PDB entries.

```bash
python3 scripts/gen_ccd_cache.py -c [ccd_cache_dir] -n [num_cpu]
```