protenix

name	protenix
description	Structure prediction using Protenix, ByteDance's open-source PyTorch reproduction of AlphaFold 3. Use this skill when: (1) Predicting protein/DNA/RNA/ligand/ion complex structures, (2) Need AF3-level accuracy with open-source code, (3) MSA-free fast prediction (--no-use-msa), (4) Multi-seed ensemble predictions, (5) Alternative to Chai or Boltz for validation. For QC thresholds, use protein-qc. For Chai prediction, use chai. For Boltz prediction, use boltz.
license	MIT
category	design-tools
tags	["structure-prediction","validation","af3","open-source"]
source	https://github.com/hgbrian/biomodals

Protenix Structure Prediction

Protenix is an open-source PyTorch reproduction of AlphaFold 3 by ByteDance. Supports: protein, DNA, RNA, ligand (SMILES), ion.

Prerequisites

Requirement	Minimum	Recommended
Python	3.12+	3.12
GPU VRAM	24GB	48GB (L40S)

How to run

FASTA input (simplest)

# Single protein
echo ">protein|A
MAWTPLLLLLLSHCTGSLSQPVLTQPTSLSASPGASARFTCTLRSGINVGTYRIYWYYQQKPGSLP" > test.faa

modal run modal_protenix.py --input-faa test.faa

Protein complex

cat > complex.faa << 'EOF'
>protein|A
MKTAYIAKQRQISFVKSHFSRQLERRLEQLKQLEQQ...
>protein|B
MVLSPADKTNVKAAWGKVGAHAGEYGAEALERMFL...
EOF

modal run modal_protenix.py --input-faa complex.faa --seeds "42,43,44"

With ligand (SMILES)

cat > protein_ligand.faa << 'EOF'
>protein|A
MKTAYIAKQRQISFVKSHFSRQLE...
>ligand|caffeine
CN1C=NC2=C1C(=O)N(C(=O)N2C)C
EOF

modal run modal_protenix.py --input-faa protein_ligand.faa

With MSA (higher accuracy, slower)

modal run modal_protenix.py --input-faa complex.faa --use-msa

JSON input (native Protenix format)

[
    {
        "name": "my_complex",
        "sequences": [
            {"proteinChain": {"sequence": "MKTAYIAKQRQISFVK...", "count": 1}},
            {"proteinChain": {"sequence": "MVLSPADKTNVKAA...", "count": 1}}
        ]
    }
]

modal run modal_protenix.py --input-json complex.json

Key parameters

Parameter	Default	Description
--input-faa	-	FASTA input file
--input-json	-	JSON input (native format)
--seeds	"42"	Random seeds, comma-separated
--use-msa	off	Enable MSA (higher accuracy)
--no-use-msa	on	Skip MSA (faster)

Supported entity types

FASTA header	Type	Example
>protein|A	Protein chain A	standard proteins
>dna|A	DNA chain	ATCGATCG
>rna|A	RNA chain	AUCGAUCG
>ligand|name	Small molecule	SMILES string as sequence
>ion|name	Metal ion	ZN, MG, CA

Output format

output/
├── predictions/
│   ├── my_complex_seed-42_sample-0.cif    # Best model
│   ├── my_complex_seed-42_sample-1.cif
│   └── confidence_my_complex_seed-42.json # pLDDT, pTM, ipTM

Protenix vs Chai vs Boltz

Feature	Protenix	Chai-1	Boltz-1
Based on	AF3	Novel	Novel
MSA	Optional	No	No
DNA/RNA	Yes	Yes	Yes
Speed (no MSA)	Fast	Fast	Fast
GPU needed	L40S	A100	L40S

Typical performance

Campaign	Time (L40S)	Cost (Modal)
100 complexes	30-45 min	~$8
500 complexes	2-3h	~$35

Troubleshooting

Error	Cause	Fix
CUDA out of memory	Complex too large	Use A100-80GB
KeyError: 'iptm'	Single chain	Ensure 2+ chains
Invalid entity type	Bad FASTA header	Check protein|A, ligand|name

Next: protein-qc for filtering and ranking.