Try the new DeepSeek V4 Flash today. Frontier intelligence at a fraction of the cost. Here

Biology

BioNeMo ProteinMPNN

NVIDIA NIM for ProteinMPNN inverse folding and protein sequence design.

Model details

Developed by
NVIDIA
Model family
BioNeMo
Use case
biology
Variant
NIM
Hardware
A10G
License
NVIDIA AI Foundation Models Community License Agreement

ProteinMPNN is a deep learning model for protein sequence design that predicts amino acid sequences likely to fold into a given 3D protein backbone structure. It takes a protein structure (PDB format) as input and outputs candidate amino acid sequences (FASTA format), offering both soluble and non-soluble model variants suited to different design needs, such as high-solubility requirements for biotechnology and pharmaceutical applications. It's often used as a downstream step after a structure-generation model like RFdiffusion — taking the generated backbone and determining a possible amino acid sequence that folds into it.

Call BioNeMo ProteinMPNN on Baseten using the NIM sequence design route.

Input

1import os
2import requests
3
4MODEL_ID = os.environ["BASETEN_MODEL_ID"]
5BASETEN_API_KEY = os.environ["BASETEN_API_KEY"]
6BASETEN_URL = (
7    f"https://model-{MODEL_ID}.api.baseten.co/environments/production/sync"
8    "/biology/ipd/proteinmpnn/predict"
9)
10
11pdb_text = requests.get("https://files.rcsb.org/download/1R42.pdb", timeout=30).text
12atom_lines = [line for line in pdb_text.splitlines() if line.startswith("ATOM")][:200]
13input_pdb = "\n".join(atom_lines)
14
15payload = {
16    "input_pdb": input_pdb,
17    "ca_only": False,
18    "use_soluble_model": False,
19    "num_seq_per_target": 3,
20    "sampling_temp": [0.1, 0.3],
21}
22
23response = requests.post(
24    BASETEN_URL,
25    headers={
26        "Authorization": f"Bearer {BASETEN_API_KEY}",
27        "Content-Type": "application/json",
28    },
29    json=payload,
30    timeout=300,
31)
32response.raise_for_status()
33result = response.json()
34print(result)
35

JSON output

1{
2    "mfasta": ">T=0.1, score=1.2345, seq=0\nMKTV...\n",
3    "scores": [
4        [
5            1.2345
6        ]
7    ],
8    "probs": [
9        [
10            "<probabilities>"
11        ]
12    ]
13}