bionemo

BioNeMo

Overview

NVIDIA's software ecosystem for building, training, fine-tuning, and deploying AI models for life sciences. BioNeMo Framework provides optimized biomolecular foundation models (ESM-2, AlphaFold2, ProteinMPNN, RFdiffusion) with distributed training support, while BioNeMo NIMs offer production-ready inference microservices with REST API endpoints for scalable deployment.

Quick Pattern

Incorrect — manual protein embedding without framework:

# Raw transformer, no optimization, no pretrained weights
model = TransformerEncoder(...)
embeddings = model(tokenize(sequence))

Correct — BioNeMo ESM-2 fine-tuning with pretrained checkpoint:

from bionemo.esm2.model.finetune.finetune_regressor import ESM2FineTuneSeqConfig
from bionemo.core.data.load import load

pretrain_ckpt = load("esm2/650m:2.0")
config = ESM2FineTuneSeqConfig(initial_ckpt_path=str(pretrain_ckpt))

checkpoint, metrics, trainer = train_model(
    experiment_name="finetune_regressor",
    experiment_dir=Path(results_dir),
    config=config,
    data_module=data_module,
    n_steps_train=50,
)

Quick Command

# Pull BioNeMo container
docker pull nvcr.io/nvidia/bionemo/bionemo-framework:latest

# Run container with GPU
docker run --gpus all -it --rm \
  -v ${PWD}:/workspace \
  nvcr.io/nvidia/bionemo/bionemo-framework:latest bash

# Download ESM-2 checkpoint
python -c "from bionemo.core.data.load import load; load('esm2/650m:2.0')"

# Run ESM-2 inference
infer_esm2 --checkpoint-path /path/to/ckpt \
  --data-path data.csv \
  --results-path results/

# NIM health check
curl http://localhost:8000/v1/health/ready

# RFdiffusion via NIM
curl -X POST http://localhost:8000/biology/ipd/rfdiffusion/generate \
  -H "Content-Type: application/json" \
  -d '{"contigs": "A10-100/0 50-150"}'

Quick Reference

Supported Models

Model	Type	Task
ESM-2	Protein language model	Sequence embedding, property prediction
AlphaFold2	Structure prediction	3D protein folding
AlphaFold2-Multimer	Complex prediction	Multi-chain structure
ProteinMPNN	Inverse folding	Sequence design from structure
RFdiffusion	Generative diffusion	De novo protein structure generation
Evo2	Genomic foundation model	DNA/RNA generation, variant prediction, 1B/7B/40B
Geneformer	Single-cell transformer	Cell embeddings, type classification, GRN, 10M/106M
AMPLIFY	Protein language model	ESM-2 variant, 120M/350M, modified layers
DiffDock	Docking model	Protein-ligand binding
MolMIM	Molecular model	Small molecule generation