Skip to content

Quick Start

Three examples to see what Molfun can do --- all in under five minutes.

1. Predict a Protein Structure

Given an amino acid sequence, predict its 3D coordinates and per-residue confidence scores.

from molfun import predict_structure

result = predict_structure(
    sequence="MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVL",
    backend="openfold",
    device="cpu",
)

print(f"Coordinates shape: {result['coordinates'].shape}")  # (N_residues, 37, 3)
print(f"Mean pLDDT: {result['plddt'].mean():.1f}")

# Save directly to a PDB file
with open("prediction.pdb", "w") as f:
    f.write(result["pdb_string"])

molfun structure \
    --sequence "MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVL" \
    --output prediction.pdb

One function, one line

predict_structure is a convenience wrapper around MolfunStructureModel. For more control --- custom heads, batched inference, or access to intermediate representations --- see First Prediction.

2. Fine-Tune with LoRA

Fine-tune a pretrained model on your own data using LoRA --- only a handful of lines.

from molfun import MolfunStructureModel
from molfun.training import LoRAFinetune
from torch.utils.data import DataLoader

# 1. Load a pretrained model
model = MolfunStructureModel.from_pretrained("openfold", device="cpu")

# 2. Prepare your data (any PyTorch Dataset / DataLoader)
train_loader = DataLoader(your_dataset, batch_size=2, shuffle=True)

# 3. Fine-tune with LoRA
strategy = LoRAFinetune(rank=8, alpha=16.0, lr_lora=1e-4, lr_head=1e-3)

metrics = model.fit(
    train_loader=train_loader,
    strategy=strategy,
    epochs=10,
    tracker="wandb",           # optional: track with Weights & Biases
    checkpoint_dir="./ckpts",  # optional: save checkpoints
)

# 4. Save the fine-tuned model
model.save("my_finetuned_model")

molfun fit \
    --data ./my_dataset \
    --strategy lora \
    --rank 8 \
    --epochs 10 \
    --tracker wandb \
    --output ./my_finetuned_model

Only ~0.5% of parameters are trained

With rank 8, LoRA adds low-rank adapters to the attention layers while keeping the full trunk frozen. This makes fine-tuning feasible on a single GPU with limited data. See First Fine-Tuning for a detailed walkthrough.

3. Predict Binding Affinity

Estimate how strongly a protein binds to a small-molecule ligand.

from molfun import predict_affinity

result = predict_affinity(
    sequence="MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVL",
    ligand_smiles="CC(=O)Oc1ccccc1C(=O)O",  # aspirin
    device="cpu",
)

print(f"Binding affinity: {result['binding_affinity_kcal']:.2f} kcal/mol")
print(f"Confidence: {result['confidence']:.2f}")

molfun affinity \
    --sequence "MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVL" \
    --ligand "CC(=O)Oc1ccccc1C(=O)O" \
    --device cpu

Bonus: Molecular Properties

Compute physicochemical properties from sequence alone --- no model weights needed.

from molfun import predict_properties

props = predict_properties(
    sequence="MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVL",
    properties=["molecular_weight", "isoelectric_point", "hydrophobicity", "charge"],
)

for name, value in props.items():
    print(f"{name}: {value:.2f}")

What is Next?

Goal Guide
Understand each prediction step in detail First Prediction
Run a full fine-tuning workflow with evaluation First Fine-Tuning
Explore real-world use cases Tutorials
Dive into the API reference MolfunStructureModel