Four Fine-Tuning Strategies
Head Only (~50K params), LoRA (~600K), Partial (~5M) and Full (~93M) — each suited to a different dataset size.
All projectsMolFun
Open-source framework
MolFun
Protein Fine-Tuning
An open-source framework for fine-tuning and adapting pre-trained protein structure prediction models like OpenFold. Four fine-tuning strategies, a modular registry and experiment tracking built in.
- PyTorch 2.x
- CUDA 12+
- Triton
- Python 3.10+
- Open source
MolFun
Key capabilities
Modular Architecture
A registry of swappable components: attention mechanisms, block types, structure modules and embedders.
Experiment Tracking
Native WandB, Comet, MLflow, Langfuse and HuggingFace integrations, with composite tracking to several at once.
Complete ML Pipeline
Data loading, MSA handling, featurization, trajectory analysis and export to ONNX, TorchScript and the Hub.
PyTorch Native
Built on PyTorch 2.x with drop-in replacements for standard ops, GPU acceleration and zero-copy tensors.
Adapt a foundation model to your own data
Structure prediction models are trained on everything and specialised in nothing. MolFun makes it practical to specialise one on the system you actually care about, without rewriting the training stack.
Read the docs- Pick a strategy by dataset size, from under 100 samples to over 10K
- Swap attention: standard, Flash, linear or gated
- Swap blocks: Evoformer or Pairformer
- Swap structure modules: IPA or diffusion
- Export to ONNX, TorchScript or HuggingFace Hub
Technology stack
- PyTorch
- CUDA
- Triton
- Python
- Hugging Face
- Weights & Biases
- MLflow
- ONNX
Using MolFun, or thinking about it?
It is open source and actively developed. Issues, ideas and collaborations are all welcome.