Theory

This section provides the scientific and mathematical background behind the methods used in Molfun. Whether you are a biologist wondering why attention mechanisms matter for proteins, or an ML engineer curious about the geometry inside the structure module, these pages will give you the full picture.

---

Protein Structure Prediction

The AlphaFold2/OpenFold architecture from the ground up: MSA processing, the Evoformer trunk, Invariant Point Attention, the diffusion approach in AlphaFold3, confidence heads, and how to add custom prediction heads.

Fine-Tuning Methods

Full fine-tuning, partial freezing, LoRA, QLoRA, IA3, BitFit, and beyond. Mathematical foundations, best practices for rank and alpha, and why domain-specific adaptation unlocks new capabilities.