Alphafold3

AlphaFold3 is a diffusion-based all-atom structure prediction method that is widely seen as state-of-the-art.

Architecture of AlphaFold3 from Abramson et al. (1)

Architectural and ML contributions

PairFormer
PDE: predicted distance error (replacing frame aligned point error)
Non-equivariant per-atom prediction, which leads to occasional errors when predicting chirality
Cross-distillation from AlphaFold2-Multimer v2.3 to avoid hallucination of low-pLDDT regions

Abramson J, Adler J, Dunger J, Evans R, Green T, Pritzel A, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630(8016):493–500. Available from: https://doi.org/10.1038/s41586-024-07487-w

Quartz 4

Explorer

Alphafold3

Architectural and ML contributions

AlphaFold3 ipTM can distinguish between antibody binders and nonbinders

AF3 binding affinity predictions are orthogonal to those made by force fields and other neural networks

Antibody-antigen modeling by diffusion-based structure prediction is data-limited

Correct antibody-antigen prediction in AF3 and related models is partially determined by training set similarity

AF3-generation methods can be substantially shrunk without sacrificing prediction quality, except on PPIs

AF3-generation methods incorporate MSA information exclusively into the pair representation

AlphaFold3 learns intramolecular interactions faster than intermolecular interactions

AlphaFold3 performs comparably to AlphaFold2 when predicting multiple conformations of fold-switching proteins

AlphaFold3 universally predicts the active state of ligand-bound GPCRs, even when the ligand is an antagonist

AlphaFold3 uses cross-distillation from AlphaFold2 to avoid hallucinating secondary structure in low-pLDDT regions

Diffusion outperforms hallucination when using AF3-generation protein structure prediction methods

Low-pLDDT hallucinated models from AF3-generation protein structure predictors are designable

Protein-ligand co-folding methods do not generalize beyond their training set

The Boltz-2 affinity module cannot be effectively repurposed for PPI affinity prediction

The confidence metrics of AlphaFold2 are better calibrated than those of AlphaFold3

Flow matching and diffusion perform comparably for biomolecular structure prediction

Conformational sampling by AlphaFold3-generation methods can be achieved by scaling the pair representation