AlphaFold 3

Replaces the previous structure module with a diffusion model that operates directly on raw atomic coordinates. The model learns to reverse a noising process, starting from a random cloud of atoms and iteratively denoising them into a valid 3D structure. This generative approach handles the full chemical complexity of proteins, DNA, RNA, ligands, ions, and covalent modifications within a single unified framework, eliminating the need for separate modules for different molecule types.

AlphaFold 3 outputs per-atom and pairwise confidence estimates critical for experimental validation:

• pLDDT (predicted Local Distance Difference Test): A per-residue score (0-100) estimating local accuracy. Residues with pLDDT > 90 are high-confidence; those < 50 are likely disordered.
• PAE (Predicted Aligned Error): A 2D matrix quantifying the expected positional error between any two residues. Low PAE values between domains indicate rigid relative orientation; high PAE suggests flexible linkers.
• ipTM (interface predicted TM-score): Measures the confidence in the relative positioning of chains within a complex, crucial for assessing protein-protein and protein-nucleic acid interfaces.

Unlike AlphaFold 2, which required separate multimer and monomer models, AlphaFold 3 natively handles arbitrary biomolecular assemblies. A single forward pass predicts the joint structure of complexes containing:

• Multiple protein chains
• Double-stranded and single-stranded DNA
• Structured and unstructured RNA
• Small molecule ligands and cofactors
• Post-translational modifications and non-standard residues This capability is transformative for drug-target interaction modeling and understanding transcriptional machinery.

The Pairformer module replaces AlphaFold 2's Evoformer as the primary information processing block. It operates on two representations:

• Single representation: Encodes per-token (residue or atom) features
• Pair representation: Encodes pairwise relationships between tokens Crucially, the Pairformer processes these representations without explicit 3D coordinates, using triangle multiplicative updates and axial attention to reason about geometric constraints implicitly. This abstraction allows the model to learn physical plausibility before committing to atomic positions in the diffusion stage.

AlphaFold 3 is trained with a denoising score matching objective on the Protein Data Bank (PDB) and other structural databases. The training process:

1. Adds Gaussian noise to ground-truth atomic coordinates
2. Trains the network to predict the original coordinates from the noisy version
3. Conditions predictions on the sequence, multiple sequence alignment (MSA), and structural templates This generative framework naturally handles structural disorder and flexibility, producing diverse, physically plausible samples for regions with low experimental density rather than collapsing to a single overconfident prediction.

AlphaFold 3 retains a modified MSA processing pipeline to extract evolutionary co-evolutionary signals:

• Genetic search: Queries sequence databases (UniRef, BFD, MGnify) to build deep MSAs
• MSA Module: A lightweight transformer stack processes the MSA to produce a condensed representation
• Template search: Identifies homologous structures in the PDB for structural priors For protein-nucleic acid complexes, the model can leverage MSAs from both the protein and nucleic acid components, though RNA MSAs are typically shallower, making template information more critical for accurate RNA structure prediction.