AlphaFold 3

Google DeepMind / Isomorphic Labs unified diffusion model predicting the joint 3D structure of proteins, nucleic acids, small molecules and ions.

1. Core Product / Service

AlphaFold 3 (Abramson et al., Nature 2024) is a single unified deep-learning model that jointly predicts the structure of complexes spanning proteins, nucleic acids, small-molecule ligands, ions and modified residues — a major expansion beyond AlphaFold 2's protein-only scope [1]. Architecturally it replaces AF2's equivariant structure module with a diffusion module that operates directly on raw atom coordinates, without rotational frames or equivariant processing — a notable bet that data scale can substitute for hardcoded symmetry [1]. Inputs are polymer sequences plus ligand SMILES; the model is paired with a confidence head and runs against retrieval databases (PDB, MGnify, UniRef, etc.).

2. Target Users & Pain Points

Structural biologists, drug-discovery and computational-chemistry teams who previously stitched together separate tools for protein folding, docking, and nucleic-acid modelling. AF3 collapses these into one model and directly tackles protein–ligand and protein–nucleic-acid interaction prediction — the step that matters for drug design. Access is via the AlphaFold Server (non-commercial) and released model weights/code for academic use.

3. Competitive Landscape

AF3 reportedly greatly outperforms classical docking tools (e.g. Vina) and prior specialist methods on its reported benchmarks [1]. Open reproductions (Boltz, Chai) have since narrowed the access gap.

4. Unique Observations

The diffusion-over-coordinates design is the single most-cited structural choice — it is the clearest "scaling beats equivariance" data point in the broader ai-for-science equivariance debate. AF3 is also a generative model used for prediction, blurring the predict-vs-design line that aifs-biology tracks (RFdiffusion etc. sit on the design side of the same diffusion toolkit). A known limitation: generative models can hallucinate plausible-but-wrong structure in disordered regions, which downstream users must filter with the confidence outputs.

5. Financials / Funding

Not a standalone commercial entity. Developed by Google DeepMind with Isomorphic Labs (DeepMind's drug-discovery spinout); commercialization runs through Isomorphic's pharma partnerships rather than an AF3 product P&L.

6. People & Relationships

• Developed by Google DeepMind + Isomorphic Labs; lead authors incl. Josh Abramson, John Jumper et al.
• John Jumper & Demis Hassabis shared the 2024 Nobel Prize in Chemistry for the AlphaFold work.
• Related: gnome (DeepMind's materials counterpart), ai-for-science, aifs-biology.

Sources

• [1] Abramson et al., "Accurate structure prediction of biomolecular interactions with AlphaFold 3", Nature s41586-024-07487-w (2026-06-14)
• google-deepmind/alphafold3 GitHub — inference code + weights (2026-06-14)