# Inside AlphaFold 2 and AlphaFold 3

> The two architectures side by side: AlphaFold 2 refines an MSA and a pair representation through 48 Evoformer blocks and builds the chain with a structure module; AlphaFold 3 keeps a lighter 4-block MSA module, 48 Pairformer blocks and generates atom coordinates with a diffusion module, for complexes of proteins, nucleic acids, small molecules and ions.

- Page: https://www.labcanvas.com/templates/drug-discovery/inside-alphafold-2-alphafold-3
- Open in the LabCanvas editor (LabCanvas account): https://www.labcanvas.com/editor.html?template=alphafold-2-and-3-architecture
- Updated: 2026-10-14
- Category: Drug discovery & proteins
- Method: Structure prediction
- Starting materials: Sequences (and for AlphaFold 3, ligands, ions and modified residues)
- Product: Atomic coordinates with confidence (pLDDT, PAE)
- Key result: Block diagram drawn after the papers; schematic, not to scale
- Figure: 1920 × 1080 px, 21 editable layers (2 generated objects with settings)

## Steps

1. **Inputs**: Genetic and structure database searches give an MSA and templates
2. **Evoformer**: 48 blocks exchange information between the MSA and pair representations (Jumper 2021; block count from Abramson 2024)
3. **Structure module**: Builds the chain from residue frames and side-chain angles (AlphaFold 2)
4. **AlphaFold 3 trunk**: A 4-block MSA module, then 48 Pairformer blocks on single and pair representations (Abramson 2024)
5. **Diffusion**: Atom coordinates generated from noise; a confidence head predicts pLDDT, PAE and PDE (Abramson 2024)

## Strengths

- One framework from proteins to complexes
- Confidence heads alongside the structure

## Limitations

- Diagrams simplify: many layers and inputs are left out

## Uses

- Teaching slides
- Methods figures
- Reviews

## References

1. Jumper et al., “Highly accurate protein structure prediction with AlphaFold”, Nature 2021. https://doi.org/10.1038/s41586-021-03819-2 (primary)
2. Abramson et al., “Accurate structure prediction of biomolecular interactions with AlphaFold 3”, Nature 2024. https://doi.org/10.1038/s41586-024-07487-w (primary)
3. Baek et al., “Accurate prediction of protein structures and interactions using a three-track neural network”, Science 2021. https://doi.org/10.1126/science.abj8754 (further reading)
4. Krishna et al., “Generalized biomolecular modeling and design with RoseTTAFold All-Atom”, Science 2024. https://doi.org/10.1126/science.adl2528 (further reading)

Credit: After Jumper et al., Nature 2021 (doi:10.1038/s41586-021-03819-2) and Abramson et al., Nature 2024 (doi:10.1038/s41586-024-07487-w). Schematic block diagrams; block counts as stated in the AlphaFold 3 paper.

## Questions

**What does the “Inside AlphaFold 2 and AlphaFold 3” template show?** The two architectures side by side: AlphaFold 2 refines an MSA and a pair representation through 48 Evoformer blocks and builds the chain with a structure module; AlphaFold 3 keeps a lighter 4-block MSA module, 48 Pairformer blocks and generates atom coordinates with a diffusion module, for complexes of proteins, nucleic acids, small molecules and ions.

**What are the steps shown in “Inside AlphaFold 2 and AlphaFold 3”?** 1. Inputs: Genetic and structure database searches give an MSA and templates. 2. Evoformer: 48 blocks exchange information between the MSA and pair representations (Jumper 2021; block count from Abramson 2024). 3. Structure module: Builds the chain from residue frames and side-chain angles (AlphaFold 2). 4. AlphaFold 3 trunk: A 4-block MSA module, then 48 Pairformer blocks on single and pair representations (Abramson 2024). 5. Diffusion: Atom coordinates generated from noise; a confidence head predicts pLDDT, PAE and PDE (Abramson 2024).

**What are the starting materials and the product?** Starting materials: Sequences (and for AlphaFold 3, ligands, ions and modified residues). Product: Atomic coordinates with confidence (pLDDT, PAE).

**Which papers is “Inside AlphaFold 2 and AlphaFold 3” based on?** Jumper et al., “Highly accurate protein structure prediction with AlphaFold”, Nature 2021 (doi:10.1038/s41586-021-03819-2); Abramson et al., “Accurate structure prediction of biomolecular interactions with AlphaFold 3”, Nature 2024 (doi:10.1038/s41586-024-07487-w); Baek et al., “Accurate prediction of protein structures and interactions using a three-track neural network”, Science 2021 (doi:10.1126/science.abj8754); Krishna et al., “Generalized biomolecular modeling and design with RoseTTAFold All-Atom”, Science 2024 (doi:10.1126/science.adl2528).

**Can I edit the “Inside AlphaFold 2 and AlphaFold 3” template?** Yes. Open it in the LabCanvas editor: all 21 layers are editable, including 2 generated objects (structures, plots, equipment) with their own settings. Change the text, colours and layout, then export PNG, TIFF, SVG or PDF.

**Do I need an account to use this template?** Not to view it here. To open and edit it, sign in to LabCanvas or create a free account (the free plan keeps up to 3 designs). Downloading the figure is part of LabCanvas Pro (US$19 a month or US$199 a year).