AlphaFold3

The paper changes AlphaFold from a protein-structure predictor into a unified biomolecular complex predictor with an atom-level diffusion architecture.

Metadata

Field	Value
Paper	Accurate structure prediction of biomolecular interactions with AlphaFold 3
Authors	Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore, Andrew J Ballard, Joshua Bambrick, Sebastian W Bodenstein, David A Evans, Chia-Chun Hung, Michael O’Neill, David Reiman, Kathryn Tunyasuvunakool, Zachary Wu, Akvilė Žemgulytė, Eirini Arvaniti, Charles Beattie, Ottavia Bertolli, Alex Bridgland, Alexey Cherepanov, Miles Congreve, Alexander I Cowen-Rivers, Andrew Cowie, Michael Figurnov, Fabian B Fuchs, Hannah Gladman, Rishub Jain, Yousuf A Khan, Caroline M R Low, Kuba Perlin, Anna Potapenko, Pascal Savy, Sukhdeep Singh, Adrian Stecula, Ashok Thillaisundaram, Catherine Tong, Sergei Yakneen, Ellen D Zhong, Michal Zielinski, Augustin Žídek, Victor Bapst, Pushmeet Kohli, Max Jaderberg, Demis Hassabis, John M Jumper
Year	2024
Venue	Nature
Paper	Nature
DeepMind page	AlphaFold
Code	google-deepmind/alphafold3
Status	full note started

One-Line Takeaway

AlphaFold3 replaces the AlphaFold2-style structure module with a diffusion-based coordinate generator and expands the input/output contract from protein chains to biomolecular complexes containing proteins, nucleic acids, small molecules, ions, and modified residues.

Question

AlphaFold2 answered a protein monomer structure prediction problem:

$$\text{protein sequence}\to \text{protein structure}.$$

Many biological and drug-discovery questions are not monomer problems. They ask for a joint structure:

$$\text{proteins}+\text{DNA/RNA}+\text{ligands}+\text{ions}+\text{modifications}\to \text{complex geometry}.$$

The architecture question:

Can one model represent mixed biomolecular entities and generate their joint atom-level coordinates?

The paper’s answer:

$$\text{unified atom representation}+\text{pair reasoning}+\text{diffusion coordinate generation}\Rightarrow \text{joint biomolecular structure prediction}.$$

Main Claim

The narrowed architecture claim is:

$$\text{entity-aware biomolecular input}+\text{token/pair trunk}+\text{atom-level diffusion module}\Rightarrow \text{accurate complex prediction across molecular types}.$$

This is different from a docking-only claim. A docking pipeline often assumes:

$$\text{fixed receptor}+\text{ligand conformer/search}\to \text{pose}.$$

AlphaFold3 instead predicts the joint complex:

$${\hat{X}}_{\text{complex}}=f_{\theta }(\mathcal{E}),$$

where $\mathcal{E}$ is a set of molecular entities and ${\hat{X}}_{\text{complex}}$ is the predicted coordinate set.

Architecture Contract

Component	Role
entity input	proteins, nucleic acids, ligands, ions, modified residues
atom/token features	represent mixed molecular objects
pair representation	stores relationships between residues, atoms, or tokens
trunk	reasons over biomolecular context before coordinate generation
diffusion module	denoises atom coordinates into a complex structure
confidence heads	estimate reliability and interaction quality
sampling/refinement	produces candidate structures through generative denoising

The output is atom-level geometry:

$$\hat{X}\in {\mathbb{R}}^{N_{\text{atoms}}\times 3}.$$

For any global rigid transform:

$$X^{\prime }=RX+t,$$

the structural claim should be unchanged. Complex prediction therefore has the same rigid-transform symmetry issue as protein structure prediction, but with more heterogeneous objects.

Why It Is an Architecture Paper

AlphaFold3 should be read as an architecture paper because it changes the modeling contract:

Axis	AlphaFold2	AlphaFold3
primary target	protein structure	biomolecular complex structure
output route	structure module with residue frames	diffusion-based atom coordinate generation
molecular scope	mainly proteins, then multimer variants	proteins, nucleic acids, ligands, ions, modified residues
downstream relevance	protein modeling	protein modeling, interaction modeling, SBDD
core architectural shift	Evoformer plus structure module	trunk plus diffusion module

This is not merely “AlphaFold2 with more input types.” The generative coordinate module changes how structure is produced.

Input Object Contract

A biomolecular complex can be represented as:

$$\mathcal{C}=(\mathcal{E},A,B,F),$$

where:

• $\mathcal{E}$ is the set of molecular entities;
• $A$ is the set of atoms;
• $B$ stores bonds or connectivity;
• $F$ stores features such as residue type, atom type, chain, entity type, charge, or modification.

For proteins:

$$e_{\text{protein}}=(s_1,\dots ,s_L).$$

For ligands:

$$e_{\text{ligand}}=(G_{\text{mol}},\text{chemistry features}),$$

where $G_{\text{mol}}$ is a molecular graph.

For nucleic acids:

$$e_{\text{NA}}=(b_1,\dots ,b_L),$$

where $b_i$ is a nucleotide token.

The hard part is not just concatenating inputs. The model must preserve molecular identity, connectivity, symmetry, and interaction context.

Unified Coordinate Prediction

AlphaFold3 predicts a joint coordinate set:

$$\hat{X}=\{{\hat{x}}_1,\dots ,{\hat{x}}_N\},{\hat{x}}_i\in {\mathbb{R}}^3.$$

The coordinate set includes atoms from multiple entities. A protein-ligand complex is not:

$${\hat{X}}_{\text{protein}}\text{ first, then dock ligand}.$$

It is closer to:

$$({\hat{X}}_{\text{protein}},{\hat{X}}_{\text{ligand}})=f_{\theta }(\text{protein},\text{ligand},\text{context}).$$

This matters because interactions can affect relative placement and local conformations.

Diffusion Module

Diffusion models learn to recover clean data from noise. For coordinates:

$$X_t={\alpha }_t{X}_0+{\sigma }_tϵ,ϵ\sim \mathcal{N}(0,I).$$

The denoising model predicts either the clean coordinates, the noise, or a related update:

$${ϵ}_{\theta }(X_t,t,c),$$

where $c$ is conditioning information from the trunk.

A reverse denoising step can be written abstractly as:

$$X_{t-1}=g_{\theta }(X_t,t,c).$$

For AlphaFold3 reading, the important contract is:

$$\text{noisy atom cloud}+\text{biomolecular context}\to \text{less noisy atom coordinates}.$$

The model generates a structure through iterative coordinate refinement rather than a single deterministic residue-frame output.

Why Diffusion Helps Here

Biomolecular complexes have multi-modal uncertainty:

• ligands can bind in different poses;
• side chains can rearrange;
• loops can move;
• complexes can have ambiguous interfaces;
• multiple conformations may be plausible;
• input context can be incomplete.

Diffusion gives an architecture for conditional structure generation:

$$p_{\theta }(X\mid \mathcal{E})$$

rather than only a point estimate:

$$\hat{X}=f_{\theta }(\mathcal{E}).$$

In practice, evaluation still often reports best or ranked predicted structures. The paper should not be read as solving conformational ensemble modeling completely.

Token and Atom Representations

Mixed biomolecular modeling needs at least two levels of representation:

Level	Examples	Why Needed
token/entity level	residues, ligand tokens, nucleotides	long-range context and sequence/entity structure
atom level	atom coordinates and features	final geometry, contacts, pose, chemistry
pair level	token-token or atom-atom relations	interactions, distances, interface reasoning

A generic pair representation:

$$z_{ij}\in {\mathbb{R}}^{d_z}$$

stores a relation between object $i$ and object $j$.

For interaction prediction, pair features are central:

$$\text{binding mode}\approx \text{consistent cross-entity geometry}.$$

Pair Reasoning

A biomolecular complex can be treated as a dense relational object:

$$G=(V,E),E=V\times V.$$

Pair state:

$$z_{ij}\leftrightarrow \text{relation between object }i\text{ and }j.$$

For a protein-ligand complex, cross-entity pairs are crucial:

$$z_{i,a}^{\text{protein-ligand}}$$

where $i$ is a protein residue/atom and $a$ is a ligand atom/token.

The architecture must distinguish:

Relation	Meaning
protein-protein	interface, multimer packing
protein-ligand	pocket pose and contacts
protein-DNA/RNA	sequence-specific binding, backbone geometry
ligand internal	molecular graph and conformer validity
ion/modified residue	coordination and local chemistry

Relation to AlphaFold2

AlphaFold2:

$$s,M,T\to \text{MSA/pair trunk}\to \text{structure module}\to {\hat{X}}_{\text{protein}}.$$

AlphaFold3:

$$\mathcal{E}\to \text{biomolecular trunk}\to \text{diffusion module}\to {\hat{X}}_{\text{complex}}.$$

The key shift:

$$\text{protein-specific coordinate prediction}\to \text{general biomolecular coordinate generation}.$$

Relation to Docking

Classical docking often separates receptor preparation, ligand conformer generation, search, and scoring:

$$\text{receptor}+\text{ligand}\to \text{pose candidates}\to \text{score/rank}.$$

AlphaFold3 is closer to joint prediction:

$$\text{protein}+\text{ligand}\to \text{complex structure}.$$

This distinction matters:

Docking View	AlphaFold3 View
receptor may be fixed	complex can be predicted jointly
ligand pose search explicit	coordinate generation learned
scoring function central	model confidence and learned structure distribution central
usually ligand-focused	all biomolecular entities share a prediction framework

But AlphaFold3 should not be treated as replacing every docking workflow. Docking tasks often involve screening millions of ligands, protonation/tautomer states, induced fit, water networks, affinity ranking, and assay-specific validation.

Relation to Diffusion Models

Generic diffusion:

$$x_T\sim \mathcal{N}(0,I),x_{t-1}=g_{\theta }(x_t,t,c).$$

AlphaFold3-style structure diffusion:

$$X_T\sim \text{noisy atom coordinates},X_{t-1}=g_{\theta }(X_t,t,\mathcal{E}).$$

The conditioning object is scientific:

$$c=\text{molecular identity, connectivity, templates, sequence, pair context}.$$

The output is constrained by chemistry and geometry, not just image realism.

Relation to Geometric Deep Learning

Coordinate prediction should respect rigid transforms:

$$X^{\prime }=RX+t.$$

If the input coordinate frame is rotated and translated, coordinate outputs should transform consistently:

$$f(RX+t)=Rf(X)+t.$$

Scalar confidence or interaction probabilities should be invariant:

$$c(RX+t)=c(X).$$

AlphaFold3 is a major example of generative geometric modeling for scientific structures.

Evidence Reading

The paper evaluates many interaction categories. The evidence should be separated by claim:

Evidence Type	Supports	Does Not Prove Alone
protein-ligand benchmarks	improved pose prediction for some drug-like interactions	reliable affinity ranking
protein-protein/nucleic-acid complexes	broader biomolecular complex modeling	all biological assemblies are solved
comparison to specialized tools	one model can compete across categories	every category is equally strong
confidence estimates	some reliability signal	all hallucinations are detected
PoseBusters-style evaluation	physical plausibility checks	experimental binding or activity

The paper is strong because it broadens the target distribution. That also makes evaluation harder: each entity type and interaction type has different failure modes.

Benchmark Boundary

For SBDD, distinguish:

$$\text{pose accuracy}\ne \text{binding affinity}\ne \text{virtual screening enrichment}\ne \text{prospective drug discovery success}.$$

AlphaFold3 can improve pose or complex prediction while still leaving:

• affinity calibration;
• activity cliffs;
• assay transfer;
• ligand protonation;
• water/metal coordination;
• induced fit;
• conformational ensembles;
• prospective validation.

This boundary is important for public notes because AlphaFold3 attracts broad claims that can easily become too strong.

Implementation and Reproducibility Reading

When reading or trying to use AlphaFold3-like systems, separate:

Question	Why It Matters
Are model weights available?	determines reproducibility and local deployment
Is inference code available?	determines implementation auditability
Are inputs standardized?	ligands, CCD codes, SMILES, modifications, templates
Are multiple samples generated?	diffusion outputs may vary
Are confidence scores calibrated?	ranking and filtering depend on reliability
Is the benchmark leakage-controlled?	PDB time splits and homolog/template leakage matter

This is not just an architecture issue. For scientific modeling, access, input format, and evaluation protocol are part of the claim contract.

Failure Modes

Failure Mode	Why It Matters
treating a plausible generated complex as experimental truth	generative models can produce convincing but wrong structures
using predicted pose as affinity	pose and binding energy are different claims
ignoring ligand chemistry state	protonation, tautomer, stereochemistry, and charge affect binding
ignoring water and cofactors	missing context can change the interface
comparing against docking without matching inputs	some methods may use different receptor/template information
mixing server output and paper architecture claims	access route and model version may differ
assuming all molecular types are equally solved	performance varies by interaction category

Common Misreadings

”AlphaFold3 solves docking.”

No. It makes a major advance in complex structure prediction, including protein-ligand interactions, but docking workflows include screening, affinity ranking, chemical state handling, and prospective validation.

”Diffusion means it models all conformational ensembles.”

No. Diffusion is a generative architecture, but practical prediction may still produce limited samples and can miss biologically relevant states.

”One model for many molecule types means one evaluation metric is enough.”

No. Protein-protein interfaces, ligand poses, nucleic-acid complexes, and ion coordination need different checks.

”High confidence means the interaction is biologically real.”

No. A predicted complex can be geometrically plausible while missing biological context, concentration, cellular state, or assay evidence.

Later-Paper Checklist

When reading post-AlphaFold3 papers, check:

• Does the model predict joint complexes or dock into fixed structures?
• Are proteins, ligands, nucleic acids, ions, and modifications represented in one input schema?
• Is the coordinate generator diffusion-based, flow-based, or deterministic?
• Are molecular graph constraints enforced or learned?
• Are ligand protonation, tautomer, stereochemistry, and charge explicit?
• Are templates, homologs, and PDB date splits controlled?
• Are pose, affinity, and screening claims separated?
• Is confidence calibrated per interaction type?
• Are multiple samples generated and ranked?
• Is there prospective or held-out temporal validation?

Why It Matters

AlphaFold3 is a major architecture transition:

$$\text{structure prediction}\to \text{interaction structure generation}.$$

For this wiki, it connects:

$$\text{AlphaFold2}\to \text{diffusion models}\to \text{protein-ligand modeling}\to \text{structure-based drug discovery}.$$

It is exactly the kind of paper that should be cross-listed: architecture, computational biology, protein modeling, and SBDD all need it, but the canonical note can live once in papers/architectures.

Limitations

AlphaFold3 should be treated as a powerful structure predictor, not a complete biological oracle.

Important limitations:

• generated structures still need experimental or orthogonal validation;
• binding affinity is not identical to pose quality;
• ligand chemistry state can dominate real-world behavior;
• water, ions, cofactors, and missing context can matter;
• flexible and disordered regions remain difficult;
• model access and reproducibility constraints affect scientific use;
• benchmark performance may not transfer to every target class.

The defensible claim:

$$\text{AlphaFold3}\Rightarrow \text{major advance in joint biomolecular structure prediction}.$$

The overclaim to avoid:

$$\text{AlphaFold3}\Rightarrow \text{drug discovery solved}.$$

Connections

• AlphaFold2
• PoseBusters
• Protein modeling papers
• Structure-Based Modeling Papers
• Computational Biology papers
• Diffusion model
• Flow matching
• 3D structure
• Interaction prediction
• Protein structure prediction
• Protein-ligand representation contract
• Protein-ligand interaction
• Receptor-ligand preparation
• Docking workflow
• Interaction modeling
• Protein-ligand docking
• Architecture papers