AlphaFold: A Complete Overview

What Is AlphaFold?

1. How It Works

The Problem It Solves

AlphaFold2 Architecture (the breakthrough version, 2021)

AlphaFold3 (2024) — The Next Generation

• Diffusion-based architecture — replaces the Evoformer structure module with a diffusion model (similar to image-generation AI), generating 3D atomic coordinates from noisy inputs
• Unified biomolecular modeling — predicts proteins, DNA, RNA, small molecules (ligands), ions, and post-translational modifications in a single model
• Protein–ligand docking — directly predicts how drugs bind to their protein targets
• Protein–nucleic acid complexes — predicts how proteins interact with DNA and RNA
• Multi-chain assemblies — handles large macromolecular complexes

• Still struggles with intrinsically disordered regions and proteins with multiple conformational states
• Not ideal for predicting dynamic conformational changes (it gives a single predicted state)
• Access to the full model weights has been restricted — only a web server is available for non-commercial use, which sparked controversy in the open-science community

Recent review: Krokidis et al. (2025) — AlphaFold3: An Overview of Applications and Performance Insights [PMID: 40332289]

2. Impact on Biology & Medicine

Structural Biology

• Solved structures for proteins that resisted decades of experimental work
• Accelerated cryo-EM studies — AlphaFold models are used as starting templates for fitting into electron density maps
• Enabled structural genomics at scale — whole-proteome coverage for any organism

Drug Discovery

• Target identification — rapidly models disease-relevant proteins (including previously "undruggable" targets)
• Virtual screening — AlphaFold3's docking capabilities allow predicting small molecule binding poses
• Antibody/epitope design — predicts protein surface features for vaccine and antibody development
• Major pharma companies (including Isomorphic Labs, a DeepMind spinout) are using it in active drug pipelines

Infectious Disease & Neglected Diseases

• Modeled structures of pathogens with no prior structural data (e.g., neglected tropical disease proteins)
• Used extensively in COVID-19 research (spike protein interactions, protease inhibitors)

Evolutionary Biology

• Revealed structural homologs across distantly related organisms, uncovering ancient evolutionary relationships invisible at sequence level

Nobel Prize (2024)

3. The AlphaFold Database

• Contains >200 million protein structure predictions — essentially the entire UniProt database
• Covers proteomes of humans, model organisms, pathogens, and more
• Freely available at AlphaFold Database (EMBL-EBI)

1. Search by gene name, UniProt ID, or organism on the web interface
2. Download individual structures as PDB or mmCIF files
3. Bulk downloads available for entire proteomes
4. Structures can be opened in PyMOL, ChimeraX, UCSF Chimera, or any molecular viewer
5. The pLDDT coloring (blue = confident, red = low confidence) guides interpretation

https://alphafold.ebi.ac.uk/api/prediction/{UniProt_ID}

4. The Broader AI Protein Ecosystem

Summary

AlphaFold: Newest Updates (2025–2026)

1. AlphaFold3 Source Code Released (Late 2024 → Impact in 2025–2026)

2. AlphaFold Database (AFDB) 2025 Redesign

3. AlphaFold3 Benchmarked at CASP16 (2025–2026)

• Better confidence/pLDDT estimation — improved model selection compared to AF2
• Protein complexes — slightly outperforms AF2 for multi-chain assemblies
• Easier targets — the AF3 server ranks above AF2-based methods
• User-friendliness — competitive with state-of-the-art in all CASP categories from a single web server

• Hard targets — AF2 with massive sampling can match or beat AF3
• Heteromeric complexes — stoichiometry prediction accuracy remains limited for large assemblies
• Nucleic acid predictions — accuracy is relatively low (though competitive with best available tools)
• Model ranking — selecting the best model from multiple candidates remains a persistent weakness
• Truncated sequences, irregular secondary structures, interaction-induced conformational changes

4. What Remained Unsolved (Active Frontiers)

• Single-domain fold prediction is essentially solved — no incorrect folds predicted across all CASP16 evaluation units
• Remaining hard problems: disordered regions, multi-state/dynamic conformations, large heteromeric complex stoichiometry, protein–RNA interactions at high accuracy

Summary Table

AlphaFold 1 vs AlphaFold 2 vs AlphaFold 3

At a Glance

1. AlphaFold 1 (2018)

• Used at CASP13 — first time a DeepMind system competed in protein structure prediction
• Architecture combined convolutional neural networks (CNNs) with gradient boosting to predict inter-residue distances
• Generated a distance map (how far apart pairs of residues are), then used energy minimization to fold the protein into 3D
• Did NOT directly output 3D coordinates — used the predicted distances as constraints for a separate optimization step
• Outperformed all competitors at CASP13 by a large margin — but still far from experimental accuracy
• Established that deep learning could meaningfully crack the folding problem

2. AlphaFold 2 (2021)

What changed architecturally:

• Replaced the CNN+gradient boosting pipeline with an end-to-end differentiable deep learning system
• Introduced the Evoformer — a transformer that jointly processes:
  • MSA (Multiple Sequence Alignment): evolutionary information from related protein sequences in other species
  • Pair representation: all pairwise residue–residue relationships
• Added a Structure Module that directly outputs 3D Cα coordinates using equivariant neural networks (respects rotational/translational symmetry)
• Recycling: the predicted structure is fed back into the network and refined iteratively (3 cycles)
• Output includes pLDDT (per-residue confidence, 0–100) and PAE (Predicted Aligned Error — for domain orientation confidence)

Performance at CASP14:

• Achieved ~92.4 GDT_TS on free-modeling targets — within experimental error for most proteins
• So dominant that many CASP categories effectively became obsolete
• Described as solving "a 50-year-old grand challenge of biology"

What it could NOT do:

• Only predicts single proteins or homo-oligomers (initially)
• No support for DNA, RNA, or small molecule ligands
• Gives a single static structure — no conformational dynamics
• Struggles with intrinsically disordered proteins (IDPs)
• ColabFold and multimer extensions came later as community add-ons

3. AlphaFold 3 (2024)

What changed architecturally:

• Keeps the Evoformer for sequence/evolutionary processing
• Replaces the Structure Module with a diffusion model (like image-generation AI — generates 3D atom coordinates from a noisy distribution)
• Uses a unified token representation for all molecule types — proteins, nucleic acids, small molecules, ions, and post-translational modifications are all treated the same way
• Handles joint co-folding of mixed complexes (e.g., a drug molecule docked inside a protein, with a DNA strand also present)

New capabilities vs AF2:

CASP16 performance (2025–2026 independent benchmark):

• Better than AF2 for easier monomer targets and protein complexes
• Comparable to AF2 (with massive sampling) on hard targets
• Nucleic acid predictions — competitive but still low absolute accuracy
• Heteromeric complex stoichiometry — still a weakness
• Model selection (picking the best of 5 predicted structures) remains a known gap

Architecture Evolution Summary

AF1 (2018)
  MSA → CNN → Distance map → Energy minimization → 3D structure
  (two disconnected stages)

AF2 (2021)
  MSA + Pair representation → Evoformer → Structure Module → 3D Cα coords
  (end-to-end, one unified system)

AF3 (2024)
  MSA + Pair representation → Evoformer → Diffusion Model → Full-atom 3D coords
  (unified across ALL molecule types)

Impact Comparison

Bottom Line

• AF1 → proved deep learning could tackle protein folding
• AF2 → solved single-protein structure prediction to near-experimental accuracy; transformed structural biology
• AF3 → extended the paradigm to the entire biomolecular universe — proteins, nucleic acids, small molecules, and their interactions — enabling true computational drug discovery