Chapter 33 — Case Study 2: AlphaFold and the Protein Folding Problem

"The 2020 release of AlphaFold 2 by DeepMind solved, in one stroke, the most important computational problem in biology. From sequence, predict structure — a problem that had occupied tens of thousands of scientists for sixty years was now answered with near-experimental accuracy. Within two years, ~200 million proteins had their structures predicted and made publicly available. The age of structural ignorance ended." — paraphrase from a 2024 review

The protein folding problem — given an amino acid sequence, predict the 3D structure — was identified as a central problem in computational biology in the 1960s. Anfinsen's experiment (1961, Nobel 1972) had proved that sequence determines fold; the challenge was to predict it. For decades, progress was slow. In 2020, AlphaFold solved it. This case study traces the problem and its solution, with focus on the chemistry.

The protein folding problem (1961-2020)

The challenge

A typical 100-residue protein has 10⁶⁰ possible conformations (3 conformations per residue, 100 residues). Random search through this space would take longer than the age of the universe. Yet proteins fold in milliseconds to seconds.

The folding question has two parts: 1. Forward (kinetic): how does a protein find its native fold quickly? 2. Inverse (predictive): given a sequence, what is the native fold?

The forward problem (Levinthal's paradox) was solved conceptually by recognizing that folding follows funneled energy landscapes (partial folding intermediates).

The inverse problem (sequence → structure) was harder. From the 1960s through 2010s, progress was slow: - Homology modeling (use a known structure of a similar protein): worked for some proteins. - Threading methods (try fitting the sequence onto known structures): improved, but limited. - Ab initio methods (compute energy from physics): possible only for small proteins. - Coevolutionary methods (residues that mutate together likely contact each other): mid-2010s; promising.

Each year, the CASP (Critical Assessment of Protein Structure Prediction) competition tracked progress. In CASP10 (2012), the best predictions had ~30% accuracy on hard targets. CASP13 (2018) saw modest improvements.

Then in 2020, CASP14:

AlphaFold 2: the breakthrough

DeepMind's AlphaFold 2 (Jumper, Hassabis, et al., Nature 2021) achieved near-experimental accuracy at CASP14. The model: - Was trained on the Protein Data Bank (~200,000 known protein structures). - Used multiple sequence alignments (MSAs) from related proteins. - Used coevolutionary signals (residues that mutate together). - Applied a transformer-style neural network architecture. - Outputs predicted 3D coordinates with confidence scores (pLDDT, predicted local distance difference test).

For ~95% of proteins, AlphaFold 2 predicted structure with backbone accuracy < 1.5 Å — essentially the resolution of experimental X-ray crystallography for small proteins.

This was a step-change. The CASP community recognized it immediately as the solution to the protein folding problem.

The 2022 release: AlphaFold DB

In 2022, DeepMind released the AlphaFold Database (https://alphafold.ebi.ac.uk/) with structures predicted for ~200 million proteins — essentially every known protein from every sequenced organism.

Highlights: - All ~20,000 human proteins predicted. - All proteins from major model organisms (E. coli, yeast, fly, mouse, etc.). - Many "dark proteins" with no known relatives now structurally characterized.

The database is freely available, downloadable, and updated continuously. Within months of release, it became the most-used resource in structural biology.

What AlphaFold "knows" about chemistry

AlphaFold's network learned, implicitly from training data, the chemistry of protein folding:

  1. Hydrophobic packing: nonpolar side chains cluster in the protein interior.
  2. Hydrogen bonding: backbone amides and polar side chains form specific H-bond networks (α-helices, β-sheets).
  3. Electrostatics: oppositely charged side chains form salt bridges; same-charged repel.
  4. Disulfide bonds: Cys-Cys pairings in oxidatively-folded proteins.
  5. Geometric constraints: bond angles, dihedral angles, chiral preferences.
  6. Coevolutionary signal: residues that contact each other tend to mutate together (because mutating one alone breaks the contact; mutating both compensates).

AlphaFold did not "learn" the physics from first principles. Instead, it learned to recognize the patterns of folded structures from data. But the patterns it learned correspond to the physical chemistry described in this chapter.

AlphaFold's limitations

AlphaFold is excellent for: - Single-chain proteins with homologs in the PDB. - Stable, well-defined folds. - Globular proteins.

AlphaFold struggles with: - Intrinsically disordered proteins (IDPs) — proteins without a single defined structure. - Proteins with novel folds — no homologs in training data. - Conformational dynamics — proteins that change shape during function. - Multi-protein complexes — interactions between different proteins (AlphaFold-Multimer addresses some of this). - Membrane proteins with unusual lipid interactions. - Protein-ligand complexes — small-molecule binding modes (improving with AlphaFold 3).

AlphaFold 3 (2024)

DeepMind released AlphaFold 3 in 2024, extending capabilities to: - Protein-protein complexes. - Protein-ligand binding (small-molecule docking). - Protein-DNA / protein-RNA interactions. - Post-translational modifications.

AlphaFold 3's accuracy on these new targets is impressive but more variable than AlphaFold 2's protein folding accuracy. It is now widely used in drug discovery for predicting protein-ligand poses.

Impact on biology and medicine

AlphaFold has transformed:

Drug discovery

  • Target structure determination: previously took years of X-ray crystallography. Now predicted in seconds.
  • Lead compound design: structure-based design accelerated.
  • Druggable target identification: more proteins now have known structure → more targets.

Basic biology

  • Functional characterization of unknown proteins: structure suggests function (similar fold = likely similar function).
  • Evolutionary analysis: comparing predicted structures across organisms.
  • Mechanism elucidation: many biochemical questions answered structurally.

Vaccine design

  • Antigen structure prediction: useful for designing immunogens (e.g., COVID vaccines).
  • Antibody-antigen complex modeling: with AlphaFold-Multimer.

Enzyme engineering

  • Designing new enzymes (or modifying existing ones) for industrial purposes.
  • Predicting effects of mutations on stability and activity.

The Nobel Prize 2024

The 2024 Nobel Prize in Chemistry was awarded to Demis Hassabis and John Jumper (DeepMind, AlphaFold) for "computational protein design and structure prediction" — sharing the prize with David Baker (University of Washington, Rosetta) for related work on de novo protein design.

This was the first major Nobel Prize in chemistry given for an AI achievement. It marked a watershed moment: machine learning + organic chemistry can solve grand-challenge problems.

What's left to do

Even with AlphaFold, big challenges remain: - Predicting dynamics: how proteins move and respond to ligands. - Predicting effects of mutations: how disease variants alter structure. - De novo protein design: creating new proteins with desired functions (Baker lab's specialty). - Predicting protein-protein and protein-ligand interactions at scale. - Predicting RNA structure (AlphaFold for RNA is in development).

The next decade of structural biology will be about applying AlphaFold's structural predictions, rather than trying to determine more structures experimentally.

Take-home

  • The protein folding problem (sequence → structure) was the central open problem in computational biology for 60 years.
  • AlphaFold 2 (DeepMind, 2020) solved it with near-experimental accuracy.
  • The AlphaFold Database (2022) provides predicted structures for ~200 million proteins, freely available.
  • AlphaFold uses deep learning + multiple sequence alignment + coevolutionary signal + geometric reasoning.
  • Implicit chemistry learned: hydrophobic packing, hydrogen bonding, electrostatics, disulfides, geometric constraints.
  • Limitations: intrinsically disordered proteins, novel folds, dynamics, multi-protein complexes.
  • AlphaFold 3 (2024) extends to protein-ligand, protein-DNA/RNA, and modifications.
  • Nobel Prize 2024 to Hassabis, Jumper, Baker for computational protein structure work.
  • Impact: drug discovery, basic biology, vaccine design, enzyme engineering all transformed.
  • Mastery of Chapter 33 amino acid and protein chemistry is what AlphaFold "implicitly knows" — and what underlies its predictions.