Protein Structure Prediction and AI

Protein structure prediction is the computational determination of a protein’s three-dimensional shape from its amino acid sequence. For half a century, this was considered one of biology’s hardest unsolved problems. In 2020, DeepMind’s AlphaFold 2 achieved near-experimental accuracy — a solution that has since reshaped the life sciences more than any other computational development in decades.

Why Protein Structure Matters

Proteins are the molecular machines of life. They are assembled from sequences of 20 amino acids, but their 3D folded shape — not their sequence — determines their function:

• Enzymes catalyze biochemical reactions based on the geometry of their active sites
• Receptor proteins bind other molecules in shape-complementary pockets
• Structural proteins provide mechanical support based on their 3D architecture
• Antibodies recognize pathogens via shape-specific binding

Knowing a protein’s structure is therefore essential for:

• Drug design: Designing molecules that fit into a protein’s binding pocket
• Disease understanding: Misfolded proteins underlie Alzheimer’s, Parkinson’s, and prion diseases
• Enzyme engineering: Designing catalysts for industrial chemistry and biofuels
• Vaccine development: Understanding pathogen surface proteins

The Protein Folding Problem

A protein with 100 amino acids has ($\sim$ 3 possible conformations per residue) $= 3^{100} \approx 5 \times 10^{47}$ possible backbone configurations. Even at $10^{13}$ configurations per second, exhaustive search would take longer than the age of the universe.

Yet proteins fold reliably in milliseconds in biological conditions. Anfinsen’s dogma (supported by experiments) states that the primary sequence (amino acid order) contains sufficient information to determine the native 3D structure.

The challenge: recover that structure computationally.

AlphaFold 2: The Breakthrough

AlphaFold 2 (Jumper et al., Nature 2021) won the CASP14 (Critical Assessment of protein Structure Prediction) competition with a median score of 92.4/100 on the global distance test — roughly equivalent to experimental X-ray crystallography accuracy for many proteins. Scores of ~90+ are considered “solved.”

Architecture

AlphaFold 2 introduces two novel data representations:

1. Multiple Sequence Alignment (MSA): Aligning the target sequence against thousands of evolutionarily related proteins. Co-evolutionary patterns (residues that mutate together) reveal which amino acids are spatially close in the 3D structure.

2. Pair Representation: A 2D matrix capturing relationships between every pair of residues — essentially a learned distance map.

These feed into Evoformer blocks — specialized Transformer modules that enable attention to flow both within rows and columns of the MSA and pair representations simultaneously, allowing each representation to update the other through iterative refinement.

$$\text{MSA} \xrightarrow[\text{row/column attention}]{\text{Evoformer}} \text{MSA}’ + \text{Pair}’$$

A Structure Module then uses invariant point attention (IPA) — attention that is equivariant to rotations and translations — to directly predict 3D atomic coordinates from the refined representations.

Recycling

The predicted structure is fed back as additional input for multiple refinement rounds — allowing the model to iteratively correct early errors.

AlphaFold Database

DeepMind and EMBL-EBI released the AlphaFold Protein Structure Database (2021, expanded 2022):

• Over 214 million protein structures — covering nearly the entire known proteome of every organism with a sequenced genome
• Freely accessible, transforming structural biology overnight
• Used in 1M+ research projects within two years of release

AlphaFold 3 and Biomolecular Interactions

AlphaFold 3 (Abramson et al., Nature 2024) extends predictions beyond proteins to:

• Protein–DNA and protein–RNA complexes
• Protein–small molecule (ligand) interactions — directly relevant to drug discovery
• Protein–protein interactions (complexes)
• Covalent modifications (glycosylation, phosphorylation)

AlphaFold 3 uses a diffusion-based architecture rather than direct coordinate regression, generating structural ensembles that capture conformational flexibility.

Other Notable Systems

ESMFold is particularly significant: it uses evolutionary scale modeling (ESM) — a protein language model pre-trained on the entire UniRef database — as a foundation. Structural information is extracted directly from the learned protein representations without MSA construction.

Protein Language Models

Analogous to LLMs for text, protein language models (PLMs) are trained on massive databases of amino acid sequences using masked language modeling:

$$P(\text{masked residue} | \text{rest of sequence})$$

Pre-trained PLMs learn evolutionary constraints, structural preferences, and functional correlates from sequence alone. Fine-tuning enables:

• Property prediction: Thermostability, solubility, activity
• Function annotation: Predicting enzyme class or pathway
• Fitness modeling: Predicting mutational effects

Key PLMs: ESM-2 (650M–15B parameters, Meta), ProtTrans (ProtBERT, ProtT5), Ankh.

Inverse Protein Folding: Designing New Proteins

The complementary task to structure prediction is inverse folding — given a desired 3D structure, design an amino acid sequence that will fold into it. This enables de novo protein design:

• ProteinMPNN (Baker Lab): Message passing neural network that designs sequences for target backbone structures; used to design novel enzymes, binders, and nanomaterials
• RFDiffusion: Diffusion model for generating entirely new protein backbone structures with desired properties
• Chroma: Generative model for programmable protein design conditioned on structural and functional specifications

Experimental validation in 2022 confirmed AI-designed proteins that bind influenza hemagglutinin and Clostridioides difficile toxins — novel functionalities not found in nature.

Impact on Drug Discovery

The combination of AlphaFold-quality structure prediction and generative molecular design enables a new paradigm:

1. Identify a disease target gene
2. Predict its protein structure with AlphaFold
3. Identify the binding pocket computationally
4. Use generative chemistry (DiffSBDD, AutoDock) to design molecules that fit the pocket
5. Predict binding affinity with ML models
6. Synthesize and test top candidates

This cycle is now measured in weeks rather than years for the computational portion.

Open Challenges

• Disordered regions: ~30% of human proteins are intrinsically disordered — they have no fixed 3D structure. AlphaFold flags these (low pLDDT scores) but cannot predict ensemble behavior
• Conformational dynamics: Proteins are not static; they flex and change shape. Predicting the full energy landscape remains challenging
• Multi-state proteins: Some proteins have multiple functionally distinct conformations (e.g., active/inactive states of GPCRs)
• Large complexes: Protein complexes with many chains and cofactors still challenge prediction accuracy

Further Reading

• Jumper et al. (2021), Highly Accurate Protein Structure Prediction with AlphaFold — Nature
• Abramson et al. (2024), Accurate Structure Prediction of Biomolecular Interactions with AlphaFold 3 — Nature
• Rives et al. (2021), Biological Structure and Function Emerge from Scaling Unsupervised Learning to 250 Million Protein Sequences (ESM)
• Watson et al. (2023), De Novo Design of Protein Structure and Function with RFdiffusion — Nature