Protein AI Series Part 0: What Is Protein Structure Prediction?
The Technical Evolution of Protein AI — A Record of Key Design Decisions
This is Part 0 of a 10-part series tracing the architectural choices behind modern protein structure prediction and design models.
Introduction
Every protein in biology begins as a linear chain of amino acids. Yet this one-dimensional sequence encodes a three-dimensional structure that determines the protein’s function — catalyzing reactions, transducing signals, recognizing pathogens. The central question of structural biology has long been: given only the amino acid sequence, can we predict the 3D structure?
This series traces how AI models have answered that question — and how their answers have evolved from AlphaFold2 (2021) to the latest generation of co-folding and design models in 2025–2026. Each installment examines a specific architectural decision point and compares how different model families responded. Before diving into those comparisons, this Part 0 establishes the shared vocabulary and conceptual framework that every subsequent post builds on.
1. Protein Structure Fundamentals
1.1 Amino Acids and the Peptide Bond
Proteins are linear polymers of amino acids. Nature uses 20 standard amino acid types, each sharing an identical backbone but carrying a distinct side-chain (R group):
1
2
3
4
5
H O
| ‖
H₂N─Cα──C─OH
|
R (side-chain, unique per amino acid type)
Adjacent amino acids are linked by a peptide bond between the carbonyl carbon of residue $i$ and the amide nitrogen of residue $i+1$:
1
2
...─[NH─Cα─C(═O)]─[NH─Cα─C(═O)]─[NH─Cα]─...
residue i residue i+1 residue i+2
The repeating N–Cα–C pattern forms the backbone. The variable R groups attached at each Cα are the side-chains. Every residue contributes four backbone heavy atoms: N, Cα, C, and O.
1.2 Torsion Angles: The Internal Coordinates of Structure
Rather than specifying Cartesian coordinates for every atom, protein structure can be compactly described by torsion (dihedral) angles — the rotation around each covalent bond.
Backbone torsion angles (three per residue):
\[\phi_i: \quad \text{C}_{i-1} - \text{N}_i - \text{C}_{\alpha,i} - \text{C}_i \qquad \in [-\pi, +\pi]\] \[\psi_i: \quad \text{N}_i - \text{C}_{\alpha,i} - \text{C}_i - \text{N}_{i+1} \qquad \in [-\pi, +\pi]\] \[\omega_i: \quad \text{C}_{\alpha,i} - \text{C}_i - \text{N}_{i+1} - \text{C}_{\alpha,i+1} \qquad (\approx \pi, \text{trans})\]Since bond lengths and bond angles are nearly fixed across all proteins, specifying $(\phi, \psi)$ for each residue (with $\omega \approx \pi$) is sufficient to reconstruct the full backbone trace.
Side-chain torsion angles $\chi_1, \chi_2, \chi_3, \chi_4$ (0–4 per residue depending on amino acid type) describe the rotational state of the side-chain. Discretized combinations of $\chi$ angles are called rotamers.
The Ramachandran plot — a 2D scatter of $(\phi, \psi)$ values — reveals that only certain regions of torsion-angle space are sterically permitted:
1
2
3
4
5
6
7
8
ψ
↑
| ■■■■ ← β-sheet region (φ ≈ -120°, ψ ≈ +130°)
| ■■■■
|
| ■■■■ ← α-helix region (φ ≈ -60°, ψ ≈ -45°)
| ■■■■
+──────────────→ φ
Structure prediction models must produce $(\phi, \psi)$ pairs that fall within these physically allowed regions.
1.3 Secondary Structure
Local, repeating backbone conformations give rise to secondary structure elements:
| Element | Hydrogen Bond Pattern | Geometry | Residues per Turn |
|---|---|---|---|
| α-helix | C=O(i) ↔ N-H(i+4) | Right-handed helix, rise 5.4 Å | 3.6 |
| β-sheet | Inter-strand C=O ↔ N-H | Extended strands, parallel or antiparallel | — |
| Loop / Coil | Irregular | Flexible, often solvent-exposed | — |
Loops are the most structurally diverse and hardest to predict. Antibody CDRs (Complementarity Determining Regions), the antigen-binding loops, are canonical examples of structurally variable loops.
1.4 Tertiary, Quaternary Structure, and Rigid Body Frames
Tertiary structure is the full 3D arrangement of a single polypeptide chain — secondary structure elements packed together through hydrophobic interactions, disulfide bonds, hydrogen bonds, and ionic interactions.
Quaternary structure arises when multiple chains assemble into a complex (e.g., an antibody: 2 heavy + 2 light chains).
A key representational concept introduced by AlphaFold2 is the rigid body frame. Each residue is treated as a rigid body defined by:
\[T_i = (R_i, \vec{t}_i), \qquad R_i \in \mathrm{SO}(3),\ \vec{t}_i \in \mathbb{R}^3\]where $R_i$ is a $3 \times 3$ rotation matrix encoding the residue’s orientation, and $\vec{t}i$ is a translation vector encoding its position. The frame is constructed from the three backbone atoms N, $\text{C}\alpha$, C using a Gram-Schmidt-like procedure. Side-chain atoms are then placed relative to this frame via the predicted $\chi$ angles.
2. The Sequence-to-Structure Problem
2.1 Levinthal’s Paradox
A protein of $L$ residues has roughly $2L$ backbone torsion degrees of freedom ($\phi, \psi$ per residue). Even discretizing each angle into just 3 states yields $3^{2L}$ possible conformations. For a modest 100-residue protein:
\[3^{200} \approx 10^{95}\]Exhaustive enumeration is computationally intractable — yet real proteins fold on the millisecond-to-second timescale. This is Levinthal’s paradox: the conformational search space is astronomically large, but nature navigates it efficiently through the energy landscape encoded by the sequence.
2.2 Experimental Structure Determination
Before AI, 3D structures were determined experimentally:
| Method | Resolution | Throughput | Limitations |
|---|---|---|---|
| X-ray crystallography | ~1–2 Å | Months–years | Requires crystallization; bias toward rigid, ordered structures |
| Cryo-EM | ~2–4 Å | Weeks–months | Large complexes preferred; resolution varies across map |
| NMR spectroscopy | ~2–5 Å | Months | Limited to small proteins (~30 kDa) |
The Protein Data Bank (PDB) contains ~220K experimentally determined structures — a tiny fraction of the ~250M+ known protein sequences. This gap between sequence space and structure space is the fundamental motivation for computational structure prediction.
2.3 Why AI Succeeded Where Physics Struggled
Classical molecular dynamics (MD) simulates atomic forces at femtosecond ($10^{-15}$ s) timesteps. Folding a 100-residue protein (folding time ~ms) requires $\sim 10^{12}$ steps — feasible on specialized hardware (e.g., Anton) but impractical at scale. Moreover, force field inaccuracies accumulate over long trajectories.
AI approaches sidestep direct physics simulation by learning the sequence → structure mapping from experimental data. The key insight: evolution has already explored sequence-structure relationships across billions of years. By analyzing patterns in known sequences and structures, neural networks can infer the mapping without simulating the physical folding process.
3. The Common Architecture: Trunk + Structure Generation
Nearly every modern protein structure prediction model shares a two-stage architecture. Understanding this shared blueprint is essential for following the rest of this series.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
┌───────────────────────────────────────────────────────────────┐
│ INPUTS │
│ Sequence → one-hot / PLM embedding │
│ MSA → coevolutionary statistics │
│ Templates → known similar structures (optional) │
└──────────────────────────┬────────────────────────────────────┘
↓
┌───────────────────────────────────────────────────────────────┐
│ INPUT EMBEDDING │
│ Sequence → s_i ∈ R^{c_s} (single representation) │
│ MSA → Outer Product Mean → z_ij ∈ R^{c_z} (pair repr.) │
│ Templates → pair features added to z_ij │
└──────────────────────────┬────────────────────────────────────┘
↓
┌───────────────────────────────────────────────────────────────┐
│ TRUNK (the "representation engine") │
│ ┌─────────────────────────────────────────────────────────┐ │
│ │ Evoformer (AF2) / Pairformer (AF3, Boltz) / ... │ │
│ │ × N blocks (48–64) │ │
│ │ │ │
│ │ Iteratively refines: │ │
│ │ z_ij : pair representation (L × L × c_z) │ │
│ │ s_i : single representation (L × c_s) │ │
│ │ │ │
│ │ × R recycling iterations (typically R = 3) │ │
│ └─────────────────────────────────────────────────────────┘ │
│ │
│ Output: z_trunk, s_trunk │
└──────────────────────────┬────────────────────────────────────┘
↓
┌───────────────────────────────────────────────────────────────┐
│ STRUCTURE GENERATION MODULE │
│ │
│ AF2: IPA Structure Module (deterministic, 8 steps) │
│ AF3/Boltz: Diffusion Module (stochastic, 200 steps) │
│ NP3: Flow Matching Decoder (stochastic, 40 steps) │
│ BioEmu: Langevin Dynamics (stochastic, ~500 steps) │
│ │
│ Output: 3D atomic coordinates │
└──────────────────────────┬────────────────────────────────────┘
↓
┌───────────────────────────────────────────────────────────────┐
│ OUTPUT HEADS │
│ Confidence: pLDDT, pTM, pAE, pDE │
│ (Optional) Affinity: K_d, ΔΔG │
│ (Optional) Distogram: pairwise distance distributions │
└───────────────────────────────────────────────────────────────┘
3.1 Pair Representation $z_{ij}$
The pair representation is an $L \times L \times c_z$ tensor (typically $c_z = 128$) where each entry $z_{ij} \in \mathbb{R}^{c_z}$ encodes the relationship between residues $i$ and $j$ — their spatial proximity, relative orientation, and interaction type.
Initialization combines multiple signals:
- Relative positional encoding: function of $|i - j|$
- Outer product mean from MSA: column-pair statistics capturing coevolutionary signal
- Template pair features: Cα–Cα distances from known similar structures
After the Trunk refines $z_{ij}$ through dozens of layers, it serves as a structural blueprint: “$z_{ij}$ encodes that residues $i$ and $j$ should be within 5 Å, oriented in a particular way.” The Structure Generation Module reads this blueprint to produce 3D coordinates.
Memory cost scales quadratically:
\[\text{Memory} = L^2 \times c_z \times \text{sizeof(dtype)}\]For $L = 2000$, $c_z = 128$, BF16: $2000^2 \times 128 \times 2 \approx 976\;\text{MB}$ — just for the pair representation alone. This $O(L^2)$ scaling is the primary bottleneck limiting the size of structures that can be processed.
3.2 Single Representation $s_i$
The single representation $s \in \mathbb{R}^{L \times c_s}$ (typically $c_s = 384$) captures per-residue properties: amino acid identity, local structural environment, and aggregated information from MSA columns or PLM embeddings.
In the Pairformer architecture (AF3, Boltz-2), $s_i$ is updated via Attention with Pair Bias:
\[\text{Attn}(Q, K, V, z) = \text{softmax}\!\left(\frac{QK^\top}{\sqrt{d}} + W_z \cdot z_{ij}\right) V\]where $Q = W_q s$, $K = W_k s$, $V = W_v s$. The pair representation $z_{ij}$ acts as a learned bias on the attention weights, injecting structural context directly into the self-attention over residues.
3.3 The Trunk: Where Representations Are Built
The Trunk is the computational core of every model — the module that transforms raw input features into rich structural representations $(s, z)$. Its design has been the primary axis of architectural innovation:
| Generation | Trunk Architecture | Key Feature |
|---|---|---|
| AF2 (2021) | Evoformer | MSA + pair jointly processed |
| AF3 (2024) | Pairformer | MSA separated; cleaner pair-only trunk |
| Pairmixer (2026) | Attention-free | Triangle multiplication only; 4× faster |
Part 2 of this series examines this evolution in detail.
3.4 Structure Generation: From Deterministic to Generative
The Structure Generation Module converts the Trunk’s output $(s, z)$ into 3D atomic coordinates. This is where the most dramatic paradigm shifts have occurred:
| Paradigm | Representative | Mechanism | Nature |
|---|---|---|---|
| IPA | AF2 | SE(3)-equivariant point attention over rigid frames | Deterministic |
| EDM Diffusion | AF3, Boltz-1/2, Chai-1 | Iterative denoising: $D_\theta(x_t, \sigma; z) \approx \mathbb{E}[x_0 \mid x_t]$ | Stochastic |
| SE(3) Diffusion | RFdiffusion, FrameDiff | Noise on frames $(R, t)$ in SE(3) | Stochastic |
| Flow Matching | NP3, Proteina, FrameFlow | ODE: $\frac{dx}{dt} = v_\theta(x, t)$, linear interpolation path | Stochastic/ODE |
| Langevin | BioEmu | Score-based sampling from $p(x \mid \text{seq}) \propto e^{-E(x)/kT}$ | Stochastic |
Part 3 dives deep into these five paradigms.
4. Evaluation Metrics
To compare models, the field relies on a standard set of metrics. Understanding these is essential for interpreting benchmarks throughout the series.
4.1 Structural Accuracy Metrics
RMSD (Root Mean Square Deviation) measures the average atomic displacement between predicted and true structures after optimal superposition:
\[\text{RMSD} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \| \vec{x}_i^{\,\text{pred}} - \vec{x}_i^{\,\text{true}} \|^2}\]Typically computed over Cα atoms. Interpretation: $< 1$ Å excellent, $< 2$ Å good, $< 5$ Å approximate fold correct.
lDDT (Local Distance Difference Test) evaluates local structural accuracy without requiring global superposition. For all pairs of atoms within 15 Å in the true structure, it checks whether their predicted distances are preserved within tolerance thresholds:
\[\text{lDDT} = \frac{1}{|\mathcal{P}|} \sum_{(i,j) \in \mathcal{P}} \frac{1}{4} \sum_{\tau \in \{0.5, 1, 2, 4\}} \mathbb{1}\!\left[ |d_{ij}^{\text{pred}} - d_{ij}^{\text{true}}| < \tau \right]\]where $\mathcal{P}$ is the set of atom pairs within 15 Å in the reference. Range: $[0, 1]$, higher is better. Unlike RMSD, lDDT is robust for proteins with flexible domains.
TM-score (Template Modeling score) provides a length-normalized measure of global structural similarity:
\[\text{TM-score} = \frac{1}{L} \sum_{i=1}^{L_{\text{aligned}}} \frac{1}{1 + (d_i / d_0)^2}\]where $d_0 = 1.24 \sqrt[3]{L - 15} - 1.8$ is a length-dependent normalization. Range: $(0, 1]$; $> 0.5$ indicates the same fold, $> 0.7$ high similarity.
4.2 Complex / Docking Metrics
DockQ is a composite score for protein–protein interface quality, combining:
- $F_{\text{nat}}$: fraction of native contacts preserved
- $L_{\text{rms}}$: ligand RMSD (after interface alignment)
- $i_{\text{rms}}$: interface RMSD
Range: $[0, 1]$; $> 0.23$ acceptable, $> 0.49$ medium, $> 0.80$ high quality.
4.3 Model Confidence Scores
Modern models output self-assessed confidence alongside their predictions:
| Score | What It Estimates | Range | Interpretation |
|---|---|---|---|
| pLDDT | Per-residue lDDT | [0, 100] | > 90 very high; < 50 likely disordered |
| pTM | Global TM-score | [0, 1] | Overall fold confidence |
| pAE | Pairwise aligned error | Å | Inter-chain relative positioning |
| pDE | Pairwise distance error | Å | Atom-pair distance accuracy |
Confidence scores are critical for ranking: models like AF3 and Boltz-2 generate multiple structures with different random seeds, then rank them by confidence to select the best prediction.
4.4 Conformational Diversity Metrics
ConfBench (introduced by NP3) evaluates a model’s ability to distinguish apo (ligand-free) and holo (ligand-bound) conformations:
\[\text{score} = \frac{\text{RMSD}_{\text{alt}} - \text{RMSD}_{\text{ref}}}{\sqrt{(\text{RMSD}_{\text{alt}}^2 + \text{RMSD}_{\text{ref}}^2 + \text{RMSD}_{\text{ref-alt}}^2) / 2}}\]where $+1$ means the correct state is predicted, $0$ means indistinguishable, and $-1$ means the wrong state. Current results: AF2-Multimer ~30% (near random), NP3 ~52% — conformational diversity remains an open challenge (Part 6).
5. Series Roadmap
This post established the foundation. The remaining parts each pose a core architectural question and trace how models have diverged in answering it:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
Parts 1–3: How has the prediction pipeline evolved?
Part 1 ─ How to read evolutionary information?
MSA vs PLM vs Hybrid
Part 2 ─ How to reason about residue-pair relationships?
Evoformer → Pairformer → Pairmixer
Part 3 ─ How to output 3D structure?
IPA → Diffusion → Flow Matching
Parts 4–5: From prediction to design
Part 4 ─ Proteins only, or all biomolecules?
The rise of co-folding + open-source competition
Part 5 ─ How to turn a prediction model into a design model?
Four strategies compared
Parts 6–7: Open problems and the future
Part 6 ─ One structure or many?
The conformational diversity problem
Part 7 ─ Where is the field heading?
Points of convergence and unsolved problems
Each post follows a consistent structure: a core question → why it matters → how different models answered differently → a comparison table → current consensus and remaining open questions.
Next: Part 1 — How to Read Evolutionary Information? MSA vs PLM vs Hybrid
We examine the first major design decision: should a model rely on multiple sequence alignments, protein language models, or both?
Part of the series: The Technical Evolution of Protein AI — A Record of Key Design Decisions