Even More Protein Structure Prediction

10/21/2026

print view

%%html
<script src="https://bits.csb.pitt.edu/preamble.js"></script>

AF2: from sequence information to structure

Jumper et al. (2021)

AlphaFold Key Takeaways

• Evolution provides structural information. Templates provide known geometry; coevolving residues suggest interactions.

• AlphaFold2 integrates sequence, evolutionary, and geometric information. The Evoformer updates MSA and pair features; the structure module updates residue frames.

• Geometry matters in both the architecture and the loss. IPA uses spatial relationships; FAPE compares positions in residue-local frames.

• Assess local accuracy and global arrangement separately. pLDDT estimates local accuracy; PAE assesses confidence in relative placement; pTM is overall structure.

• Structure prediction is not a simulation of protein folding. A predicted structure does not reveal folding pathways or conformational populations.

How Important is the MSA?

Neff = normalized nonredundant sequences

Coverage = how much of the target does this template provide structural information for?

Single sequence fails

Image credit: S. Ovchinnikov

But a template can rescue

Image credit: S. Ovchinnikov

AF2Rank

Providing a single sequence and candidate structure let's us repurpose AF2 as a structure evaluator instead of generator.

Image credit: S. Ovchinnikov

AF2Rank

AlphaFold's confidence in its prediction correlates with the quality of the input template (single sequence case). This ranking of structures outperforms the previous state-of-the-art scoring.

Roney & Ovchinnikov (2022)

A proposed role for the MSA

The MSA may guide search; learned structural preferences help refine candidates.

Roney & Ovchinnikov (2022)

AF2 predicts monomer folds well

What about complexes?

AlphaFold Multimer

• AF2 architecture retained: Evoformer, structure module, and recycling; all chains are predicted jointly from their sequences and specified stoichiometry.

• Trained on complexes: Uses multichain biological assemblies, with training crops that include interacting regions from different chains.

• Explicit chain identity: Distinguishes residues within a chain, across chains, and in identical copies of a protein.

• Losses adapted to interfaces: Uses a larger interchain FAPE cutoff (30 Å versus 10 Å within chains), along with adjustments to clash and chain-placement penalties.

• Paired and unpaired MSAs: Combines each chain’s evolutionary information with homologs paired across species to reveal potential interchain coevolution.

• Sample the MSA. Data augmentation and balancing. Chain with many homologs dos not dominate.

• Interface confidence: Introduces ipTM to assess interchain arrangement.

Evans et al. (2022). Protein complex prediction with AlphaFold-Multimer. bioRxiv.

Alphafold Multimer

Pair homologs across chains

Paired rows can expose cross-chain coevolution. Paralogs complicate pairing.

Evans et al. (2022)

Read the inter-chain PAE blocks

Good individual chains can still have an uncertain relative placement.

Synthetic teaching example. Darker green = lower PAE.

Jumper et al. (2021) · Evans et al. (2022)

ipTM

ipTM (interface predicted TM-score) measures confidence in the relative placement of chains. Basically cross-chain pTM.

The problem: a confident interface can have low ipTM when long disordered tails or noninteracting domains have uncertain positions.

ipSAE (interaction prediction Score from Aligned Errors) focuses the score on confidently positioned regions.

Dunbrack, R. L., Jr. (2025). Rēs ipSAE loquuntur: What’s wrong with AlphaFold’s ipTM score and how to fix it. bioRxiv.

Conformational Sampling with AlphaFold

Change the evolutionary evidence supplied to AF2 while keeping its trained weights fixed.

Approach	What changes?	Example
Random subsampling	Use shallow MSAs containing different random subsets of homologs	del Alamo et al.
Sequence clustering	Group similar homologs and predict separately with each cluster’s MSA	AF-Cluster
Targeted MSA edits	Replace selected alignment columns with alanine, targeting contacts identified in an initial structure	SPEACH_AF

Different subsets or edits can shift predictions toward alternative conformations.

Greater diversity can also include misfolded structures. Prediction frequencies are not automatically equilibrium populations.

del Alamo et al. (2022). Sampling alternative conformational states of transporters and receptors with AlphaFold2. eLife.
Wayment-Steele et al. (2024). Predicting multiple conformations via sequence clustering and AlphaFold2. Nature.
Stein & Mchaourab (2022). SPEACH_AF: Sampling protein ensembles and conformational heterogeneity with Alphafold2. PLOS Computational Biology.

Shallow MSAs can reveal other states

LAT1: compare predictions with inward- and outward-facing structures.

del Alamo et al. (2022)

Change rows or mask columns

Rows: which homologs?  Columns: which positions?

Kalakoti & Wallner (2025)

AF-Cluster: group homologs first

Predict the same query using different sequence clusters.

Wayment-Steele et al. (2024)

AF-Cluster

Wayment-Steele et al. (2024)

A fold-switching success: RfaH

The C-terminal region changes between α-helical and β-sheet folds.

Wayment-Steele et al. (2024)

Unseen fold switches remain difficult

A separate held-out set: only 1 of 7 switches recovered.

Sa1 predictions recover the α/β fold and miss the three-helix bundle.

Chakravarty et al. (2024). AlphaFold predictions of fold-switched conformations are driven by structure memorization. Nature Communications 15, 7296.

Case study: prospective ligand discovery

Question: can unrefined AF2 receptor models support discovery of new ligands?

Lyu et al. docked the same large libraries against AF2 and experimental receptor structures, then tested selected compounds.

Target	AF2 hits / tested	Experimental-structure hits / tested
σ₂ receptor	64/119 (54%)	70/138 (51%)
5-HT₂A receptor	42/161 (26%)	51/223 (23%)

Hits: >50% radioligand displacement at 1 μM (σ₂) or 10 μM (5-HT₂A). Rates concern selected compounds, not the full library.

Lyu et al. (2024)

A discovered ligand gets a structural test

Cryo-EM of the new 5-HT₂A agonist Z7757 supports the predicted binding pose.

Lyu et al. (2024)

AlphaFold3: Going beyond proteins

AF3: proteins + nucleic acids + ligands + ions + modifications.

Abramson et al. (2024)

AF3 - a focus on the pair representation

Molecular inputs → learned representations → coordinate diffusion.

Abramson et al. (2024)

Trace input features, MSA/template processing, pairformer, diffusion, confidence. The trunk is recycled; a fixed trunk output can condition repeated diffusion samples. Distinguish changing the whole model seed from changing only diffusion noise.

AF2 → AF3: what changes?

	AF2	AF3
Representation trunk	Evoformer: MSA + pairs	Short MSA processing + Pairformer
Coordinate output	Residue frames and torsions	Atomic coordinates
Structure generation	Structure-module refinement	Repeated denoising

AF3 retains evolutionary input and triangle-based pair reasoning.

Abramson et al. (2024)

The Pairformer retains pair reasoning

Abramson et al. (2024)

AF3’s coordinate denoiser

Atom/token attention; random rotations and translations; no hard-coded rotational equivariance.

Abramson et al. (2024)

Diffusion training: corrupt coordinates

$$\mathbf{x}_{\sigma}=\mathbf{x}_0+\sigma\boldsymbol\epsilon,\qquad\boldsymbol\epsilon\sim\mathcal N(0,I)$$

The input includes the noise level and molecular context.

Abramson et al. (2024) · Karras et al. (2022)

Diffusion: learn to denoise, then generate

The idea: learn to recover clean data from noisy examples, then use that to generate new samples.

Training: the correct structure is known

$$ \underbrace{\mathbf{x}_{\sigma}=\mathbf{x}_0+\sigma\boldsymbol{\epsilon}}_{\text{add Gaussian noise}} \quad\xrightarrow{\;D_\theta(\mathbf{x}_{\sigma},\,\sigma,\,c)\;} \underbrace{\widehat{\mathbf{x}}_0}_{\text{predict clean coordinates}} $$

• Train at many noise levels $\sigma$, comparing predictions with the known coordinates $\mathbf{x}_0$.
• The denoiser $D_\theta(\mathbf{x}_{\sigma},\,\sigma,\,c)$ receives both the noise level and conditioning information $c$.

Generation: start from random coordinates

1. Predict clean coordinates using the trained denoiser.
2. Take a step toward that prediction and reduce the noise level.
3. Repeat: large-scale arrangement → local atomic geometry.

In AF3: $\mathbf{x}$ contains atom coordinates; $c$ includes molecular identities and Pairformer features. These guide every denoising step.

Different random starts can produce different candidate structures.

Abramson et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500.

Diffusion does not guarantee state coverage

AF3 predicts closed conformations in both apo and holo conditions. Grey is experimental structure.

Abramson et al. (2024)

AF3 improves(?) local monomer accuracy

Below the diagonal favors AF3. Global-fold gains are less consistent.

Peng et al. (2025)

What if homolog information is missing?

AF3 improves forced single-sequence prediction, but both models struggle.

Peng et al. (2025)

ESM: a language model for proteins

ESM = Evolutionary Scale Modeling. ESM-2 learns to predict masked amino acids.

• Amino acids are tokens; self-attention uses context across the sequence.
• Train on many unaligned sequences, using the original residues as labels.
• Trained on 65 million unique protein sequences (ESM2)

Lin et al. (2023)

ESM learns more than residue probabilities

Residue embeddings: learned vectors that depend on sequence context.

Structural signal: contact information can be read out from attention patterns.

The language model was pretrained using sequence labels.

Contacts predicted from attention weights of ESM model.

Lin et al. (2023)

ESMFold: turn representations into structure

ESM-2 representations → supervised folding head → atomic coordinates.

The folding head learns from experimental and AF2-predicted structures.

Inference uses the target sequence without an MSA search.

Lin et al. (2023)

ESM Folding Block

Performance

Mahtha, Venkadesan & Mohanty (2026). Comparative evaluation of the prediction accuracy of AlphaFold and ESMFold for monomeric and dimeric proteins. NAR Genomics and Bioinformatics 8(1), lqag002.

RoseTTAFold: couple three tracks

Exchange sequence, pair, and 3D information throughout the network.

Baek et al. (2021)

RoseTTAFold All-Atom: broaden the chemistry

Add atom–bond graphs and chirality to model mixed molecular assemblies.

Krishna et al. (2024)

One timeline: expanding the prediction task

Milestone	What changed?
2020–21: AF2, RoseTTAFold	Accurate monomer structures; joint geometric reasoning
2021: AF-Multimer	Train explicitly on complexes
2022–23: ESMFold	Language-model features support single-sequence prediction
2024: AF3, RFAA, Chai-1, Boltz-1	Broader molecular composition
2025: Boltz-2	Structure prediction plus an affinity component

Structure Prediction $\ne$ Protein Folding

Many proteins fold in vitro ... but it can get complicated

Molecular chaperones are a group of unrelated protein families whose roles include preventing protein aggregation, assisting in correct folding and assembly of proteins, or disassembling complexes or aggregates.

Chaperone Example: Heat Shock Protein 70

Chaperone Example: Heat Shock Protein 70

v= py3Dmol.view('1dkx',style='cartoon:color~magenta'); v.setStyle({'chain':'B'},{'stick':{'colorscheme':'whiteCarbon'}}); v.setStyle({'chain':'A','resi':'1-507'},{'cartoon':{'color':'lightgreen'}});v.zoomTo({'chain':'B'}); v.show()

3Dmol.js failed to load for some reason. Please check your browser console for error messages.

Chaperones as Drug Targets

Geldanamycin HSP90 Inhibitor

• Extremely challenging

• Cancer: increased mutation burden and protein expression in cancer cells mean they are more dependent on chaperones for proper protein folding

  • Targeting heat shock proteins induces degradation of mutated proteins
  • Inhibitors commonly bind to ATP binding pocket
  • Toxicity issues

• Neurodegenerative Diseases: Alzheimer’s, Parkinson’s, and Huntington’s disease all involve misfolded protein aggregates.

  • Upregulating chaperones may reduce aggregation (Arimoclomol - Niemann–Pick disease)

Takeaways

• Multimer: joint prediction needs interface-specific assessment.
• Sampling: recovering states is different from predicting populations.
• AF3: molecular conditioning + atomic denoising expands the scope.
• ESMFold: sequence pretraining changes what is needed at inference.

Choose the model and validation for the biological question.