12 Session 12: AlphaFold and Generative Models of Protein Structure
Sources for this page: Arne Elofsson’s AlphaFold lecture (AlphaFold 2026, undervisning) and diffusion-models lecture (slides/Diffusion models (undervisning).pdf), which this page follows; every paper named below is the primary source for its numbers, cited where it is used. Data and models are fetched live from the RCSB PDB, the AlphaFold Protein Structure Database (CC BY 4.0) and the ESM Atlas API. All figures were drawn or computed for this course, and every number from our own code was read back from the notebook’s output. The “four regions” heuristic in Section 12.4 is taken from the lecture; it has not yet been published as a paper.
12.1 Learning goals
By the end of this session you should be able to:
- explain why long-range information is essential for structure prediction, and how evolutionary covariation provided it;
- describe what AlphaFold2 represents internally (an MSA representation and a residue-pair representation that update each other, and a geometry-aware structure module) and why attention and pair representations suit proteins;
- interpret pLDDT and PAE correctly, and explain why confidence is not the same as correctness;
- explain why complexes are harder than single chains, and why a confident complex model does not show that two proteins interact in the cell;
- explain at a high level what a diffusion model does and what AlphaFold3 changed;
- say what current structure predictors cannot tell us.
In Session 6 we treated structure as the answer, and we learned to inspect and compare it. Now we ask how that answer can be predicted from sequence. This session closes the course, and nearly every earlier session contributes a piece to it:
| From | Idea used in modern structure prediction |
|---|---|
| Sessions 4-5 | alignments and profiles: the evolutionary record of a family |
| Session 6 | structure comparison (TM-score), distance and contact maps, conformational states |
| Session 7 | nonlinear models, and secondary structure from local sequence |
| Session 8 | honest evaluation on proteins unlike the training data |
| Session 9 | training deep networks |
| Session 10 | pairwise maps treated as images, residual networks, invariance |
| Session 11 | attention and protein language models |
The session starts from one question: what information is missing if we try to predict a residue’s structure from its local sequence alone? The answer is the contacts and interactions with residues far away in the sequence. The rest of the session is therefore about where this long-range information comes from, how AlphaFold2 learned to use it, and what we can trust in the result.
12.2 Where does long-range structural information come from?
12.2.1 The folding problem
For many proteins, the sequence determines the structure (Anfinsen, 1973, Science 181:223). A protein cannot, however, find that structure by trying every conformation. With even a few states per residue, a random search would take longer than the age of the universe, while real proteins fold in milliseconds to seconds (Levinthal’s paradox, 1969). Dill & MacCallum (Science 338:1042, 2012) summarised fifty years of work as three questions, the last of which is “can we devise a computer algorithm to predict protein structures from their sequences?”
12.2.2 Three generations of structure prediction
- Physics and templates (1970s onwards). The first approaches either simulated the atoms or borrowed from known structures. Molecular dynamics simulates the atoms, which is possible only for small, fast-folding proteins, even on special-purpose computers. Homology modelling copies the structure of a relative, and threading finds the known fold that a sequence fits best (Session 5’s profile idea applied to structures). Rosetta assembles the chain from short fragments of known proteins with a knowledge-based energy function. Blind tests held every two years since 1994, the CASP experiments (Moult et al. 1995, Proteins 23:ii), showed steady progress for proteins with templates, and little for the rest.
- Contact prediction (2011-2019). Coevolution in large protein families gave long-range contacts, then distances, cleaned up by deep networks (below).
- End-to-end learned geometry (2020). AlphaFold2 learned the whole path from alignments to atoms.
What all these approaches lacked was long-range information. Local sequence is not enough, because the same ten residues can be a strand in one protein and a helix in another (Session 7’s chameleon sequences).
12.2.3 Coevolution: long-range restraints from evolution
The second generation found this long-range information in evolution. If two residues touch, a mutation in one is often compensated by a mutation in the other, so across a large family of homologs the two residues co-evolve. The idea is old (Göbel et al. 1994, Proteins 18:309), and three developments made it work two decades later:
- Better statistics. Simple correlation between two alignment columns is swamped by phylogeny and by indirect effects (A-B and B-C contacts make A and C look coupled). Direct coupling analysis (DCA; Weigt et al. 2009, PNAS 106:67; Morcos et al. 2011, PNAS 108:E1293) solves this by fitting one global statistical model to the whole alignment. The model is the maximum-entropy model that reproduces the single-column and column-pair frequencies of the alignment (Session 1), and it separates direct couplings from indirect ones. The idea itself is a decade older and was long overlooked. Alan Lapedes and colleagues identified both problems, phylogeny and indirect couplings, in 1999 (Lapedes et al., IMS Lecture Notes 33:236). In a 2002 manuscript they fitted this same maximum-entropy model to protein alignments and showed that it predicts contacts better than simple correlation, and also the effects of combined mutations. The manuscript was not published in a journal and only appeared on arXiv in 2012 (Lapedes, Giraud & Jarzynski, arXiv:1207.2484), after DCA had made the idea famous. Because these pairwise couplings capture epistasis, the same kind of model also predicts mutation effects better than an independent profile (EVmutation; Hopf et al. 2017, Nat. Biotechnol. 35:128). These are the interactions Session 9 measured in GB1, now learned from an alignment.
- Bigger families. Sequence databases grew far faster than structure databases, so families became large enough for the statistics (hundreds to thousands of effective sequences).
- Machine learning on top. Contacts are not independent pixels. Helices give bands, and strand pairs give lines along or across the diagonal, so a network that looks at a whole neighbourhood of the map (Session 10’s convolution) can clean up the noisy coevolution signal. In the notebook, 3 × 3 windows centred on real contacts show only 146 of the 256 possible patterns. When the centre pair is in contact, a neighbouring pair is in contact 23-42% of the time, against a base rate of 3.0%. The first discussion problem asks you to read these patterns, which are exactly what a learned pair representation exploits.
The figure shows this progress on one protein, the bacterial response regulator CheY, using an alignment of 11,084 homologs. For each method, the 128 highest-scoring residue pairs (L = 128, the length of the protein) are plotted over the true contacts. Precision rises from 0.49 (mutual information) to 0.73 (DCA), 0.89 (a deep-learning contact predictor) and 0.99 (the contacts in the AlphaFold2 model):
The next step was from contacts to distances. AlphaFold 1 (Senior et al. 2020, Nature 577:706) and trRosetta (Yang et al. 2020, PNAS 117:1496) predicted distance distributions between residue pairs with deep residual networks, and they folded proteins by minimising against these distributions. The blind CASP assessments show how sudden the change was:
From CASP9 to CASP11 (2010-2014), contact precision stayed around 20-27%. It then reached 47% in CASP12 (2016) and more than 70% in CASP13 (2018). In CASP14, AlphaFold2 made contacts a by-product of predicting the whole structure directly.
The Stockholm group (this course’s teachers) contributed to this generation:
- PconsC2 (Skwark, Raimondi, Michel & Elofsson, 2014, PLoS Comput Biol 10:e1003889) used “a deep learning approach to identify protein-like contact patterns”, namely random forests applied iteratively over an 11 × 11 window of first-round predictions. It helped most for β-sheet proteins.
- PconsFold (Michel et al., 2014, Bioinformatics 30:i482) folded proteins from these contacts, improving model TM-scores by 33% over EVfold. PconsC3 (2017) predicted accurate contacts for “more than 4000” Pfam families of unknown structure.
- RaptorX-Contact (Wang et al. 2017, PLoS Comput Biol 13:e1005324, Session 10) treated the whole contact map as an image for a deep residual network. PconsC4 (Michel, Menéndez Hurtado & Elofsson, 2019, Bioinformatics 35:2677) made this fast and “hassle-free”, with no GPU needed.
- AlphaFold 1 (Senior et al. 2020, first shown at CASP13 in 2018; Nature 577:706) and trRosetta (Yang et al., 2020, PNAS 117:1496) predicted distance distributions (and, in trRosetta, orientations) instead of binary contacts. They folded by minimising against them. By CASP13 most single-domain proteins could be predicted “quite accurately”.
Historically, complexes were predicted by docking separately determined or modelled components, whereas modern models predict the chains and their interface jointly. The older route is summarised here for reference.
Docking asks how two proteins of known structure fit together. It followed its own path, along the same three routes as structure prediction.
Shape and physics. Katchalski-Katzir et al. (1992, PNAS 89:2195) put both proteins on a 3D grid. They scored shape complementarity for every relative translation at once with the fast Fourier transform (FFT), and they sampled the rotations exhaustively. This became the standard engine, and later methods refined it:
- GRAMM used low-resolution grids that tolerate errors in the structures (Vakser 1995, Protein Eng 8:371);
- ZDOCK added electrostatics and desolvation (Chen, Li & Weng 2003, Proteins 52:80).
All of them generate thousands of candidate poses, so the hard part is ranking them. This is the ranking problem again.
Data. HADDOCK drives docking with experimental information, such as mutations or NMR data, marking the interface residues (Dominguez, Boelens & Bonvin 2003, J Am Chem Soc 125:1731).
Templates. Complexes can also be modelled from a related complex of known structure. Templates “are available to model nearly all complexes of structurally characterized proteins” (Kundrotas, Zhu, Janin & Vakser 2012, PNAS 109:9438). Petras Kundrotas, the first author, teaches on this course.
CAPRI, the Critical Assessment of PRedicted Interactions, has been docking’s CASP since 2001 (Janin et al. 2003, Proteins 52:2). DockQ was developed to score its models (Basu & Wallner 2016, PLoS One 11:e0161879). Until 2018 the picture was clear. Targets with a template were predicted “very good to excellent”, but for 11 difficult targets without one, “medium and high quality models [were] submitted for only 3” (Lensink et al. 2019, CASP13-CAPRI, Proteins 87:1200).
Coevolution across an interface. Residues in contact between two proteins co-evolve too, provided we can pair up which sequence in one family interacts with which sequence in the other. This is easy for bacterial proteins encoded next to each other, but hard in general. Ovchinnikov, Kamisetty & Baker (2014, eLife 3:e02030) and Hopf et al. (2014, eLife 3:e03430) used such paired alignments to predict interfaces. Combined with folding, this idea led to fold and dock:
- Fold and dock. Joining the alignments of two proteins lets the folding machinery predict how they bind. Pozzati et al. (2022, Bioinformatics 38:954) did this with trRosetta, and their PconsDock score separated correct from incorrect models 98% of the time. With AlphaFold2 the same idea became FoldDock (Section 12.4), and folding and docking had become one problem.
12.2.3.1 Further reading
- Dill & MacCallum (2012), The protein-folding problem, 50 years on
- Lapedes, Giraud & Jarzynski (written 2002, posted 2012), Using sequence alignments to predict protein structure and stability with high accuracy
- Morcos et al. (2011), Direct-coupling analysis of residue coevolution captures native contacts across many protein families
- Skwark et al. (2014), Improved contact predictions using the recognition of protein like contact patterns
12.3 How AlphaFold2 combines the course’s ideas
AlphaFold2 (Jumper et al. 2021, Nature 596:583) was “the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known”. At CASP14, its median backbone accuracy was 0.96 Å, against 2.8 Å for the next best method. It is trained end to end to map sequence-derived, evolutionary and template information (the target sequence, an alignment of its homologs, and optionally structures of relatives) to atomic coordinates. This distinguishes it from single-sequence methods such as ESMFold, which we meet below.
12.3.1 The data flow
AlphaFold2 carries two representations through the whole network:
- the MSA representation, with a vector for every residue of every aligned sequence;
- the pair representation, with a vector for every pair of residues. It is a learned, much richer version of the contact and distance maps above (and of the distance map of β-globin in Session 6).
The Evoformer updates both representations, many times over. Three of its operations carry the main ideas:
- MSA attention. Within each aligned sequence, residues attend to each other (Session 11’s attention, biased by the current pair representation). Within each column, sequences attend to each other, so the network learns which homologs are informative for each position.
- MSA → pair. For every pair of columns, the way the two columns vary together across the family updates the pair representation. This is a learned version of coevolution.
- Triangle updates. Pair (i, j) is updated with information passing through every third residue k. Residue-pair relationships are not independent: if i is close to k and j is close to k, that constrains how i and j can be related. The authors were inspired by geometric constraints such as the triangle inequality on distances, but the updates are learned operations, not rules that enforce it.
The paper’s own figure shows how these operations fit together in one Evoformer block (panel a). Read it from left to right. The MSA representation (top row; s sequences × r residues × c channels) passes through row-wise attention, which is biased by the pair representation (the arrow from below), column-wise attention and a transition, a small feed-forward network applied to each vector. The outer product mean then sends what the alignment shows down to the pair representation (bottom row). This passes through the four triangle operations and its own transition. Every step adds its result to its input (the ⊕ symbols), the residual connections of Session 11’s Transformer. The block is repeated 48 times, each with its own weights. Panels b and c show the pair representation as the edges of a graph, and which edges each triangle update combines to update the edge ij:
The structure module then builds coordinates. Each residue is a rigid frame (the position and orientation of its backbone). Invariant point attention reasons about these frames in 3D in a way that does not depend on how the whole protein happens to be rotated or translated. This is the invariance idea of Session 10, now in three dimensions. Finally, recycling runs the whole network again with its own outputs as extra inputs, which refines the prediction a few times.
In short, AlphaFold2 kept the ingredients of the second generation (alignments, coevolution, contacts, distances, physical plausibility) and turned each into a learnable operation inside one network, trained on the final coordinates.
The numbers below are from the paper and its Supplementary Information.
1. Inputs. Three kinds of information go in.
- An MSA. Genetic databases are searched for homologs. At inference the network reads a representation of 512 rows: cluster centres of the alignment, plus template information added as extra rows. Up to 5,120 further sequences enter through a lighter extra-MSA stack (4 blocks, with a cheaper “global” column attention).
- Templates. Up to 4 structures of related proteins, found in the PDB. Only two of the five released models use them, and the network works well without them.
- The sequence itself, including the relative position of every pair of residues along the chain.
These are embedded into two representations that are carried through the whole network:
- the MSA representation: one 256-number vector for every residue in every sequence (sequences × residues × 256);
- the pair representation: one 128-number vector for every pair of residues (residues × residues × 128). This is a learned, much richer version of Section 12.2’s contact and distance maps.
2. The Evoformer: 48 blocks. Each block updates both representations and lets them talk to each other. In order (figure, panel b):
- Row-wise gated self-attention with pair bias. Within each aligned sequence, every residue attends to every other residue (Session 11’s attention). The attention weights get a bias from the pair representation: what the network currently believes about residues i and j steers how much they attend to each other.
- Column-wise gated self-attention. Within each column, every sequence attends to every other sequence. The network learns which homologs are informative for this position.
- MSA transition. A small two-layer network applied to every vector.
- Outer product mean. For every pair of columns (i, j), the outer product of their vectors is averaged over all sequences and added to pair (i, j). This is the learned version of coevolution: how two columns vary together across the family.
- Triangle multiplicative updates, “outgoing” and “incoming” edges. Pair (i, j) is updated from the pairs (i, k) and (j, k) for every third residue k. In the authors’ words, “many constraints must be satisfied including the triangle inequality on distances. On the basis of this intuition, we arrange the update operations on the pair representation in terms of triangles”.
- Triangle self-attention, around the starting node and around the ending node. The same triangle idea with attention: pair (i, j) attends to all pairs (i, k), with a bias from pair (j, k).
- Pair transition. A two-layer network on every pair vector.
Every step is added to its input (residual connections, Session 10’s ResNet idea) and uses layer normalisation. Dropout is applied during training. The MSA Transformer (Rao et al. 2021) used the same row-and-column attention at about the same time. The idea of learning directly from raw alignments had been explored in rawMSA (Mirabello & Wallner, 2019, PLoS One 14:e0220182).
3. The structure module: 8 layers. It turns the representations into coordinates. Its input is the pair representation, plus a single representation: the first row of the MSA representation (the target sequence), projected to 384 numbers per residue.
- Each residue is a rigid frame: the position and orientation of its backbone N, Cα and C atoms, as a rotation and a translation. All frames start at the origin with the same orientation, which the authors call “black hole initialisation”.
- Invariant point attention (IPA) updates the single representation. It is attention that combines three terms: the residue vectors, the pair representation, and 3D points attached to each residue’s current frame. Because distances between points do not change when the whole protein is rotated or moved, the result is invariant to global rotations and translations. The network cannot cheat by learning where proteins usually sit in space.
- After each layer, every frame is updated by a small rotation and translation. The 8 layers share their weights, like 8 steps of one refinement.
- Finally, 7 torsion angles per residue (backbone ω, φ, ψ and side-chain χ1-χ4) are predicted and turned into all-atom coordinates.
4. Recycling. The whole network is run again with its own outputs fed back as extra inputs: the predicted structure, the pair representation and the first MSA row. At inference it runs 4 times, that is, 3 recycles, with the same weights. Each pass refines the last.
5. Training.
- The main loss is FAPE, frame aligned point error. Every atom is compared with the true structure in the local frame of every residue, and errors are clamped at 10 Å. Unlike a global RMSD, this penalises wrong local geometry everywhere, and it cannot be satisfied by a mirror image.
- Auxiliary losses keep the intermediate representations meaningful:
- a distogram head predicts binned residue-residue distances from the pair representation (64 bins, 2-22 Å): AlphaFold 1’s output, now just a side task;
- a masked-MSA head predicts residues hidden in the alignment (15%), like Session 11’s BERT;
- a pLDDT head learns to predict the model’s own local accuracy (50 bins);
- fine-tuning adds penalties for impossible geometry, such as clashes and bad bond lengths.
- Data. PDB structures released before 30 April 2018. The network was then used to predict about 350,000 sequences from Uniclust30, and these predictions were added to the training data (self-distillation).
- Scale. Training used crops of 256 residues, then 384 in fine-tuning, on 128 TPU v3 cores: about a week of initial training plus about four days of fine-tuning in the simplified protocol described in the paper.
- PAE was added by a further short fine-tuning, to the “pTM” models, with a head that predicts the aligned error between every pair of residues (64 bins, 0-31.5 Å).
6. Five models. Five networks were trained with different random seeds, two with templates and three without. All five are run, and the predictions are ranked by pLDDT.
12.3.1.1 Further watching
Highly accurate protein structure prediction with AlphaFold (Simon Kohl, DeepMind), a detailed walk through AlphaFold2’s architecture:
12.3.2 Confidence: pLDDT and PAE
A predicted structure is only useful if we know how far to trust it. AlphaFold2’s confidence measures are therefore as important as the structure itself, and they answer two different questions:
- pLDDT (0-100, per residue) asks how trustworthy is the local structure around this residue? Above 90 is very high, 70-90 confident, and below 50 very low.
- PAE (predicted aligned error, per pair of residues) asks if I align the model on residue (or region) A, how certain is the position of B relative to A? Low PAE within a domain and high PAE between two domains means that each domain may be right while their relative orientation is unknown. This matters for multi-domain proteins, for flexible regions and, below, for complexes.
Low pLDDT means low confidence, not necessarily disorder. Intrinsic disorder is one common reason, but other reasons are too little evolutionary information, several conformational states, an uncertain domain orientation, or simply the limits of the model. Moreover, confidence is not correctness, because a confidence score is the model’s own estimate, learned from its training data. The second discussion problem tests how well pLDDT tracks the real error. Checking this on famous proteins is harder than it looks, because well-studied proteins are likely to have close relatives in the training data (Session 8).
In the notebook. The AlphaFold DB model of human β-globin (mean pLDDT 97.2) matches the crystal structure 4HHB with TM-score 0.988 and RMSD 0.51 Å. For p53, 30% of the residues have pLDDT below 50. The DNA-binding domain averages 95.3, while the N-terminal transactivation region (47.6) and the C-terminus (45.8) are not confident. For p53 that low confidence matches biology, since these regions are known from experiments to be intrinsically disordered:
For fun: in Protein Fighter, a game from Sergey Ovchinnikov’s lab, the health bar is the fighter’s pLDDT and the small map under it is its PAE.
Scale and tools. AlphaFold predicted almost the entire human proteome (Tunyasuvunakool et al. 2021, Nature 596:590), and the AlphaFold Database made predicted structures available for essentially every protein in UniProt (Varadi et al. 2024, NAR 52:D368). ColabFold (Mirdita et al. 2022, Nat Methods 19:679) replaced the slow alignment search with MMseqs2 and put AlphaFold in a Colab notebook. ESMFold predicts from a single sequence, using the ESM-2 language model (Session 11) instead of an alignment. In the notebook it predicts Ephrin-A5 in seconds, with TM-score 0.941 against the crystal structure (1SHW). In 2024, Demis Hassabis and John Jumper shared half of the Nobel Prize in Chemistry “for protein structure prediction”, and the other half went to David Baker “for computational protein design”.
Known limits (lecture). AlphaFold cannot tell whether a sequence folds at all, or how a mutation changes the structure. It may or may not find alternative conformational states. The arrangement of domains in large flexible or repeated proteins is often arbitrary. It “does not know what a membrane is”. In a protein-DNA complex it can place another protein where the DNA should be.
12.4 From single chains to interactions
12.4.1 Complexes are harder
AlphaFold2 was trained on single chains, yet it can predict complexes when the chains are given together (FoldDock; Bryant, Pozzati & Elofsson 2022, Nat Commun 13:1265). Later, AlphaFold-Multimer (Evans et al. 2021, bioRxiv) was retrained on complexes. Complexes are harder than single chains for four reasons:
- Paired evolutionary information helps interfaces. Residues in contact across an interface co-evolve too, if the alignments of the two families can be paired (which sequence interacts with which).
- Interfaces need their own confidence. AlphaFold-Multimer introduced ipTM, a confidence score for the interfaces, and FoldDock’s pDockQ combines the interface pLDDT with the number of interface contacts. A model is compared with the true complex using DockQ, which runs from 0 to 1, with at least 0.23 counting as acceptable.
- Interface confidence is harder than chain confidence. Each chain can be right while their relative placement is wrong, which is exactly what the PAE between chains shows.
- A plausible complex is not evidence of interaction. A structure predictor can suggest how two supplied molecules could interact. It does not establish whether they interact in the cell, or when. That is one of the most important messages of this session.
Predicted complexes can still lead to biology when they are combined with other evidence. All-against-all modelling of 14 fertilisation proteins pointed to “a pentameric complex involving sperm IZUMO1, SPACA6, TMEM81 and egg JUNO, CD9” at the egg-sperm fusion synapse (Elofsson et al. 2024, eLife 13:RP93131).
12.4.2 The ranking problem
Generating a good model has become easier than knowing which of your models is good. In short, sampling quality is not ranking quality, as two examples show:
- Antibody-antigen complexes have no coevolution between the partners. On 110 complexes unseen in training, sampling 200 models per complex raised the average DockQ of the best model “from below 0.3 to above 0.5”. The models’ own confidence, however, often “cannot identify the best predicted structures”, so the top-ranked model is much worse than the best one (Fromm, Ludaic & Elofsson 2026, Bioinformatics 42:btag136).
- Confidence depends on what you include. Trimming a full-length construct to its interacting domains raises ipTM “even though the structure prediction of the interaction is unchanged”. ipSAE counts only residue pairs with good PAE, and it separates true from false complexes better (Dunbrack 2025, bioRxiv).
The common thread is Session 8’s lesson on a grand scale. Performance on examples similar to the training data says little about genuinely new ones, and confidence scores do not always know the difference.
12.4.3 When can a complex prediction be trusted? A practical heuristic
How, then, can we decide when to trust a complex prediction? The lecture sorts protein interactions into four regions. This is a heuristic for thinking about the question, not an established classification:
| Region | Character | Example | Current methods |
|---|---|---|---|
| I, tractable | Stable interfaces, strong coevolution | Obligate heterodimers (Ephrin-A5:EphB2) | Work well |
| II, ambiguity and competition | A plausible interface is easy; the right partner or stoichiometry is hard | Proteasome paralogs; Izumo1 and its competing partners | Struggle to rank the correct partner |
| III, missing coevolution | No joint selective pressure | Antibody-antigen, host-pathogen | AlphaFold3-class models help, partly through memorisation |
| IV, missing context or dynamics | Transient, conditional, disordered | p53 transactivation domain binding MDM2 | Static structure prediction is the wrong tool |
The lab tests this heuristic. Each group makes one AlphaFold3 Server prediction per region and checks whether the model’s confidence tracks how hard each problem really is.
What the scores do not tell you. pLDDT is the confidence in the local structure, PAE the confidence in the relative position of two parts, and ipTM or ipSAE the confidence in an interface. All of them are the model’s own estimates for the chains it was given. None of them shows that the proteins actually bind, that they bind each other rather than a competitor that was left out, or that the predicted complex is the one formed in the cell. A confident interface is a candidate hypothesis, to be tested with other evidence.
12.4.3.1 Further reading
- Jumper et al. (2021), Highly accurate protein structure prediction with AlphaFold
- Bryant, Pozzati & Elofsson (2022), Improved prediction of protein-protein interactions using AlphaFold2
- Fromm, Ludaic & Elofsson (2026), Evaluating deep learning based structure prediction methods on antibody-antigen complexes
12.5 Generative structure modelling
12.5.1 Diffusion in one page
The newest structure predictors are generative models, and most of them are built on diffusion. A diffusion model learns to generate data by learning to remove noise, in three steps:
- Forward process. Real data are noised by adding a little Gaussian noise step by step, until only noise is left. This step needs no learning.
- Training. A network is shown a noisy example and its step number, and it is trained to predict the noise that was added.
- Generation. Starting from pure noise, the network is applied many times, and each step removes a little noise.
The idea goes back to Sohl-Dickstein et al. (2015). Ho, Jain & Abbeel (arXiv:2006.11239, 2020) made it practical, and it is behind image generators such as DALL-E. Two relatives deserve one sentence each. Flow matching learns a smooth path from noise to data and needs fewer steps. A variational autoencoder generates data by decoding samples from a learned low-dimensional space. The third discussion problem asks whether a diffusion model learns a distribution or memorises its training examples.
12.5.2 AlphaFold3: what changed
AlphaFold3 (Abramson et al. 2024, Nature 630:493) brings diffusion into structure prediction. It predicts “the joint structure of complexes including proteins, nucleic acids, small molecules, ions and modified residues”. Conceptually, four things changed:
- Broader chemistry. Its published model was designed to represent proteins, nucleic acids, ligands, ions and modified residues jointly, with residues, nucleotides and single ligand atoms all as tokens.
- A pair-focused trunk. The Pairformer replaces the Evoformer. It works mostly on the pair representation and much less on the alignment.
- Generative coordinates. A diffusion module replaces the structure module. It starts from random atom positions and denoises them, conditioned on the trunk.
- All atoms at once. The model works on all atoms at once, which is what lets it handle any molecule made of atoms.
Diffusion brings sampling into structure prediction, and this has two consequences. First, many seeds can be tried, which brings back the ranking problem above. Second, the model can hallucinate plausible-looking structure in regions that should be disordered, which the authors had to counter in training. At CASP16, “AlphaFold3 performs slightly better for protein complexes. However, when massive sampling is applied to AlphaFold2, the difference disappears” (Elofsson 2026, Proteins 94:154).
12.5.2.1 Where the field is in 2026 (optional, and will age quickly)
AlphaFold3 was released without training code, so open reproductions followed quickly:
Boltz-1: fully open (MIT licence).
Boltz-2: adds binding-affinity prediction, approaching free-energy perturbation while being “at least 1000× more computationally efficient”.
Chai-1: can run without alignments.
Protenix.
OpenFold3: preview release.
RoseTTAFold All-Atom (Krishna et al. 2024, Science 384:eadl2528) is an independent architecture.
RNA. Ten methods were benchmarked on RNAs and RNA complexes not seen in training. AlphaFold3 and Boltz-1 lead, but “accuracy depends greatly on similarity to the training set”, and the methods “recognize recurring motifs rather than generalize to novel RNA folds” (Ludaic & Elofsson 2026, NAR 54:gkag813).
The field moves fast, and three systems from 2026 show three different ways of moving beyond AlphaFold3.
ESMFold2 (Candido et al. 2026, bioRxiv; Biohub; MIT licence, open weights). It joins a large protein language model, ESMC (the successor to Session 11’s ESM-2), to an AlphaFold3-style diffusion structure module, and the whole model has about 6.6 billion parameters. The original ESMFold was built for single chains, whereas ESMFold2 predicts complexes of proteins, nucleic acids and ligands, from single sequences or with an optional alignment. The authors report that it “exceeds the performance of established methods for biomolecular complex prediction across benchmarks, including for the interactions of antibodies with their targets” (Candido et al. 2026). The notebook runs it on one complex from each of the four regions above, with five seeds each (figure). For every complex, the seed with the highest ipTM is a good model (DockQ 0.61-0.85). For the 1996 D1.3 antibody, however, three of five seeds put the antibody in the wrong place. Moreover, no single ipTM threshold separates right from wrong across complexes, since correct Fab : MHC-I models (ipTM 0.73) score below wrong D1.3 models (0.78).
Figure 12.8: ESMFold2 on one complex per region, five seeds each, scored with DockQ against the crystal structure (0.23 acceptable, 0.49 medium, 0.80 high). Left: all seeds; the bar marks the seed with the highest ipTM. Right: ipTM against DockQ. These are well-known complexes; p53 : MDM2 has been in the PDB since 1996, so a good model here does not show that Region IV is solved. IsoDDE, the Isomorphic Labs Drug Design Engine (Isomorphic Labs 2026, technical report), is proprietary and comes from DeepMind’s drug-discovery sister company. The company claims:
- on the Runs N’ Poses protein-ligand benchmark, more than double AlphaFold3’s accuracy for the ligands least similar to the training data;
- on 334 new antibody-antigen complexes, 2.3 times AlphaFold3 and 19.8 times Boltz-2 at high accuracy (DockQ > 0.8);
- binding-affinity predictions that can beat physics-based free-energy methods;
- binding pockets found from sequence alone.
There is no code, no model and no independent test, so these are claims, not results. Notice where the claimed gains are: on the examples unlike the training set, where AlphaFold3 is weakest.
OpenDDE (company version) (2026, GitHub, Apache licence), developed by Aureka Research, is an open-source, all-atom biomolecular foundation model. It turns co-folding into a scalable engine for structure prediction, design and optimisation in drug discovery. Its developers report higher performance on antibody-antigen complexes.
OpenDDE (single-developer version) (2026, GitHub, MIT licence) is a community project by a single developer, not affiliated with Isomorphic Labs. It is not a new model. Instead, it chains existing open tools behind one web interface:
- AlphaFold3 for complexes;
- P2Rank for pockets;
- Boltz-2 for affinity;
- ImmuneBuilder for antibodies;
- ChEMBL and RDKit for known ligands;
- a language-model assistant.
It has no benchmarks of its own. What it shows is what an “engine” is: a pipeline of separate models, each of which can, and should, be evaluated on its own.
The contrast between open and closed systems matters for science. Open tools such as ESMFold2, Boltz-2 and OpenFold3 can be tested by anyone, on data they have never seen. A closed system, in contrast, can only be judged by what its owners choose to report.
12.5.3 Designing new proteins
The same generative machinery can also run in reverse, from a desired structure to a new sequence. A typical design pipeline has four steps:
- RFdiffusion (Watson et al. 2023, Nature 620:1089) generates a new backbone, for example one that binds a chosen target surface;
- ProteinMPNN (Dauparas et al. 2022, Science 378:49) designs a sequence for that backbone;
- AlphaFold2 or ESMFold refolds the designed sequence, to check that it folds into the intended structure;
- experiments test whether it does, and whether it works. Refolding is only a computational consistency check, not experimental validation. Hundreds of RFdiffusion designs were characterised in the lab, including an influenza-haemagglutinin binder whose cryo-EM structure was “nearly identical to the design model”.
12.5.3.1 Further reading
- Abramson et al. (2024), Accurate structure prediction of biomolecular interactions with AlphaFold 3
- Watson et al. (2023), De novo design of protein structure and function with RFdiffusion
- Elofsson & Ben-Tal (2026), What will be the future of computational biology for macromolecules in the era of AI?
12.5.3.2 Further watching
Flow-Matching vs Diffusion Models explained side by side (AI Coffee Break with Letitia), used in the diffusion lecture:
12.6 The course in one line
Modern structure prediction combines alignment and evolution with neural networks, attention, pair representations and geometry. The scientific warning that goes with it is that a model is a hypothesis with confidence estimates, not an experiment.
12.7 Check your understanding
12.8 The course notebook
notebooks/session12-alphafold-and-generative-models-in-code.ipynb computes the numbers on this page:
- Contact prediction for CheY by mutual information and DCA, computed from an alignment, and compared with ESM-2 and AlphaFold2; 3 × 3 contact patterns in eight real structures.
- AlphaFold DB models of β-globin and p53 with pLDDT and PAE; the β-globin model against its crystal structure; ESMFold on Ephrin-A5; ESMFold2 models of five complexes scored with DockQ.
- The forward noising of a diffusion model, and a small diffusion model generating points from noise.
It runs on a CPU in a few minutes. Open the notebook in Google Colab.
12.9 The lab
Full lab: the lab is adapted from the previous course’s AlphaFold2 lab and the 2026 AlphaFold3 Server exercise. Your group runs one of four targets on the AlphaFold Server, one for each region of the four-regions heuristic on this page:
- Region I: Ephrin-A5 : EphB2 (1SHW);
- Region II: Izumo1 : Juno (5F4E);
- Region III: antibody : antigen (8TQ7);
- Region IV: p53 peptide : MDM2 (1YCR).
Work through notebooks/session12-lab.ipynb, replacing every todo(...) call with your own code:
- fetch your target’s chain sequences from the PDB and submit them to the AlphaFold Server (20 jobs per day and account);
- read AlphaFold DB models: pLDDT, the PAE matrix, which models are disordered, and how closely a model matches its crystal structure;
- load your AlphaFold Server result, plot its PAE heatmap, and compute pTM, ipTM and the interface score ipSAE;
- pool the class’s results and compare ipTM with ipSAE across the four regions;
- superimpose your model on the crystal structure in PyMOL.
12.9.1 Submitting this lab
This lab is administered through Canvas, not through this book. See the course’s Canvas page for the current quiz, the deadline and how to submit.
12.10 What’s next
What remains is the capstone project. It returns to predicting secondary structure, the problem this course met in Session 7, now with every tool from Sessions 5-12 at hand.