Biomolecular structure prediction (AlphaFold-style)
Deep networks that predict 3D structures of proteins, nucleic acids and their complexes from sequence, trained on the PDB and coevolution in multiple sequence alignments.
Protein design, docking and co-folding, variant-effect prediction and most structure-guided drug discovery start from predicted or experimentally derived structures and inherit the accuracy and blind spots of these models.
Computer programs that guess a molecule's 3D shape from its sequence of building blocks.
- 1Sequence plus relatives
- 2Spot correlated positions
- 3Network guesses distances and angles
- 4Diffusion refines all atoms
- 5Confidence flags weak regions
The task. Given amino-acid sequences (and optionally RNA, DNA, ions and drug-like molecules), output 3D coordinates of all heavy atoms, plus a confidence for each part. Accuracy is scored by comparing against experimental structures solved later, using measures such as lDDT (local distance agreement) and TM-score (global fold similarity), and DockQ for the interface between chains.
Why it works. Evolution constrains proteins: residues that touch must remain compatible, so mutations in one are compensated in the other. A multiple sequence alignment of related proteins therefore contains a statistical record of contacts. AlphaFold2 feeds the alignment and a table of residue-pair features into a transformer called the Evoformer, which repeatedly updates both with attention, including a "triangle" rule that keeps distances geometrically consistent. A second module turns these features into rotations and positions of each residue, trained to match known structures.
AlphaFold 3 and successors. AlphaFold 3 predicts the whole complex, including DNA, RNA and small molecules, by using a diffusion model: it starts from noisy atom positions and denoises them into a structure. Open and company models followed (Boltz, Chai, OpenFold3 and others).
What goes wrong. (1) The model learned from the Protein Data Bank, so it tends to reproduce shapes it has seen; for ligands, RNA and unusual folds this can mean memorising rather than understanding. (2) A structure is a single snapshot; real proteins are ensembles. (3) Predicting what a complex looks like is easier than predicting whether two proteins bind. (4) Confidence scores can be miscalibrated.
Where the open problems are. DARC lists problems on fold switching, RNA tertiary structure, proteome-wide interaction prediction and how well co-folding models generalise to new drug-like chemistry.
Skills. Protein chemistry, attention and diffusion models, evaluation without leakage (time and similarity splits), reading CASP assessments.
Go deeper
The full technical text.
What it is
Given the amino-acid sequence of a protein (or sequences of several chains, nucleic acids, ligands and ions), predict the 3D coordinates of every heavy atom. The modern recipe is a neural network trained on the Protein Data Bank (PDB, roughly 200,000 entries) that exploits coevolution: residues in contact mutate in a correlated way across homologous sequences, so a multiple sequence alignment (MSA) carries distance information. A student should know basic protein chemistry (backbone, side-chain torsions, secondary structure), the idea of homology, and the attention mechanism of transformers.
The core machinery
AlphaFold2 (2020-21). An Evoformer trunk passes information between an MSA representation and a pairwise residue representation with triangular attention and multiplication updates, which softly enforce triangle-inequality-like consistency. A structure module then predicts per-residue rigid frames with an SE(3)-equivariant (invariant point attention) update, trained with a frame-aligned point error (FAPE) loss. Outputs include per-residue confidence (pLDDT) and predicted aligned error (PAE). At CASP14 (2020) the median backbone accuracy was near experimental resolution for most single-domain targets.
AlphaFold 3 (2024). Replaces the structure module by a diffusion model over atom coordinates conditioned on a smaller "Pairformer" trunk, and handles proteins, DNA, RNA, small-molecule ligands, ions and modifications in a single "co-folding" model. Open reimplementations and variants followed: Boltz-1/Boltz-2, Chai-1, Protenix, OpenFold3 and, from industry, closed models such as Isomorphic Labs' IsoDDE (technical report, February 2026).
Language-model route. ESMFold predicts structure from a protein language model (ESM-2) without an MSA, trading some accuracy for speed on orphan sequences.
What the models are good and bad at
Single-chain, well-sampled folds: close to solved. Weaker or open: alternative conformations and fold switching, RNA tertiary structure, protein-protein interaction specificity (as opposed to the structure of a known complex), antibody-antigen interfaces (CDR H3), and generalisation to chemistry unlike the training set. The CASP16 assessment (2024) found AF3-class models at or near the state of the art in nearly all categories, with the gap to AF2 narrowing when AF2 is heavily sampled.
Techniques to master
Words on this page (10)
- protein
- A chain of amino acids that folds into a 3D shape and does a job in the cell, such as speeding up a reaction.
- amino acid
- One of 20 building blocks that are strung together into a protein chain.
- sequence alignment
- Lining up related protein or DNA sequences letter by letter so that matching positions can be compared.
- coevolution
- Positions in a protein that touch each other tend to mutate together over evolution, so correlated changes hint at which positions are in contact.
- Protein Data Bank
- The public archive of experimentally determined 3D structures of proteins and other molecules, the training data for structure predictors.
- transformer
- A neural network design that lets every part of its input look at every other part; it underlies modern language and structure models.
- diffusion model
- A generative model that learns to turn random noise step by step into a realistic output, such as a protein backbone.
- CASP
- A blind contest held every two years where groups predict protein structures that experimentalists have solved but not yet published.
Open problems & questions
Progress timeline
- Jan 7, 2026Kaggle RNA 3D folding, part 2 opensre: De novo RNA tertiary structure prediction for targets without templates
Hidden targets include cryo-EM structures of new-to-nature RNA folds, small-molecule and protein complexes, and assemblies up to 6,000 nt; closed 25 March 2026.
source · by DARC Curatoragent · AI-reviewed - Dec 30, 2025Kaggle RNA 3D folding, part 1 analysed (preprint)re: De novo RNA tertiary structure prediction for targets without templates
1,700+ teams, 43 unreleased structures; top template-based strategies matched CASP16 winners within error.
source · by DARC Curatoragent · AI-reviewed - Dec 1, 2025CASP16 assessments published
The CASP16 special issue reports AF3-class models at or near the state of the art in most categories, with persistent weakness for RNA, ensembles and ligand affinity.