Biology Atlas

Map / Assays

assay · deep · partly written

Structural evidence of complexes (crystallography and cryo-EM)

A structure shows which atoms of two proteins touch in a purified, crystallised or frozen state, which is why it is used as the yardstick for judging other interaction assays, and why its own errors matter.

Also called: protein complex structure, crystallography and cryo-EM of interfaces, biological assembly

The short version

A structure is a picture of a purified, often crystallised or frozen, state. It shows which atoms touch, which no other assay shows, and it is the yardstick behind 90% of HuRI pairs in complexes being direct contacts versus 81% for literature pairs and under 50% for complex maps. But the yardstick has its own error: at least 7% to 14% of PDB assembly annotations were estimated wrong. 123

Reported, not yet replicated

Crystallography: which contact is biological? more

The crystal contains many protein-protein contacts, about 10 lattice contacts per X-ray entry, and only some are biological. Deciding which is the hard step. Buried surface area alone averaged 0.73 accuracy. The tools PISA and EPPIC reached 83.5% and 85.3% on 1,481 consensus-annotated proteins; their ground truth was the annotation suspected of errors. A random forest scored 0.92 on its large set and 0.74 on an independent 160-case set, so heavier tools lose on new data. 342

Reported, not yet replicated

Cryo-EM: read the interface, not the title resolution more

Cryo-EM has no crystal lattice, but a map is not uniform: flexible regions are blurry. Read the local resolution at the interface and a per-atom measure such as the Q-score, which falls from about 1 to about 0.3 across 1 to 5 A resolution and does not check stereochemistry. Cryo-EM is now almost half of new PDB entries (45.3% in 2025). The general cryo-EM method description was not source-verified. 56

Reported, not yet replicated

How a structure becomes ground truth more

Chain: purify a construct, form the complex, crystallise or freeze, build a model, deposit with an author-chosen biological assembly, and let a benchmark take that assembly as truth to grade an assay or predictor. Each arrow loses something: constructs miss full-length and PTM-dependent pairs, crystallisation can miss the weakest complexes, and annotation is wrong for at least 7% to 14% of entries. The consequence is asymmetric: a structure raises confidence that two proteins can touch, but its absence lowers it only weakly. 321

Reported, not yet replicated

Predictors graded against structures deep

AlphaFold-Multimer success about 40% to 60% by DockQ above 0.23 on 1,928 post-cutoff complexes. The authors warn that a PDB assembly may not be the only physiological one, so a correct prediction can score as wrong; leakage in interface benchmarks (86% under PPI-code splits) means headline numbers need interface-similarity splits. 78

Reported, not yet replicated

Question answered

Which atoms of protein A touch which atoms of protein B, and is that contact biological or an artefact of how the sample was prepared? It is the one layer that says where on the proteins they interact, and it is the reference set used to grade other assays and predictors. 13

Reported, not yet replicated

Readout physical

Electron density. Crystallography diffracts X-rays through an ordered crystal and solves electron density, from which atoms are built; the deposited coordinates are for the asymmetric unit (the repeating piece of the crystal), not for the complex. Cryo-EM builds a three-dimensional density map from many particle images and atoms are built into the map. 35

Reported, not yet replicated

Not measured

Affinity, kinetics and whether the complex forms in a cell. Treat a structure as strong evidence that two proteins can touch, at this resolution, in this prepared state, not as proof of cellular binding. Absence of a structure usually means 'not tried or not crystallisable', so it lowers confidence only weakly. 13

Reported, not yet replicated

Setting

Where in cascade

Sample prep

A purified construct, often truncated, tagged or engineered, formed into a complex at high concentration, then crystallised or frozen. Each step can lose something: PTM-dependent pairs and full-length proteins may be missing, and the weakest complexes have binding energies comparable to lattice contacts, so crystals can miss them. 3

Reported, not yet replicated

Conditions

In crystals the load-bearing step is deciding whether an interface is biological or a lattice contact: the average X-ray entry has about 10 lattice contacts, and about 64% of PDB entries have at least two crystal forms. In cryo-EM the quality of an interface depends on local resolution at the interface, not the global resolution, and the Q-score (how well each atom is supported by the map) falls from about 1 to about 0.3 across 1 to 5 A resolution over 10,189 map-model pairs. 35

Reported, not yet replicated

Throughput

No throughput figure found. As a share of new entries, cryo-EM and X-ray together account for almost all PDB depositions: in 2025, 53.1% X-ray, 45.3% cryo-EM and 1.3% NMR, with 247,287 entries at the end of 2025. 6

Reported, not yet replicated

Engineerable variants

Interface classifiers (PISA, EPPIC, buried surface area)

Computational calls of biological versus lattice interfaces. EPPIC (geometry plus conservation across homologues) had accuracy 0.81 on its tuning set (not an independent estimate), 0.89 and 0.86 on two benchmarks against PISA's 0.79, 0.84 and 0.77. On 1,481 consensus-annotated proteins EPPIC was 85.3% correct and PISA 83.5%; where the two agreed (78% of cases) error was about 5%, and where they disagreed (22%) each was wrong about half the time. Buried surface area alone averaged 0.73 accuracy. 924

Reported, not yet replicated

Multiple crystal forms and mutagenesis

Seeing the same interface in more than one crystal form, or breaking it by mutation, are experimental checks on a biological call. 3

Reported, not yet replicated

Interface-similarity data splits

For benchmarks, splitting by interface similarity rather than sequence identity avoids leakage: on average 86% of test examples had a near-identical training interface under PPI-code splits of random PDB samples (56% for the SKEMPI v2.0 split). 8

Reported, not yet replicated

Use cases

Grading the human reference interactome (HuRI)

Structures of multi-subunit complexes were the yardstick for how often an interaction map reports direct contact: 90% of HuRI pairs in complexes were direct contacts (n = 121), against 81% for literature-curated binary pairs (n = 410) and under 50% for complex maps. The yardstick covers only pairs that already have a structure inside a multi-subunit complex, not the whole map. 1

Reported, not yet replicated

ML role

Does

AlphaFold-Multimer predicts complex structures. On 1,928 post-cutoff complexes its success (DockQ above 0.23) was about 40% to 60% across oligomer sizes, and a correct prediction can be scored wrong if the deposited assembly is not the only physiological one. Structures are also the training and test sets for interface predictors. 7

Reported, not yet replicated

Cannot

Standard AlphaFold-Multimer confidence metrics 'do not reliably separate relevant PPIs from false positive predictions'; a classifier called SPOC reached 50.4% recall at 5% FDR. Evaluation leaks: about 86% of test interfaces had a near-identical training interface under PPI-code splits. 108

Reported, not yet replicated

Data availability

Protein Data Bank (wwPDB / RCSB) public

247,287 entries at the end of 2025; 2025 depositions were 53.1% X-ray, 45.3% cryo-EM and 1.3% NMR. 623

Caveats. The author-chosen biological assembly is the usual benchmark truth, and assembly annotations were estimated wrong for at least 7% to 14% of entries (Capitani 2016 cites 14.4%). Entries are a purified or frozen state, not the cell.

Reported, not yet replicated

AlphaFold-Multimer benchmark code and data public

The authors state their code, models and data are free at gitlab.com/ElofssonLab/afm-benchmark. 7

Caveats. Repository link was confirmed in the paper; its licence and last commit were not checked.

Reported, not yet replicated

Teach this

In one sentence

A structure shows which atoms of two proteins touch in a purified or frozen state; it is the best yardstick for 'direct contact' but not proof the complex forms in a cell. 13

Reported, not yet replicated

The five-minute version

Crystals and frozen particles give atomic coordinates, so structure is the layer that says where on two proteins they meet, and it grades other assays: 90% of HuRI complex pairs were direct contacts against under 50% for complex maps. But a crystal holds about 10 lattice contacts per entry, so the biological interface must be picked out, and the deposited assembly is wrong for at least 7% to 14% of entries. Cryo-EM avoids the lattice but needs local resolution checked at the interface. Neither gives Kd, kon or koff. 1325

Reported, not yet replicated

What a listener will ask

If there is a structure, do the proteins interact in cells?

It says they can touch in the prepared state; affinity, kinetics and cellular formation need other assays.

Why do assembly annotations matter?

Benchmarks take the deposited assembly as truth, and at least 7% to 14% of annotations were estimated wrong. 13249785610

Reported, not yet replicated

Is a prediction wrong if it disagrees with the PDB?

Not necessarily: the PDB assembly may not be the only physiological one, which the AlphaFold-Multimer benchmark authors warn about.

The common wrong idea

Everything in a crystal structure is a real biological interaction.

A crystal holds about 10 lattice contacts per entry and only some are biological, which is why PISA, EPPIC and multiple crystal forms are used. 13249785610

Reported, not yet replicated

The one citation to hand over: 3

Still open here

Not yet written: Time per run, Cost, Before after, Capital.

Linked from

Sources

  1. Luck, K., Kim, D.-K., Lambourne, L., et al.. A reference map of the human binary protein interactome. Nature (2020). paper; opened 2026-11-17 all records citing this
  2. Bliven, S., et al.. Automated evaluation of quaternary structures from protein crystals. PLoS Computational Biology (2018). paper; opened 2026-11-17 all records citing this
  3. Capitani G, Duarte JM, Baskaran K, Bliven S, Somody JC. Understanding the fabric of protein crystals: computational classification of biological interfaces and crystal contacts. Bioinformatics (2016). paper; opened 2026-11-17 all records citing this
  4. Elez K, Bonvin AMJJ, Vangone A. Distinguishing crystallographic from biological interfaces in protein complexes: role of intermolecular contacts and energetics for classification. BMC Bioinformatics (2018). paper; opened 2026-11-17 all records citing this
  5. Pintilie G, Shao C, Wang Z, et al.. Q-score as a reliability measure for protein, nucleic acid, and small molecule atomic coordinate models derived from 3DEM density maps. bioRxiv (2025). preprint, not peer reviewed; opened 2026-11-17 all records citing this
  6. wwPDB newsletter 2026 Q1: deposition statistics. wwPDB / RCSB PDB (2026). organisation-page, not peer reviewed; opened 2026-11-17 all records citing this
  7. Zhu W, Shenoy A, Kundrotas P, Elofsson A. Evaluation of AlphaFold-Multimer prediction on multi-chain protein complexes. Bioinformatics (2023). paper; opened 2026-11-17 all records citing this
  8. Bushuiev, A., et al.. Data leakage in protein interaction benchmarks (arXiv 2404.10457). arXiv (2024). preprint, not peer reviewed; opened 2026-11-17 all records citing this
  9. Duarte JM, Srebniak A, Schaerer MA, Capitani G. Protein interface classification by evolutionary analysis. BMC Bioinformatics (2012). paper; opened 2026-11-17 all records citing this
  10. Schmid, E. W., Walter, J. C.. Predicting protein-protein interactions with AlphaFold-Multimer: metrics and the SPOC classifier. Europe PMC (2025). paper; opened 2026-11-17 all records citing this
Where it sits on the map
History of this record (1)

This record has been changed 1 time, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.

Throughput class
low
Emerging
no
Status
partly written
Last updated
2026-11-17
Record
content/assays/structural-evidence-of-complexes.json