3D lab
Look at the molecules: the cancer protein HER2, the tiny antibodies that grip it, and the answer sheet we are being graded against.
New to this? What is a nanobody, an epitope and a structure?Plain-language primer
A is a very small antibody from camels and alpacas. A protein such as HER2 is a long chain of building blocks (residues) folded into a 3D shape. When a nanobody binds, a few residues on each partner sit close together. Those are the interface, the thing we must predict. Scientists see it by freezing the complex and imaging it () or by X-ray crystallography.
What does the colour code mean?
Magenta marks HER2 contact residues (the epitope). Teal and violet shades mark nanobody loops. Yellow marks nanobody framework residues that also touch the target. Orange marks the surface HER2 shares with its natural partners. Every layer has its own switch.
Does this page work offline?
The molecules are downloaded from the public RCSB archive when you open a scene, so they need an internet connection. Nothing large is stored on this site.
The 3D scenes
Choose a scene, then use the switches. Click any residue in the lists to fly the camera to it.
Switch layers on and off
Where does an antibody bind?
Residues touched by 3 or more public antibodies (click one)
Pick a residue to fly the camera to it.
What am I looking at, and what does the heat map really say?Most of HER2 has never been touched by a public antibody
This is the outside part () of HER2, residues 23 to 652, 630 in all. The first 22 letters of the full protein are a : a shipping label that is cut off before the protein is finished, so it is never in any structure. Older structure files also count from the cut point, so their numbers are 22 lower than the standard () numbers used on this page. We convert for you.
Across the 25 public binders in our panel, 213 of 630 residues are touched by at least one antibody and 417 are touched by none. Never-touched does not mean unbindable. It means nobody published a structure there, so a model trained on public structures has nothing to learn from.
The dimer patch
HER2 teams up with relatives (, , ). The surface that those partners cover in three solved pairs is about 51 residues, mostly in domain II (residues 221 to 351), plus a small rim on domain I. The four public nanobody footprints avoid it, except one residue (105) for 47D05. The challenge antigen is probably a single copy, so the patch is likely open. That is an inference, not a verified fact.
source
research/q1_her2_biology/erbb_family_dimerisation_signalling_and_adjacent_proteins_egfr_her3_her4.md
Where the heat-map numbers come from
Counts are from 25 public binder structures. Natural partner footprints are not included in the count. Approved-drug coverage of domains I to III is undercounted because several clinical antibodies have no public structure.
source
research/q1_her2_biology/her2_ecd_per_residue_antibody_coverage.jsonresearch/q1_her2_biology/her2_therapeutic_antibodies_adcs_epitopes_trastuzumab_pertuzumab_margetuximab_and_binding_footprints.md
source
research/q1_her2_biology/her2_public_complexes.jsonresearch/q1_her2_biology/her2_ecd_per_residue_antibody_coverage.json
All footprints, areas and coverage counts on this page come from these files. HER2 residue numbers are standard UniProt numbers (P04626).
What the answer key looks like
After the deadline, the organisers compare our guesses with a table made from the real structure. Each HER2 residue that touches the nanobody gets a row with a contact area (how many square angstroms of its surface touch the partner, measured with ). Nanobody residues get rows too. The page below uses public structures as stand-ins, because the real keys stay secret until judging. Our table is computed by our own script and will not match the organisers' key exactly.
Each bar is one HER2 residue along the ectodomain (left is the start). Bar height is contact area, the dashed line is your cut-off. Coloured band: domains I to IV.
| chain | seqNum | resName | fact | weight | |
|---|---|---|---|---|---|
| A | 34 | ARG | 1 | 3.2 | |
| A | 35 | LEU | 1 | 50.7 | |
| A | 36 | PRO | 1 | 4.7 | |
| A | 37 | ALA | 1 | 13.3 | |
| A | 39 | PRO | 1 | 76.2 | |
| A | 40 | GLU | 1 | 55.5 | |
| A | 42 | HIS | 1 | 2.0 | |
| A | 43 | LEU | 1 | 8.1 | |
| A | 64 | TYR | 1 | 7.6 | |
| A | 66 | PRO | 1 | 49.7 | |
| A | 67 | THR | 1 | 24.1 | |
| A | 68 | ASN | 1 | 3.1 | |
| A | 90 | GLN | 1 | 23.9 | |
| A | 134 | PRO | 1 | 0.1 | |
| A | 378 | PRO | 1 | 50.5 | |
| A | 382 | ASP | 1 | 26.4 | |
| A | 411 | TYR | 1 | 1.3 | |
| A | 413 | SER | 1 | 22.1 | |
| A | 414 | ALA | 1 | 7.7 | |
| A | 415 | TRP | 1 | 9.4 | |
| A | 417 | ASP | 1 | 33.6 | |
| A | 446 | GLN | 1 | 46.5 | |
| A | 447 | GLY | 1 | 46.8 | |
| A | 448 | LEU | 1 | 11.0 | |
| A | 449 | GLY | 1 | 16.8 | |
| A | 469 | HIS | 1 | 8.2 | |
| A | 470 | HIS | 1 | 78.9 | |
| A | 472 | THR | 1 | 77.1 | |
| A | 473 | HIS | 1 | 36.8 | |
| A | 496 | THR | 1 | 2.1 | |
| A | 497 | ALA | 1 | 42.3 | |
| A | 498 | ASN | 1 | 12.1 | |
| A | 499 | ARG | 1 | 4.5 | |
| A | 500 | PRO | 1 | 27.2 |
Columns follow the challenge file format: chain, seqNum, resName, fact (1 = interface), weight (contact area in Ų). Chain A is HER2 and chain B is the nanobody. Only the HER2 side is shown here. We show standard UniProt numbers, whereas the real file's numbering for chain A is not confirmed in our notes.
Why contact area matters for scoringA few big contacts carry most of the weight
The area-weighted score rewards finding the residues with the most contact. Slide the number below: with only the biggest few residues, how much of the total area do you already hold?
For 27A05, the 10 biggest residues hold 65% of the contact area, from 34 residues in total. These are H470 (domain III), T472 (domain III), P39 (domain I), E40 (domain I), L35 (domain I), P378 (domain III), P66 (domain I), G447 (domain III), and more.
source
research/q1_her2_biology/her2_public_complexes.json
Per-residue Voronota-LT areas for each public structure.
Why the epitope is a tiny part of HER2About 3 to 6 residues in every 100
On the five public nanobody structures, the epitope averages 3.2% of the observed HER2 residues when counted at 4 Å and 5.6% when counted by Voronota-LT. So a guesser that marks every residue as an interface scores badly: only a few in a hundred are positives. That low is what the score on the leaderboard has to beat.
source
research/q1_her2_biology/her2_ectodomain_architecture_domains_I_to_IV_and_structures_summary.json
Field base_rate_VHH in the summary file.
Nanobody anatomy explorer
A nanobody () is a small antibody made of one chain, about 120 letters long. Seven stretches matter: four steady stretches that form the fold, and three floppy loops called 1, 2 and 3 that reach out and grip the target. The loop surface that touches the target is the . CDR3 varies most in length.
Click a letter to see its position and region. Challenge numbering is sequential, with 1 = M (the two leading letters M and A are not part of the standard numbering).
How long is CDR3?
Purple bars are the three challenge nanobodies, grey bars are the public ones. Lengths use one rule for every sequence: from the residue after the cysteine that closes framework 3 up to the tryptophan of W-G-x-G.
Why CDR3 length decides how hard a nanobody isThe longest loop (ONU23) is the one structure predictors most often get wrong
From the hive notes: no public nanobody has a CDR3 within 70% identity of the three challenge sequences, so nothing can be copied from the public record. The challenge CDR3s run from 12 letters (ONU13) to 20 (ONU23). Long, kinked loops are where structure predictors fail most, so ONU23 is the likeliest to be mis-modelled and ONU13 the safest. ONU11 and ONU23 also carry the Phe/Phe pattern at positions 42 and 52 that is read as a sign of a kinked CDR3. ONU13 sits in between. These are predictions from published patterns, not observations of our three targets.
source
research/q2_vhh_immunology_structure/vhh_structure_framework_cdr_loops_cdr3_length_geometry_and_numbering_schemes.mdresearch/q2_vhh_immunology_structure/_QUEEN_DIGEST.md
How the regions were drawn, and how far to trust the borders
The challenge numbers residues one by one and does not use the standard antibody numbering (, ). Our region borders for the challenge sequences come from an anchor-rule script (find the conserved cysteines and tryptophans), not from the dedicated numbering software ANARCI. They may be off by one residue in framework 3 and CDR2. Re-running ANARCI is on the to-do list. For the public structures, CDR1 and CDR2 borders are the research worker's IMGT-like anchored ranges.
source
research/q2_vhh_immunology_structure/vhh_structure_framework_cdr_loops_cdr3_length_geometry_and_numbering_schemes.mdresearch/_meta/sepiq_verbatim.md
Sequences are verbatim from the challenge pages. The leading MA and trailing HHHHHH are shown but are not antibody.
The nanobody in 3D
This is nanobody 2Rs15d (entry 5MY6) on its own. Switch loops on and off, and the yellow sticks are the residues that touch HER2.
source
research/q1_her2_biology/her2_public_complexes.json
Data tables in this section cover 5 public nanobody chains plus the three challenge clones.
What this page cannot tell you
Limits of the structuresFour public nanobody structures, three of them very similar
Only four public HER2-nanobody structures exist (entries 5MY6, 7QVK, 9MTE and 9MTX; 9MTE holds two nanobodies). Entries 9MTE and 9MTX were released on 2025-11-19, so they are leakage-free stand-ins for the hidden test complexes. Two of the nanobodies (27A05 and 29E09) cover similar ground on domains I and III, so the examples are not fully independent.
Why some residues look missing in a scene
Some floppy stretches are never resolved in a structure. The domain I loop at residues 121 to 134 is missing in essentially every entry, and domain IV's tail is missing in many. Residues in those stretches cannot be drawn.
source
research/q1_her2_biology/her2_ectodomain_architecture_domains_I_to_IV_and_structures.md
Why footprint sizes differ between methods
The 4 Å list counts a residue if any atom comes within four angstroms of the partner. The Voronota-LT list counts any non-zero contact area, so it is larger. The challenge key uses Voronota-LT contact areas, so the wider list is the closer match.