SEPIQ 2026 Intelligence

Can we predict where a tiny antibody sticks to a cancer protein?

SEPIQ 2026 is a blind contest: for three antibodies ( nanobodies) that grip the cancer-linked protein , predict which residues on each side touch. The answers are 3D structures nobody outside the organisers has seen.

14 days leftNo score feedback from the portal5 uploads allowedPrizes $3,000 / $2,000 / $1,000
source
  • BRIEF.md
  • research/q8_compute_validation_execution/_QUEEN_DIGEST.md
What is SEPIQA prediction contest run by SEPIQ and COGNANO: rank every residue by how likely it is part of the contact patch, scored by averaged over the three antibodies.
What is at stakePrize money and an invitation to co-author the consortium paper, and a test of the idea that antibody sequence data can reveal 3D contacts.
What we knowThe antibody side is predictable from known loop habits. The HER2 side is near chance without a real 3D structure prediction. The big sequence dataset did not help on the HER2 side.
What we do nextUpload the safe baseline, run structure predictions for the three targets, blend only what beats the baseline on clean practice tests, freeze on 28 Oct.
Show the evidence behind these four boxes
  • Track A, scoring, prizes and dates are from the organiser pages as read on 2026-10-17 (
    source
    • BRIEF.md
    ).
  • The data-representation idea (that the sequence data hold per-residue contrast for HER2) was tested four ways and failed (
    source
    • hive/EMPEROR.md
    ).
  • The plan and its stop rules are in the runbook (
    source
    • research/q8_compute_validation_execution/day_by_day_runbook_to_2026_10_31_and_failure_mode_register.md
    ).

Six numbers

14
days left
Track A closes 31 Oct 2026; our last upload is planned for 30 Oct
Why this number

The organisers do not state a time zone, so the plan treats 30 Oct 06:00 UTC as the safe last upload. Results are announced 15 Dec 2026.

source
  • BRIEF.md
  • research/q8_compute_validation_execution/day_by_day_runbook_to_2026_10_31_and_failure_mode_register.md
3
test complexes
ONU11, ONU13 and ONU23 are the only targets we are graded on
Why this number

With only three answers, a score gap smaller than about 0.1 is mostly luck, so the plan stays simple and robust. The structures that define the truth are released only after judging.

source
  • BRIEF.md
  • research/q8_compute_validation_execution/_QUEEN_DIGEST.md
3-5%
of HER2 surface is the answer
Base rate for the ; guessing at random scores about this
Why this number

Measured on five public practice complexes: 2.9% to 5.5%, mean 4.2% of the 618 modelled HER2 residues. The 12.6% figure quoted elsewhere applies to one entry only (8GZ5, a different region of HER2). On the antibody side the rate is higher, 11% to 22%.

source
  • model/README.md
19,367
labelled antibodies
of 1,843,875 sequences; 7,316 Yes, 12,051 No
Why this number

The labels say whether an antibody enriched during , not where it touches HER2. We tested four ways to squeeze per-residue information from the data and all four failed, so the HER2 side must come from structure.

source
  • hive/EMPEROR.md
0.18-0.23
best-known method scores
: organiser baseline 0.176; best published surface predictor 0.232
Why this number

The 0.176 is the organisers' own test on one entry; random guessing reaches it about 16% of the time, so it is only about 1.2 times random. The 0.232 (DiscoTope-3.0) is for antibody-blind prediction at 6.4% base rate, about 3.6 times random. Neither number is a prediction for ONU11, 13 or 23.

source
  • research/q6_epitope_ml_methods/_QUEEN_DIGEST.md
  • BRIEF.md
0.389
our floor score
Practice-set PR-AUC of the safe baseline; random would give about 0.10
Why this number

Split by side: HER2 surface 0.082 (random 0.051, not statistically distinguishable), antibody side 0.692 (random 0.203). Nearly all of the gain comes from knowing which antibody loops usually bind. These are proxy numbers on five known complexes, not a forecast, and the portal returns no score.

source
  • model/README.md

The five submissions

  1. S1
    Oct 20
    Safe baseline (CDR and surface prior)
    Built and validated locally; not uploaded
  2. S2
    Oct 24
    Baseline plus structure-prediction blend, only if it helps
    Planned
  3. S3
    Oct 26
    Best blend of the arms that passed their gates
    Planned
  4. S4
    Oct 29
    The real candidate, after QA
    Planned
  5. S5
    Oct 30, by 06:00 UTC
    Reserve or hedge, only if S4's stored files are fine
    Planned
How each upload is decidedGates are measured on public practice complexes, since the portal shows no score

Each step is promoted only if it beats the safe baseline on the practice tests by a pre-set margin; otherwise the simpler file goes up. Which of the five uploads the organisers actually score (last or best) is not stated, so the real candidate is uploaded on 29 Oct and a reserve on 30 Oct.

source
  • research/q8_compute_validation_execution/day_by_day_runbook_to_2026_10_31_and_failure_mode_register.md
  • model/README.md

Top five risks

Residue numbering off by a few positionsProbability High, impact High

If our numbering is shifted, every score lands on the wrong residue and the portal will not warn us.

What we do: Round-trip test and a strict validator (built; self-test of 24 cases passes).

source
  • research/q8_compute_validation_execution/day_by_day_runbook_to_2026_10_31_and_failure_mode_register.md

Register row 2 of 21. Probability and impact are the queen's judgement, not measured.

Practice tests look better than they areProbability High, impact High

Two of our four public practice structures were seen by the prediction engines during training, which inflates scores.

What we do: Judge on the two clean structures (9MTE, 9MTX) and state the leakage openly.

source
  • research/q8_compute_validation_execution/day_by_day_runbook_to_2026_10_31_and_failure_mode_register.md

Register row 9 of 21. Probability and impact are the queen's judgement, not measured.

Upload fails halfwayProbability Medium, impact High

The portal sends three files in separate calls, so one or two can be saved and the rest lost.

What we do: Upload the baseline early to test the flow; retry the whole set; keep receipts.

source
  • research/q8_compute_validation_execution/day_by_day_runbook_to_2026_10_31_and_failure_mode_register.md

Register row 1 of 21. Probability and impact are the queen's judgement, not measured.

Our scorer is wrongProbability Medium, impact High

We do not know exactly how the organisers score, and a mistake in our copy would mislead every decision.

What we do: Check our scorer against a library and report every scoring variant.

source
  • research/q8_compute_validation_execution/day_by_day_runbook_to_2026_10_31_and_failure_mode_register.md

Register row 19 of 21. Probability and impact are the queen's judgement, not measured.

All three antibodies get the same answerProbability Medium, impact High

If the model paints one patch for all three antibodies, a single wrong guess is paid for three times.

What we do: Predict each antibody separately; keep a low-variance hedge for S5.

source
  • research/q8_compute_validation_execution/day_by_day_runbook_to_2026_10_31_and_failure_mode_register.md

Register row 10 of 21. Probability and impact are the queen's judgement, not measured.

What red and amber mean

Red: high probability and high impact. Amber: medium probability and high impact. The five shown are the highest-ranked of 21 rows in the register, chosen by us from probability times impact; the full register is in the source file.

Go deeper