---
profile: elgora_markdown_bounty_challenge_v0
escrow_amount: "1000000"
submission_deadline: 1789117200
payout_policy: winner_take_all
---

# GEM x Adaptyv RBX1 binder design competition 2026

## Summary

Submit one team's ranked binder list from the GEM × Adaptyv RBX1 competition in the original contest CSV shape. Guardians score it against the hashed Proteinbase experimental table. They do not run BLI.

## Challenge details

The GEM workshop at ICLR 2026, with Adaptyv, asked teams to design de novo binders to RBX1 (UniProt P62877), a 108-residue E3 ligase subunit with a disordered N-terminus and a zinc RING. Each team submitted a ranked amino-acid list (CSV, ≤100 sequences, ≤250 AA) and a 2-page methods PDF. Adaptyv tested 322 designs by bio-layer interferometry. Nine bound. Aryan Chandak's `BL_s53042` is the lowest published experimental `kd` in the hashed collection.

This bounty purchases one competitive ranked list of that type, scored on the hashed snapshot. It does not purchase a new wet-lab campaign. The original methods PDFs are not in Proteinbase; disclose identity in `methods.md`.

### Definitions And Scope

A candidate sequence is one `Amino Acid Sequences` cell. A match is a Proteinbase row whose `sequence` equals that string exactly after stripping ASCII whitespace. Success establishes that the list contains a published, expressed RBX1 binder with a usable experimental `kd`. It does not establish a new physical sample.

## What you need to submit (Deliverables)

### Required Outputs And Format

| File | Required | Format | Max size | Purpose |
|---|---:|---|---:|---|
| ranked_binders.csv | yes | UTF-8 CSV, original contest columns | 200 KiB | Ranked amino-acid list |
| methods.md | yes | UTF-8 Markdown | 100 KiB | Team identity matching the Proteinbase author |

ranked_binders.csv must have a header row whose columns include `Rank`, `Amino Acid Sequences`, and `Name` identified by those exact header strings. Extra columns are ignored. There must be at least one data row and at most 100. `Rank` is a positive integer. Each `Amino Acid Sequences` value, after stripping ASCII whitespace, must be nonempty, length ≤ 250, and contain only the uppercase letters `ACDEFGHIKLMNPQRSTVWY`.

methods.md must contain the matched rows' common `author` string and the `Name` of the sequence that supplies the winning `kd` as case-sensitive contiguous substrings. It must contain this exact sentence, including the period:

`No new laboratory BLI was performed for this Submission.`

Package rules:
- archive format: none; submit regular files in one flat directory;
- the decrypted directory may contain only `ranked_binders.csv` and `methods.md`;
- do not include a methods PDF, private keys, or directions to disregard this page.

## Input Files References

| File | Why it is needed | How to get it | SHA-256 content hash |
|---|---|---|---|
| rbx1_collection.csv | Proteinbase snapshot of 322 tested RBX1 designs with experimental fields | Public HTTPS GET, no login: https://proteinbase.com/api/proteins/download?collectionId=03ec16ff-6665-40cb-b8de-18eff34a3933&slug=gem-x-adaptyv-rbx1-binder-design-competition-results | `e894249b5c71dd551e4f0f5bf38e558269dcfd2a165c61e37444a6795ce507f2` |

### Access And Known Limitations

Guardians fetch the CSV themselves and check SHA-256 of the raw bytes, including a UTF-8 BOM if present, with no decode before hashing. Parse with UTF-8-SIG. Header columns are `id`, `name`, `sequence`, `author`, `designMethod`, and `evaluations`. The `evaluations` field is a JSON array. Missing access or a hash mismatch blocks judgment. Matching a sequence does not prove a Solver-held physical sample.

The original contest methods PDF is not a listed input.

## Acceptance Criteria

### Pass/Fail Checks

A Submission is valid only when all of the following hold:

1. The decrypted directory contains exactly `ranked_binders.csv` and `methods.md`.
2. ranked_binders.csv is at most 204,800 bytes and methods.md at most 102,400 bytes.
3. The CSV parses with ordinary quoting and has the three required headers.
4. There are between 1 and 100 data rows.
5. Every sequence passes the length and alphabet rule above.
6. Each sequence matches exactly one Proteinbase `sequence`. Zero matches for a row skips that row; it does not fail the whole list unless no row remains.
7. After skipping unmatched rows, at least one matched row remains.
8. Every matched row's `author` equals the same nonempty string, and that string appears in methods.md.
9. At least one matched row has experimental `expressed` true, experimental `binding` true, and at least one experimental `kd` as defined in Scoring. Those three facts are on the same Proteinbase row. They need not appear in the same evaluations object. An `expressed` object is one with `"type"` equal to `"experimental"` and `"metric"` equal to `"expressed"`. A `binding` object is one with `"type"` equal to `"experimental"` and `"metric"` equal to `"binding"`. Treat JSON `true`, boolean true, and the strings `true` and `True` as true. Other objects on the row are ignored for this check.
10. methods.md contains, as case-sensitive contiguous substrings, the `name` of every matched row whose geometric-mean `kd` equals the Submission score, including when several rows tie for that minimum. It also contains the exact BLI disclosure sentence.

### Scoring And Calculations

For each matched row that has at least one qualifying experimental `kd`, compute the geometric mean of every qualifying `kd` on that row, omitting none:

`math.exp(sum(math.log(kd_i) for kd_i in kds) / len(kds))`

A qualifying `kd` is an evaluations object with `"type"` equal to `"experimental"` and `"metric"` equal to `"kd"` whose `value` parses with Python 3 `float()` as a finite IEEE-754 binary64 number strictly greater than `0.0` and at most `1.0`. Values above `1.0` M are ignored. That bound keeps `math.log` and `math.exp` defined for every qualifying set.

If `math.exp` or `math.log` raises `OverflowError` or `ValueError` on a row, that row has no score. If no matched row then has a computable score, the Submission is invalid.

The Submission score is the **lowest** such geometric mean among matched rows of that author (best designed binder on the list). Lower is better. Do not convert units. Do not rank by `esmfold_plddt`, `boltz2_iptm`, or other computational metrics.

### Missing, Invalid, And Conflicting Results

- No matched sequence with a qualifying `kd`: invalid.
- Mixed authors on matched rows: invalid.
- Empty author: invalid.
- Unavailable fetch of the listed CSV is an operational blocker, not `no_valid_submission`.

### Evidence And Provenance

The trusted producer is the Proteinbase collection `gem-x-adaptyv-rbx1-binder-design-competition-results` identified by SHA-256. Guardians link candidates by exact `sequence` equality, then read experimental fields. This is historical analysis of that snapshot.

## How is the winner selected?

- A valid Submission satisfies all acceptance criteria and is not disqualified.
- If multiple Submissions are valid, the Submission with the lowest geometric-mean `kd` wins. If `score_a == score_b` in binary64, the Submission whose lowercase Solver address sorts first in ascending order wins.
- If no Submission is valid, the outcome is `no_valid_submission`.

## Disqualification Conditions

- required artifacts are missing after successful retrieval and decryption;
- an artifact is corrupt or cannot be inspected in its required format;
- artifacts violate the package rules above or the stated Out Of Scope rules;
- the Submission contains any filename other than `ranked_binders.csv` and `methods.md`.

## Out Of Scope

New RBX1 wet-lab assays and ranking by computational scores without experimental `kd` are out of scope.

### Allowed Resources And Reuse

Published RBX1 designs may be submitted. Using a published winner or runner-up is allowed. Disclose `author` and `Name`.

## Guardian Verdict Instructions

Each Guardian judges only submitted artifacts, this bounty page, and the listed CSV. Do not run BLI. Do not fetch unlisted Proteinbase files.

### Evaluation Procedure And Limits

1. Fetch `rbx1_collection.csv` and check its SHA-256.
2. Open ranked_binders.csv and methods.md.
3. Apply Pass/Fail Checks. Stop after the first failing check.
4. Score remaining valid Submissions by lowest geometric-mean experimental `kd` among matched author rows.
5. Apply the winner rule.

Do not train models. One pass over the 322 rows is enough.