Question 1

How is binding affinity predicted for submitted molecules and nanobodies?

Accepted Answer

Small molecules are scored using Boltz-2, a structure prediction model that predicts binding affinity by folding the molecule with the target protein and evaluating the interaction. Nanobodies are scored using a modified BoltzGen pipeline that folds the designed sequence with the target and ranks it on ten metrics: structural confidence (iPTM, pTM, PAE), interaction quality (hydrogen bonds, salt bridges, buried surface area), and developability (liability motifs). Final ranking uses sum-of-ranks across all metrics, so consistency across all quality dimensions determines the winner.

Question 2

What formats do I submit molecules and nanobody sequences in?

Accepted Answer

Both are submitted in a single encrypted string with the format `molecules|sequences`. Small molecules are molecule IDs from the combinatorial database (e.g., `rxn:3:49485:2099:70633`), comma-separated if multiple. Nanobodies are amino acid sequences (e.g., `EVQLVESGGGLVQPGG...`), also comma-separated. Use `~` as a null placeholder if you compete in only one category (e.g., `~|sequence1,sequence2` for nanobodies only). The payload is encrypted with Bittensor Drand Timelock (preventing front-running) and uploaded to a public GitHub repo; the validator fetches and decrypts it.

Question 3

What validity checks do my submissions have to pass?

Accepted Answer

Submissions pass two filter layers. Layer 1 (on-chain): format validation (exactly one `|` separator, no whitespace/non-ASCII), sequence length constraints (typically 90–150 residues), amino acid validity (only 20 standard AAs), no duplicate sequences, cysteine count requirements (minimum 1, usually 2 for the canonical disulfide bond), homopolymer run limits, and global uniqueness (no resubmission of sequences seen before for that target). Layer 2 (NOVA Nanobody Filter): diversity check against top 50 leaderboard designs (Jaccard similarity < 0.9), IMGT numbering validation, nativeness score ≥ 0.80 (IgBLAST vs VHH database), humanness score ≥ 0.75, and TNP developability filters (CDR length, hydrophobic/charge patch limits). Any filter failure rejects your entire submission for that epoch.

Question 4

How long does a competition epoch last and when can I submit?

Accepted Answer

Each epoch lasts approximately 361 blocks. Target proteins are announced ~24 hours before the epoch start, giving you lead time to design candidates locally. Submissions must be committed on-chain within the epoch window but are not accepted in the final 10 blocks of the epoch. This prevents last-minute front-running while ensuring validators have time to decrypt, validate, score, and set weights before the next epoch begins.

Question 5

Can I use tools like BoltzGen, ODesign, or RFdiffusion to generate my nanobody candidates?

Accepted Answer

Yes. The documentation recommends BoltzGen as the primary tool (since scoring uses BoltzGen metrics) but lists alternatives like ODesign, RFdiffusion, ProteinMPNN, and RoseTTAFold. Since submissions are scored on BoltzGen metrics, designs optimized for other tools may not rank as highly; a robust strategy is to generate candidates with multiple tools, pre-screen them locally through BoltzGen, and submit the best-ranked set to maximize consistency across all scoring dimensions.

Question 6

What happens if my submission fails validation?

Accepted Answer

If any molecule or nanobody in your submission fails a validity check, your entire submission for that type is rejected and receives no score for that epoch. You do not earn rewards and the participation is burned. This is why pre-screening locally through open-source validator code is critical: replicate Layer 1 checks yourself and, if possible, run the NOVA Nanobody Filter service locally before committing to the chain.

Question 7

How is the winner determined and how are rewards distributed?

Accepted Answer

Validators calculate a final score per UID per competition type (molecules and nanobodies) by aggregating per-metric ranks. For nanobodies, each sequence is ranked independently on all 10 metrics, then ranks are summed (lower total rank wins). Multiple targets are averaged per sequence, then summed across sequences. The UID with the best (lowest in rank-based scoring) final score wins; ties are broken by earliest block submission, then push timestamp. On-chain weight distribution applies: 73.7% to UID 0 (burn), remainder distributed to winners. Validators also share scores via an external API to increase robustness against non-deterministic model outputs.

NOVA

NOVA

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup

Participate

Participate

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup