Minos
SN107AI for genomics, building the foundational data layer for genetic research
Minos runs competitive DNA mutation-finding challenges every 72 minutes, building the foundational infrastructure layer for decentralized clinical genomics.
// The foundational layer of genomics
Minos (SN107) is a Bittensor subnet that runs competitive genomic variant calling challenges. Every 72 minutes, a synthetic DNA file with hidden mutations is released, and miners race to identify those mutations as accurately as possible using standard variant calling tools.
The simple version: It's like a recurring blind genomics exam where miners submit their best method for finding mutations in DNA, and validators grade each submission by re-running it independently.
Centralized equivalent: Commercial variant analysis pipelines from Illumina DRAGEN and Sentieon, or GATK Best Practices workflows, but with open competition replacing proprietary optimization.
How it works:
- Miners download a BAM file (DNA sequence data with hidden synthetic mutations), run one of four supported variant calling tools (GATK HaplotypeCaller, Google DeepVariant, FreeBayes, or bcftools) with optimized hyperparameters, and submit only their configuration, not the output.
- Validators download the same BAM file, re-run each miner's exact configuration independently, score the resulting variant calls against truth data using hap.py, and set on-chain weights proportional to accuracy.
- The problem it solves: Genomic variant calling, detecting mutations in DNA that indicate disease risk or guide treatment decisions, requires expensive compute infrastructure and lacks open, competitive benchmarking. No market mechanism currently incentivizes continuous improvement of variant calling methods.
- The opportunity: Genomics underpins personalized medicine, oncology, and drug discovery. A decentralized, continuously improving benchmarking layer for variant calling could lower the infrastructure barrier and produce open-access reference methods for the field.
- The Bittensor advantage: The trustless verification design matters here. Validators re-run each miner's config independently rather than trusting reported outputs, so results cannot be fabricated. The incentive structure aligns miners toward real clinical-grade accuracy, not API gaming.
- Traction signals: Mainnet launched May 1, 2026. As of early May 2026, the subnet holds approximately 3.1% of total Bittensor emissions, and the token is up over 130% in 30 days. One active miner is registered, the codebase has 33 commits from 3 contributors, and the most recent commit is dated May 4, 2026.
Category: Healthcare and Medical AI | Centralized Competitor: Illumina DRAGEN, Sentieon, GATK Best Practices
Genomic variant calling sits at the intersection of compute-intensive biology and precision medicine. The field has long been dominated by a handful of well-funded commercial platforms. Minos bets that open, incentivized competition can produce variant calling configurations that match or exceed proprietary benchmarks, and that decentralized trustless verification is the right architecture for something this consequential in healthcare.
Mechanism:
The subnet's scoring cycle aligns to Bittensor's 72-minute tempo. At each round, the Minos platform generates a fresh BAM file (standard genomics format: short DNA reads mapped to a reference genome) with synthetic mutations injected at the read level using HelixForge. Miners download the file via presigned URLs, with storage backed by Hippius SN75 decentralized storage, Cloudflare R2, and AWS S3 as fallback.
Miners choose one of four supported tools: GATK HaplotypeCaller, Google DeepVariant, FreeBayes, or bcftools. They tune hyperparameters via configuration files and submit only those configurations, not the variant call outputs. This is a deliberate design choice: validators re-run each miner's exact config on the same BAM, then score the resulting VCF file against truth data using hap.py, the industry standard tool for variant benchmarking.
Scoring uses four components: Core F1 score (60%), Completeness (15%), False Positive Rate (15%), and Quality (10%). Scores are smoothed via an Exponential Moving Average to reward consistency over time. During the warmup phase, rewards split across the top three miners at 50%/30%/20%. Once any miner reaches 10 scored rounds, the subnet shifts to a winner-take-all model where the single top performer by EMA receives 100% of emissions each round. The current phase is hyperparameter optimization. According to official docs, the next phase is custom algorithm competition once the hyperparameter search space saturates.
Mining hardware requirements are modest by Bittensor standards: minimum 4 CPU cores and 8 to 16 GB RAM. No GPU is required. Community sources note that miners can run on TargonCompute infrastructure at approximately $0.10 per hour, though this is not confirmed in official documentation.
From a network metrics perspective: the subnet holds approximately 3.09 to 3.21% of total Bittensor emissions as of early May 2026. The 30-day price performance ranges from 138% to 144% depending on the data source. Net buy volume for the past 7 days is approximately 324 TAO. The pool holds around 3,865 TAO with a root proportion of 0.55, meaning the protocol subsidy still covers slightly more than half of pool depth. This is normal for a newly launched subnet.
- Very early stage: Minos launched on mainnet May 1, 2026. One active miner, 33 GitHub commits, and 3 contributors describe a subnet that is days old. Infrastructure and scoring systems are live, but the network needs substantially more participants to validate its competitive benchmarking thesis.
- Winner-take-all mechanics: In the normal phase, 100% of emissions concentrate on the single top-performing miner by EMA. This rewards the leader strongly but creates meaningful barriers for new entrants trying to displace an established top miner.
- Team transparency: No named team members are listed in official documentation or the GitHub repository. Public contact is available at contact@theminos.ai, but accountability and team depth are not publicly verifiable.
- Thin pool depth: Root in pool is approximately 3,865 TAO with a root proportion above 0.5, indicating the pool is still predominantly protocol-funded. Larger TAO movements carry slippage risk until organic staking grows the pool.
- Centralized file delivery dependency: The current file delivery layer uses AWS S3 and Cloudflare R2 alongside Hippius SN75. The decentralized storage path is active but the fallback chain still includes centralized services.
