BITTENSOR · SUBNET 107 · ROADMAP

Five Phases
To
Production

Minos is being built in five phases, starting with hyperparameter optimization of proven tools and progressing toward fully custom algorithms, consensus AI, and production-grade genomic services.

Each phase validates and expands the infrastructure, ensuring scientific rigor while enabling decentralized innovation at every step.

Phases5
CurrentPhase 01
Cadence72-min Rounds

Phase Overview

05 PHASES
01

Hyperparameter Optimization

Optimize configs for established variant callers

ACTIVE
02

Custom Genotyping Algorithms

Miners submit their own variant calling code

NEXT
03

Broader Variant Spectrum

SVs, CNVs, HLA, and difficult regions

PLANNED
04

Consensus AI

Meta-caller from top miner outputs

PLANNED
05

Production Deployment

Versioned containers, clinical reporting

VISION

Phase Details

01

Hyperparameter Optimization

ACTIVE

Miners optimize hyperparameter configurations for four established variant calling tools: GATK HaplotypeCaller, DeepVariant, bcftools, and freebayes. This phase validates the full infrastructure: challenge generation, HelixForge injection, hap.py scoring, EMA tracking, and Winner-Take-All weight distribution.

This Unlocks
01Proven accuracy benchmarking infrastructure
02Stable miner/validator coordination
03Continuously updated leaderboard of optimal hyperparameters
Keywords
GATKDEEPVARIANTBCFTOOLSFREEBAYESHAP.PY
02

Custom Genotyping Algorithms

NEXT

Miners submit novel variant calling algorithms as executable code. This unlocks algorithmic innovation: neural network callers, hybrid methods, and region-specialized algorithms. Code executes in sandboxed Docker containers.

Keywords
CUSTOM CODENEURAL NETSHYBRIDSANDBOXED
03

Broader Variant Spectrum

PLANNED

The competition expands beyond small variants to include structural variants, copy number variants, and clinically important regions like HLA genes. Scoring evolves to handle these more complex variant types that other pipelines omit due to their complexity.

Keywords
SVsCNVsHLADIFFICULT REGIONS
04

Consensus AI

PLANNED

Aggregate the best-performing miner outputs into a unified consensus model that combines the strengths of diverse approaches. This meta-caller is expected to exceed the accuracy of any individual pipeline.

Keywords
ENSEMBLECONSENSUSMETA-CALLERFUSION
05

Production Deployment

VISION

Ship the consensus caller as a production-grade service: versioned Docker images, automatic updates from subnet learnings, clinical-friendly reporting, and state-of-the-art accuracy tracking.

Keywords
DOCKERAPICLINICALPRODUCTION

Beyond Phase 5

FUTURE
Genotype ImputationBenchmark imputation algorithms against held-out high-coverage truth data
Long-Read SequencingPacBio HiFi and Oxford Nanopore variant calling benchmarks
Multi-SpeciesExpand beyond human genomes to other organisms