The methods, models, and evidence behind Bioscope.ai, published in full. Each paper links to the original publication.
Research
2024
Every result we show ships with its reference population, the covariates it was adjusted for, and an explicit confidence band. We argue that surfacing uncertainty, rather than burying it, is what makes a genomic result safe to act on.
Technical Report
Shotgun metagenomic profiles change faster than the genome and slower than the wearable feed. We characterise which compositional shifts are stable enough to act on, and how we surface them without over-claiming.
Methods
2025
Wearables, genome, microbiome, neuroimaging, and the medical record rarely speak the same language. We describe the multi-omic representation and retrieval approach that lets a single model reason across all five while citing its evidence.
We attribute the brain-age gap to specific structures and show that the trajectory across re-scans carries more clinical signal than any single estimate, with implications for early neurodegeneration screening.
2026
A raw polygenic score is a number without a frame. We present a stratification method that re-expresses scores as ancestry-matched percentiles adjusted for age, sex, and known modifiers, and quantify the clinical impact of miscalibration.
An inside look at the Bioscope interpretation pipeline: how we annotate, filter against reference populations, and rank the handful of variants worth a clinician's attention across all 3.2 billion base pairs.