About the team

Why we picked Challenge Six: Genome Firewall, how we built it, and how we measure whether it worked.

01

Why we chose Challenge Six: Genome Firewall

Antimicrobial resistance is projected to cause up to 10 million deaths per year by 2050. Clinicians increasingly have access to whole-genome sequencing but no calibrated, auditable way to turn that data into an antibiotic-response signal.

Challenge Six sits exactly at that gap. It is a defensive research tool that reads an assembled genome and emits a decision receipt: likely to work, likely to fail, or a deliberate no-call. We picked it because the safety framing (abstain rather than guess) matches how we already think about ML in high-stakes settings.

02

Meet the team

HUI
Hüseyin Umut Işık
Backend & model
EMB
Ege Mert Balçık
Bioinformatics & pipeline
JW
Joy Wang
Product manager & frontend

Three of us, one rule: don't ship what you can't defend (quod erat demonstrandum).

03

Tech stack & security considerations

Stack
FrontendReact 19, TypeScript, Vite, Tailwind v4
UI kitshadcn/ui, Lucide, Recharts
3D / vectorthree.js DNA helix, SVG backdrop
Routing / SSRTanStack Start v1
Backendgenomefw FastAPI, AMRFinderPlus
Modelsrules gate · calibrated LR / XGBoost · no-call & OOD
DeployVercel (frontend), FastAPI (backend)
Security posture
  • Sequence data is treated as data, never forwarded to an LLM or third party.
  • Species gate and QC gate reject the wrong organism or a poor assembly before any model runs.
  • Out-of-distribution safety layer downgrades to no-call when signals are weak or conflict.
  • Every decision produces a JSON receipt (rules fired, versions, thresholds).
  • Persistent safety banner and modal: research prototype, not a prescribing tool.
  • 04

    Choice, implementation & UI/UX

    We built a lab-report aesthetic (monochrome paper, highlighter accents, monospace for the technical data) so every screen reads like something you could staple into a notebook. The Analysis page presents Analyze, Evidence & Trace, and Model Validation.

    How we measure success
    Data accuracy

    Held-out AUROC ≥ 0.90 per drug and ECE ≤ 0.05 after calibration on unseen genetic groups.

    Model efficiency

    Under 90 s end-to-end per genome on commodity CPU, with an honest no-call layer.

    Business objectives

    An auditable receipt for every decision, zero silent failures, scope explicit on the UI.

    05

    Industry potential

    Clinical microbiology

    Decision support alongside phenotypic AST, flagging high-risk isolates for confirmation faster.

    Public health surveillance

    Population-scale AMR trend detection from sequenced isolates with calibrated uncertainty.

    Pharma R&D

    Prioritize candidate compounds against genotypes predicted to escape existing drugs.

    06

    What we learned and what challenged us

    Learned
    • Calibration matters more than raw accuracy for a "should we abstain?" system.
    • An evidence receipt is a UX feature, not just an ML artefact.
    • three.js and SVG make dense scientific data feel legible.
    Challenged us
    • Designing a no-call path that users trust instead of routing around.
    • Tuning the OOD layer to fire often enough to matter but not so often it is noise.
    • Keeping the safety framing visible without making the product feel timid.