ResistSense — A Genome Firewall for AMR Evidence

The description states that ResistSense is a safety-first genome firewall for antimicrobial-resistance research. It accepts an assembled E. coli FASTA file and evaluates it against five antibiotics using:

• Statistical models (e.g., calibrated logistic regression, scikit-learn)
• AMRFinderPlus
• Conformal prediction
• Out-of-distribution detection
• Independent molecular-target gates

It returns one of three outcomes:

1. Resistance-associated signal
2. Susceptibility-compatible signal
3. No-call due to insufficient or conflicting evidence

The system includes:

• Firewall Replay (separates statistical association, biological evidence, safety triggers, and final assessment)
• Prediction Autopsy (shows errors prevented and residual errors)
• Human-review worklist for no-calls
• Evidence Passport (sample, model, policy, runtime, privacy, and split provenance)
• 3D bacterial illustrations
• Downloadable machine-readable reports
• A constrained GPT-5.6 Sol auditor that checks contradictions but does not change science

Not evidenced No information on actual deployment, real-world usage, or integration into existing workflows.

The author claims ResistSense is a genome firewall for AMR evidence, designed with safety as its core principle. It positions itself as a system that knows when not to predict — avoiding binary outputs in uncertain cases.

It also claims:

• GPT-5.6 Sol audits contradictions without changing scientific outcomes.
• The system prevents 55 of 68 frozen-test base-model errors.
• It maintains high empirical conformal coverage (>94%) and selective accuracy (e.g., 99.7% for cefotaxime).
• It is a defensive research prototype, not a diagnostic device.

Inference The positioning suggests a shift from traditional AMR prediction systems toward a more cautious, evidence-driven model that prioritizes uncertainty awareness over raw accuracy.

The description states that ResistSense supports five antibiotics and evaluates assembled E. coli FASTA files.

It is built for antimicrobial-resistance research, particularly in contexts where:

• Genomic data must be interpreted with caution
• Scientific rigor requires explicit handling of uncertainty
• Users need traceable, reproducible evidence

The system includes features like Prediction Autopsy and Evidence Passport, suggesting a user base that values transparency and auditability.

Not evidenced No specific customer personas, use cases beyond the hackathon, or target institutions (e.g., hospitals, labs, research centers) are mentioned.

The description does not state any business model or pricing strategy. It is described as a research prototype, not a commercial product.

Not evidenced No revenue streams, monetization plans, or pricing information are provided.

The system uses:

• Backend: Python, FastAPI, scikit-learn, AMRFinderPlus, conformal prediction, OOD detection
• Frontend: React, TypeScript, Three.js
• AI integration: GPT-5.6 Sol via OpenAI API with structured outputs and strict input allowlists
• Deployment: Docker container on Google Cloud Run, Firestore, Nginx

Key technical features include:

• Frozen genetic groups to prevent leakage
• Structured inputs for GPT-5.6
• Deterministic fallbacks if OpenAI fails
• Transactional Firestore guard with budget controls
• Adversarial safety suite (12-case test)

Not evidenced No details on scalability, performance metrics, or production infrastructure beyond the hackathon deployment.

The project is described as a research prototype submitted to a hackathon. Accomplishments include:

• 94%+ empirical conformal coverage
• Prediction Autopsy identifies 68 frozen-test base-model errors
• Firewall prevents 55 of those errors
• 47 machine-readable release gates passed
• 73 automated backend tests pass

It also mentions:

• Permanent Cloud Run deployment remains available when development computer is offline
• GPT-5.6 boundary includes adversarial safety suite

Not evidenced No real-world adoption, user feedback, or performance data from external users.

The description does not mention competitors directly. However, it implies a space of AMR prediction tools that often compress complex genomic evidence into binary outputs — which ResistSense aims to improve upon by introducing:

• Explicit abstention
• Conformal uncertainty
• Independent molecular-target gates
• GPT-based contradiction auditing

It is positioned as a safer alternative to traditional AMR systems, especially in safety-critical contexts.

Not evidenced No comparison with existing tools or platforms in the AMR space.

1. Prototype-only status: The system is described as a research prototype, not a production-ready tool.
2. No real-world validation: No mention of external or clinical validation beyond the hackathon.
3. Single-person team: Only one member (Abel Mancilla) was involved in building and decision-making.
4. Limited scope: Supports only E. coli and five antibiotics.
5. GPT-5.6 dependency: While constrained, reliance on a proprietary LLM introduces potential risks around availability, cost, and control.
6. No commercialization path: No indication of how the project might evolve into a product or service.

1. What is the current status of the project beyond the hackathon? Is it being used in any research or clinical settings?
2. How was the dataset curated, and what steps were taken to ensure data quality and avoid leakage?
3. Can you explain how the GPT-5.6 Sol auditor is integrated into the system and how its outputs are validated?
4. What are the plans for expanding support beyond E. coli and five antibiotics?
5. How does the team plan to address scalability, cost, and availability of the AI components (e.g., OpenAI API)?
6. Are there any potential regulatory or compliance concerns with deploying this in real-world settings?