Turning sequencing files into a research surface
AWS has published a reference architecture for antimicrobial resistance surveillance that moves public sequencing data through managed bioinformatics processing and into a conversational research interface. Dated 12 October, the design uses AWS HealthOmics for assembly and screening, then normalises gene detections into a common schema. The purpose is to let researchers examine resistance patterns without first writing bioinformatics code or SQL.
Raw screening tools often produce one tab-separated result per isolate. That is practical for an individual sample but awkward for analysis across organisms, locations or time. A Lambda function converts outputs to the hAMRonization schema, and AWS Glue merges them idempotently into Apache Iceberg tables on Amazon S3 Tables. One table holds detected genes and another records isolate provenance and licensing.
Structured evidence sits beneath the conversation
Amazon Athena queries the Iceberg data, while a Strands-based agent runs in Bedrock AgentCore. The agent has four defined tools: finding resistance profiles, summarising trends, comparing isolates and looking up a gene. Each tool translates a natural-language request into a targeted query. The model does not answer directly from an unstructured memory of antimicrobial resistance; it reasons over returned rows from the prepared dataset.
That architecture matters because the conversational layer can be replaced or improved without changing the underlying evidence. Researchers can inspect the data and rerun the query independently. The Iceberg tables also preserve a transactional, columnar surface for conventional analytics, so the agent is an additional interface rather than the only route to the results.
Grounding rules are unusually explicit
The system prompt requires every claim to include the gene identifier, the detecting tool and its confidence value. If the data does not support an answer, the agent must say so instead of guessing. Bedrock Guardrails adds contextual-grounding checks and denied topics that prevent the assistant from drifting into medical or patient advice. The example uses Anthropic Claude Opus 4.7 through a cross-Region Bedrock inference profile.
These rules turn citations into part of the output contract. A response such as “resistance was detected” is insufficient without the underlying gene and method. That gives researchers a path from the conversational summary back to the assay output. It also makes evaluation clearer: reviewers can test whether each statement is traceable, not merely whether it sounds scientifically plausible.
Reproducibility is built into deployment
AWS supplies the architecture through a CDK application with no hardcoded account identifiers. It provisions ingestion, processing, tables, agent runtime and the web front end. The reference can therefore be deployed in another account and tested with the same structure. Costs vary with the number of isolates processed, Glue work and agent calls, rather than a single fixed subscription.
Reproducible infrastructure does not guarantee reproducible science. Teams still need versioned reference databases, tool versions, parameters and input-access dates. Public sequencing records may be corrected, and resistance interpretation can evolve. The agent should expose those boundaries whenever they affect a result.
Useful surveillance, not clinical advice
The architecture is aimed at population and research questions such as which resistance genes appear in Salmonella samples or how detections change over time. It is not positioned as a system for choosing an individual patient’s treatment. Keeping that distinction in the interface and guardrails reduces the chance that a research summary is mistaken for a clinical recommendation.
Australia has strong reasons to monitor antimicrobial resistance across human, animal and environmental sources. A grounded natural-language layer could help more analysts explore harmonised data, provided provenance and confidence remain visible. AWS’s design is valuable because the model sits above a queryable evidence base and is required to admit when that base cannot answer the question.
Data governance crosses technical and scientific boundaries
Public sequence data still carries provenance and licensing obligations. The architecture records isolate metadata alongside detected genes, which is important when sources have different reuse terms or sampling biases. Researchers should avoid interpreting a trend as population prevalence when the underlying submissions are uneven across places, hosts or time. The agent can make data easier to query, but it cannot repair an unrepresentative dataset.
Security also matters even without patient records. A public surveillance system could be misused to overstate a local threat or generate unsupported claims about an organism. Tool limits, clear dataset descriptions and response citations help readers check the result. Interfaces should display the query scope and sample count so a concise conversational answer does not hide the size or composition of the evidence behind it.
Researchers should also be able to export the exact query and rows used for an answer. Conversational convenience is valuable, but scientific collaboration still depends on shareable methods and reproducible tables. An export path lets another analyst check the result with conventional tools, compare it with a later dataset version and cite the analysis without relying on a transient chat transcript.