NextGenAMR — a product by Vanguard Biotech Systems
From supported WGS inputs to reviewable AMR evidence
NextGenAMR coordinates a versioned bioinformatics workflow: it processes whole-genome sequencing of a bacterial isolate within its supported scope, organises resistance-related evidence, applies quality and scope controls, preserves provenance, and produces structured outputs for professional review.
Controlled preview · Validation in progress · Not a replacement for phenotypic AST or autonomous clinical decision-making.
Detecting genomic evidence is not, by itself, a clinical interpretation. The absence of a known marker does not prove susceptibility. Outputs must be reviewed together with the clinical context and the applicable tests.
Widely-used science, inside a governed system
External components
Established open bioinformatics tools perform specific tasks — read cleaning, taxonomic classification, alignment, assembly, gene prediction and resistance-determinant detection. NextGenAMR does not develop these tools; it integrates, runs and coordinates them.
The system layer developed by VBS
The VBS-developed layer decides how inputs are accepted and verified, which profiles and versions may run, which gates must pass, how heterogeneous outputs become structured evidence, how scope is enforced, how evidence relates to the antibiotic panel, how warnings and abstentions are produced, how provenance is preserved, and how a coherent, reviewable result is produced.
Inputs & preconditions
The input is whole-genome sequencing (paired-end FASTQ) of cultured bacterial isolates within the supported scope.
The supported input is short-read, paired-end whole-genome sequencing (FASTQ) of a single cultured bacterial isolate, within the published scope. It is not a primary clinical sample and not metagenomics.
- Operational species
- Escherichia coli
- Antibiotic panel
- 9 antibiotics
- Scope version
- 2026.07.0 · 2026-07-15
- A technically accepted format is not the same as an input inside the supported scope.
- An input inside scope is not guaranteed to reach a result without abstention — quality and evidence must also be sufficient.
- A readable file can be rejected before analysis if it does not meet the required contract.
This page never invites uploads. Files for an evaluation are exchanged only through a controlled channel agreed with VBS — never samples, sequences, identifiers or patient data through the website.
The current published operational scope is Escherichia coli, with a closed, species-specific panel of 9 antibiotics. NextGenAMR is species-specific by design and built to expand; additional species are in internal development and are not yet available.
The workflow, stage by stage
Seven editorial stages, derived from the real productized workflow. Some steps are conditional: the path depends on the read layout, the organism and the outcome of earlier checks.
01Controlled intake & preflight
- Purpose
- Accept only inputs and configurations the workflow can honestly process, and pin exactly what will run.
- Input
- A supported input plus the requested scope, submitted through a controlled channel — never the public website.
- Main operations
- Detect the read layout (paired-end / single-end), check integrity, resolve the species profile, and verify that the required reference and databases are present and version-pinned.
- External tools
- Snakemake
- Controls developed by VBS
- VBS defines the input contract, the reproducible workflow definition, the reference/database preflight (fail-closed if anything required is missing) and the run identity that binds everything that follows.
- Output
- A validated, version-pinned run ready to proceed — or a clear rejection.
- Warnings & failure modes
- A readable file can be rejected here if it does not meet the contract, is out of scope, or if a required database is not available.
02Read quality & preparation
- Purpose
- Remove low-quality and adapter sequence so downstream steps work on clean reads.
- Input
- The raw sequencing reads of the isolate.
- Main operations
- Trim and filter reads and produce read-quality metrics.
- External tools
- fastp
- Controls developed by VBS
- VBS fixes the parameters, preserves the quality metrics as evidence and interprets the quality gate that decides whether the run may continue.
- Output
- Cleaned reads plus read-quality metrics recorded in the run.
- Warnings & failure modes
- Insufficient read quality is flagged and can prevent a result from being emitted.
03Species identification & scope coherenceconditional
- Purpose
- Confirm which organism the sample really contains and whether it is inside the supported scope.
- Input
- The cleaned reads.
- Main operations
- Classify reads by taxon, re-estimate the dominant species, and — when host contamination is above a threshold — screen out human reads (a conditional step).
- Controls developed by VBS
- The tools produce classifications; VBS decides how they bind to the supported profile, checks that the taxonomic signals agree, and abstains where the species is unsupported, mismatched or of low confidence.
- Output
- A confirmed species decision with its confidence, or an abstention.
- Warnings & failure modes
- Species mismatch, low confidence, missing or conflicting taxonomic evidence each lead to a safe abstention rather than a forced result.
04Reference alignment & coverage
- Purpose
- Place the reads on the species reference genome and measure how well the genome is covered.
- Input
- The cleaned (and, if applicable, host-screened) reads and the pre-staged species reference.
- Main operations
- Index the reference, align the reads, and compute coverage and alignment-quality metrics.
- Controls developed by VBS
- VBS selects the version-pinned reference (staged in advance, with no runtime internet access), records coverage metrics and carries them into the run's provenance.
- Output
- An aligned genome with coverage and quality metrics.
- Warnings & failure modes
- Low coverage or poor alignment quality is recorded and propagates into later quality decisions.
05Assembly & genome quality
- Purpose
- Reconstruct a draft genome and check it is complete and uncontaminated enough to trust downstream evidence.
- Input
- The reads recovered from the alignment.
- Main operations
- Assemble the reads into contigs and estimate assembly completeness and contamination.
- Controls developed by VBS
- VBS runs assembly in single-isolate mode, applies a genome-quality gate (completeness and contamination) and propagates any limitation to later stages — it does not treat a quality metric as a clinical guarantee.
- Output
- A draft genome plus completeness and contamination metrics.
- Warnings & failure modes
- A genome that fails the quality gate can suppress the antibiotic panel for that run.
06Gene prediction & AMR-evidence detection
- Purpose
- Turn the genome into structured evidence: genes, functional annotation and known resistance determinants.
- Input
- The draft genome.
- Main operations
- Predict genes, assign functional annotation, characterise genes not otherwise resolved, and detect known AMR determinants and point mutations.
- External tools
- PyrodigalSma3sDIAMONDAMRFinderPlus
- Controls developed by VBS
- VBS normalises the heterogeneous tool outputs into consistent evidence records with provenance — each detection stays a piece of evidence, never an automatic clinical conclusion.
- Output
- Normalised evidence records: annotated genes, unresolved genes and AMR determinants/mutations.
- Warnings & failure modes
- A detected marker is evidence associated with resistance — its meaning still depends on organism, antibiotic, data quality and review.
07Aggregation, modelled interpretation & reviewable output
- Purpose
- Combine the evidence into calibrated, reviewable output within the supported panel — or abstain.
- Input
- The normalised evidence records.
- Main operations
- Build the per-antibiotic feature representation, run the governed model layer, apply scope and quality rules, and assemble the result contract, report and provenance.
- External tools
- VBS-developed layer
- Controls developed by VBS
- This is the VBS-developed layer: a six-model calibrated ensemble per antibiotic, scope and panel mapping, warnings and abstention, and a single validated result contract that forbids categorical S/I/R — plus the provenance and integrity that make it reviewable.
- Output
- For each antibiotic in scope, a calibrated probability of resistance; structured JSON + PDF + run provenance for professional review.
- Warnings & failure modes
- The output is never a categorical S/I/R result and never a treatment recommendation; where conditions are not met the system abstains.
External components
The third-party tools NextGenAMR integrates, each with its function and the stage where it takes part. They are integrated and coordinated — not developed — by VBS. Their outputs are preserved or transformed under control, and later decisions are made on normalised contracts.
Exact tool and database versions are pinned per run and recorded in each result's provenance; here we show the public workflow version. v2.0.0
| Tool | Function | Project | Runs |
|---|---|---|---|
| SnakemakeMIT | Workflow engine that runs the stages in a defined, resumable order. VBS authors the governed workflow definition on top of it; the engine executes it. | Snakemake community | Every run |
| fastpMIT | Cleans and trims sequencing reads and reports read-quality metrics. | OpenGene | Every run |
| Kraken2MIT | Assigns reads to taxa to identify which organism the sample contains. | Derrick Wood / Johns Hopkins | Every run |
| BrackenGPL-3.0 | Re-estimates species-level abundance from the Kraken2 report to find the dominant species. | Jennifer Lu et al. | Every run |
| minimap2MIT | Aligns reads against a human reference so host reads can be screened out when present above a threshold (conditional). | Heng Li | Conditional |
| SAMtoolsMIT/Expat | Supporting utility to sort, index and convert alignment files used across the workflow. | SAMtools / htslib | Every run |
| BWA-MEM2MIT | Aligns the cleaned reads to the species reference genome. | bwa-mem2 | Every run |
| QualimapGPL-2.0 | Computes alignment and coverage quality metrics for the mapped genome. | Qualimap (CIPF/Conesa lab) | Every run |
| SPAdesGPL-2.0 | Assembles the reads into contigs (a draft genome) in single-isolate mode. | SPAdes (ABLab) | Every run |
| CheckM2GPL-3.0 | Estimates assembly completeness and contamination as a genome-quality gate. | CheckM2 (Chklovski et al.) | Every run |
| PyrodigalGPL-3.0 | Predicts protein-coding genes from the assembled genome. | Martin Larralde (Prodigal) | Every run |
| Sma3sAcademic use — see project | Assigns functional annotation to the predicted proteins. | Sma3s (Universidad Pablo de Olavide) | Every run |
| DIAMONDGPL-3.0 | Fast protein sequence search, used to characterise genes not resolved by functional annotation. | Benjamin Buchfink | Every run |
| AMRFinderPlusNCBI — public domain | Detects known antimicrobial-resistance determinants and point mutations from the predicted proteins. | NCBI | Every run |
The system layer developed by VBS
Vanguard Biotech Systems develops NextGenAMR, a bioinformatics software platform that processes supported WGS data from bacterial isolates within its published scope, organises genomic evidence related to antimicrobial resistance, preserves run-level provenance and generates outputs for professional review. NextGenAMR is currently undergoing controlled evaluation and validation. It is not presented as an autonomous diagnostic system, a replacement for phenotypic AST or a tool that selects treatments.
The value NextGenAMR adds is not a single algorithm; it is a coordinated set of contracts, controls and decisions that turn heterogeneous tools and outputs into a reproducible, reviewable execution. It does not remove error, guarantee accuracy or replace human review.
Scientific execution
- Input contract & intake checks
Defines what a valid input is and rejects anything that does not meet it before any analysis runs.
- Reproducible, versioned workflow
A governed workflow definition on top of the engine, run in isolation per execution, so the same input and versions reproduce the same steps.
- Component & database pinning
Selects and version-pins tools, reference genome and databases in advance, with no runtime internet access.
- Preflight verification
Verifies that required databases are present and integrity-checked before a run proceeds; a reduced database set cannot be presented as the full one.
- Quality gates
Read-quality and genome-quality gates that can flag a run or stop it from producing a panel.
- Scope control
Resolves the species profile and only lets the supported scope proceed; out-of-scope inputs abstain.
- Evidence normalisation
Maps heterogeneous tool outputs into consistent, provenance-carrying evidence records.
Product governance
- Panel & context mapping
Relates evidence to the supported, versioned antibiotic panel for the species.
- Governed model ensemble
A six-model calibrated ensemble per antibiotic, versioned and stamped into the result — its internal design is not published.
- Warnings & safe abstention
Emits designed warnings and abstains — emitting no call — across distinct causes rather than forcing an answer.
- Provenance & traceability
Binds tool, database and model versions, hashes, timestamps, run id, quality and scope decisions to each result.
- Validated result contract
A single validated result contract that refuses to emit a panel on an abstained run and forbids categorical S/I/R.
- Reviewable, integrity-checked outputs
Consistent JSON and PDF plus provenance, with a content hash so an independent party can confirm the report matches exactly one result.
Models, ensembles & abstention
For each antibiotic in scope, the model output is an estimate of the calibrated probability of resistance — never a categorical susceptible/intermediate/resistant (S/I/R) result.
For each antibiotic in scope, the governed layer combines several modelled signals into a calibrated probability of resistance. The internal design — the model families, parameters, features, thresholds and calibration — is not published. The system applies scope and quality rules before emitting anything.
Per antibiotic: a 6-model calibrated ensemble.
The states a run can end in
Abstention is designed behaviour, distinct from a warning, a technical failure and an out-of-scope input.
The system abstains safely — it emits no call — where a species or sample is outside its published scope, or where input quality or evidence is insufficient.
- Result
A calibrated probability of resistance is emitted for each antibiotic in the supported panel.
- Abstention
No call is emitted because the species or sample is outside the published scope, mismatched, or of low confidence. This is designed behaviour, not a hidden error.
- Not yet available
A candidate species is recognised but not yet offered because its model components are not complete.
- Quality-flagged
The run is flagged and the panel can be suppressed when read or genome quality is insufficient.
- Warning
A non-blocking flag recorded alongside a result to inform the professional review.
- Technical failure
A technical failure is reported as such; no partial or unverifiable result is served.
How the evidence connects
A flow of evidence and relationships — not clinical causality. A detected marker does not lead directly and inevitably to a treatment decision.
01Genomic input
Supported whole-genome sequencing of a cultured bacterial isolate, within scope.
02Detected evidence
Genes, resistance determinants, mutations and signals detected in the genome.
03Antibiotic context
The supported antibiotic panel each piece of evidence is related to.
04Aggregated analysis
Normalised evidence combined by the governed model layer into calibrated probabilities.
05Review output
Structured outputs and provenance prepared for professional review — not a clinical decision.
The graph shows flow and relationships, not clinical causality.
Genomic evidence vs clinical interpretation
- A detected gene, variant or signal is evidence that may be associated with resistance. Its meaning depends on the organism, antibiotic, mechanism, data quality, scope and available knowledge.
- The absence of a known signal does not automatically mean susceptibility. Evidence can also be discordant or insufficient — and NextGenAMR may warn or abstain.
- Results must be reviewed together with other tests, including phenotypic AST where applicable, by qualified professionals in the appropriate context.
NextGenAMR does not replace phenotypic antimicrobial susceptibility testing (AST). Genomic outputs must be interpreted together with AST, microbiological context and applicable procedures.
What comes out
Each run preserves provenance and execution metadata and produces structured, auditable outputs for professional review (JSON and PDF).
- Calibrated resistance probability: A per-antibiotic calibrated probability of resistance (0–1). Never a categorical S/I/R result and never a therapeutic recommendation.
- Structured JSON report: Machine-readable per-antibiotic results with evidence and provenance.
- PDF report: Human-readable report carrying the internal-validation-only and not-for-clinical-use notices and a report hash.
- Run provenance: Pipeline and database versions plus a report hash, for auditability and reproducibility.
Each run preserves provenance — tool and database versions, hashes and a report content hash — so an independent reviewer can confirm that a report corresponds to exactly one result. The PDF and interface are for professional review; they are not a definitive clinical report.
Security & deployment →Status & validation
NextGenAMR is a functional platform in controlled preview. Validation is ongoing, and the product is not presented as an autonomous diagnostic or treatment-decision system.
NextGenAMR is currently undergoing controlled evaluation and internal, non-clinical validation. Validation results have not been published and must not be inferred.
No clinical performance metrics (sensitivity, specificity, categorical agreement, VME/ME, accuracy) have been published. Performance evaluation is in progress.
Evidence & validation status →