Design Experiments From Reported Results

 Quick Analysis Plan Copied

 Long Analysis Plan

Evidence Foundation: Why Single-Marker Alerts Fail

The supplied records demonstrate that canonical mammalian-adaptation markers are absent from most circulating 2.3.4.4b isolates, yet these viruses still replicate in novel hosts. Bangladesh 2.3.4.4b duck isolates lack PB2 E627K, PB1-F2 N66S, and PA T97I Bangladesh duck H5N1 mutational analysis. Australian sub-Antarctic isolates similarly retain avian-like PB2 E627 and D701 Australian seabird H5N1 molecular markers. Meanwhile, the bovine Ohio H5 HA (H5OH) binds 3′SLN at KD 18.55 nM versus 36.45 nM (clade 1 H5VN) and 46.95 nM (clade 2.3.4.4b H5HK), suggesting that receptor affinity varies within the clade and contributes to host adaptation independently of canonical polymerase mutations H5N1 clade HA receptor binding across tissues. These findings directly motivate a constellation scoring approach that integrates signals across all gene segments.

Figure 1. Reported biolayer interferometry KD values for 3′SLN binding by recombinant H5 HAs [Source: 10.1128/jvi.00976-25]. The bovine-isolate HA binds strongest, but the authors explicitly caution that attachment does not equal replication efficiency.

Phase 1: Standardized Isolate Panel + Whole-Genome Sequencing Pipeline

Panel construction (n≥30 isolates): Stratify by host (bovine mammary, human spillover, avian/wild bird, feline, marine mammal), genotype (prioritize B3.2/E.2 and related sub-Antarctic lineages alongside North American B3.2), and sampling date (2024–2026). Include the two Western Australian genomes (EPI_ISL_20479340, EPI_ISL_20497638) as sub-Antarctic representatives GISAID accessions for WA seabird H5N1.

Sequencing workflow: Multiplex one-step RT-PCR across all eight segments, followed by Oxford Nanopore PromethION/MinION sequencing (minimum 10× coverage, typically ~10,000×) as validated in Antarctic surveillance Antarctic H5N1 nanopore sequencing protocol, plus Nanopore Rapid Barcoding (SQK-RBK114.24) with R10 flow cells and IRMA/Minimap2 consensus generation as used for the Australian isolates Australian H5N1 sequencing tools and pipeline.

Genomic annotation pipeline (Python3): MAFFT alignment → GenoFLU + ggFLU genotype assignment → FluMut molecular marker screening → per-gene mutation matrix extraction against a reference backbone → phylogenetic placement (IQ-TREE v3.1.3 ML, BEAST X v10.5.0 time-calibrated Bayesian for selected nodes) Phylogeographic pipeline for H5N1 constellation tracking.

Phase 2: Constellation Scoring Framework

Define a constellation score as a weighted composite across ≥5 gene segments, each contributing domain-specific binary or ordinal features:

Score construction: (a) Generate a binary mutation matrix M[isolate × feature]; (b) Apply elastic-net logistic regression or gradient-boosted trees (scikit-learn) with leave-one-isolate-out cross-validation to predict each phenotype; (c) The constellation score is the predicted probability from the fitted multi-segment model; (d) Benchmark against single-marker alerts (PB2 E627K alone, HA Q226L alone, etc.) using DeLong test for paired AUC comparison.

Phase 3: Matched Phenotypic Assays

Bovine udder explants: Lactating dairy cow mammary gland/teat tissue; infect at matched MOI; measure replication (RT-qPCR M-gene, TCID50 on MDCK-SIAT1) at 6/12/24/48/72 hpi. The HA attachment study established that H5 binds mammary alveolar and cistern (not ductal) epithelium and that both α2,3 (MAL-II: NeuAcα2,3Galβ1,3GalNAc) and α2,6 (SNA) sialoglycans are present Bovine mammary tissue tropism for H5N1 HA.

Human airway epithelial organoids: Differentiated primary HAE at air-liquid interface; matched MOI inoculation; identical time-course and readouts. Add interferon-β pretreatment arm (0 vs 100 IU/mL, 16 h) to isolate IFN-antagonism contribution.

IFN antagonism readouts: (i) ISG54/ISG56/IFIT1 qRT-PCR suppression ratio (infected/pretreated ÷ infected/untreated); (ii) luciferase reporter under ISRE promoter co-transfected with NS1 expression plasmid from each isolate; (iii) phospho-STAT1 Western blot.

Matched conditions: Same MOI (0.01 TCID50/cell), same medium (DMEM/F12 + TPCK-trypsin for explants; BEGM for organoids), same incubator (37°C, 5% CO₂), same time points, ≥3 biological replicates per isolate per tissue.

Phase 4: Statistical Hypothesis Test

Primary hypothesis (H₁): Constellation scores predict composite host-adaptation phenotype (normalized replication titer + IFN antagonism index) in both tissue models with AUC > 0.80, exceeding single-marker alert AUCs. Null (H₀): No significant AUC improvement of constellation over best single marker (DeLong p < 0.05 after Bonferroni correction for multiple single-marker comparisons).

Statistical plan: n≥30 isolates provides 80% power to detect AUC difference of 0.15 (α=0.05, paired design, DeLong). If single-marker AUC ≈ 0.60 and constellation ≈ 0.82, required n ≈ 25. Report exact AUC, 95% CI (bootstrap, 1000 resamples), calibration slope, and decision-curve analysis for surveillance threshold setting.

Counterpoints, Blindspots, and Known Unknowns

• The HA attachment study explicitly states that HA binding does not equal replication efficiency and that ecological factors likely drive dairy cattle outbreaks Limitations of HA binding for predicting replication — this is a fundamental constraint on any sequence-only predictor.
• Only two Western Australian and seven Bangladeshi genomes are available from these records; a 30-isolate panel requires drawing from the broader 1,324-genome reference set or GISAID EpiFlu 1324-genome reference panel for H5N1.
• The non-human mammalian adaptation study assumes constant exposure likelihood and cannot perform gain-of-function experiments Limitations of mammalian adaptation screen — sequence-only adaptation scores will miss host-level ecological/behavioral confounders.
• What would falsify the constellation approach: If isolates with identical constellation scores show >100-fold replication differences in matched explants/organoids, or if a single novel mutation emerges that alone flips phenotype regardless of background constellation, multi-gene scoring adds no surveillance value over that single alert.
• Hidden assumption: Constellation scoring implicitly assumes additive or epistatic-but-stationary mutation effects; if compensatory mutations shift with background genotype, the model will not generalize beyond the training panel without regular retraining.