The Evidence Layer for Scientific AI

 Quick Analysis Plan Copied

 Long Analysis Plan

Evidence gap: what the records can and cannot support

The supplied claim-level records contain zero livestock in vivo AMP efficacy or resistance-rate outcomes. Three records are adjacent but insufficient: (1) the cecropin-like peptide CLP1 was purified from H. illucens larvae and characterized with MIC/MBC 0.52–2.07 µM against Gram-negative bacteria only, with no animal data CLP1 antimicrobial characterization; (2) a randomized piglet trial fed 0/2.5/5% H. illucens larvae meal (n=8/group, 35 days) is genuinely in vivo livestock, but its outcomes are fatty-acid profiles, hepatic collagen, and gene expression — no pathogen-challenge or resistance endpoints Piglet HI-meal trial without infection endpoints; (3) comparative genomics documents expansion of immune genes (CecA1, PGLYRP3, GNBP3 orthogroups; CYP450 cluster of 66 copies), which is mechanistic context, not efficacy evidence BSF immune gene family expansion. BGPT inference: any published meta-analysis on this exact question would need studies not present here; proceeding to pooled effect estimates now would be fabrication.

Reproducible meta-analysis plan (ready to execute once studies are retrieved)

Step 1 — Protocol and search: pre-register (PROSPERO), define PICO (livestock species × H. illucens-derived AMP or AMP-rich fractions × in vivo challenge × efficacy/resistance outcomes), search PubMed/Scopus/WOS/CNKI; PRISMA 2020 flow; dual independent screening with kappa reporting.

Step 2 — Evidence extraction: encode each record as claim-level units (the schema above), capturing n per group, challenge organism/dose, AMP identity/dose route, mortality or shedding effect sizes (risk ratio, Hedges g), and resistance endpoints (MIC shift over serial passage, resistance-frequency per 10⁶ CFU).

Step 3 — Statistical engine (Python 3): snakemake pipeline; statsmodels/scipy for random-effects (DerSimonian–Laird + Paule–Mandel), I² and prediction intervals; PyMC for Bayesian hierarchical models when k is small; subgroup meta-regression by livestock species (poultry, pig, fish), AMP class (cecropin vs defensin vs lauric-acid fraction), and route (feed vs purified peptide). Funnel plots + Egger/p-curve for publication bias; robvis (RoB 2) for risk-of-bias weighting; leave-one-out sensitivity.

Step 4 — Resistance-rate sub-analysis: treat resistance emergence as a time-to-event or dose–response endpoint; distinguish de novo resistance under AMP pressure from cross-resistance selection — a distinction the CLP1 record cannot address since it tested only naive planktonic MICs CLP1 study limitations.

Blindspots to counteract: insect-industry sponsor bias in feed trials; small-n challenge studies (the piglet trial had n=8/group and explicitly low power for some endpoints Piglet trial power limitations); conflation of lauric-acid content with peptide AMP activity; and publication bias favoring positive antimicrobial results. What would change the conclusion: a single randomized livestock challenge study reporting AMP efficacy or resistance emergence would convert this plan from empty-evidence to executable.