This is likely a curation-scope issue rather than a detection bug. RGI/CARD only returns a hit when your sequence matches a curated ARO reference model above the model's specific identity/bitscore cutoff — it's not testing "does this gene exist," it's testing "does this gene match a known resistance-conferring variant CARD has explicitly curated."

Prokka, on the other hand, is annotating against general-purpose databases (Pfam/UniProt-style HMMs), so it'll call any lpxD homolog by gene family/function regardless of whether that specific copy carries a resistance-relevant mutation.

lpxD is part of lipid A biosynthesis, and its link to AMR (colistin/polymyxin resistance in some Enterobacteriaceae/Acinetobacter) is usually tied to specific resistance-conferring mutations or operon-level changes, not the wild-type gene itself. If your copy doesn't carry the exact variant CARD curated, RGI correctly returns no hit — even though the gene is genuinely present, which is exactly what Prokka is telling you.

A couple things worth checking:

1. Browse the ARO ontology directly on the CARD site to confirm whether lpxD is represented as a model at all, and if so, whether it's a full protein homolog model or a specific SNP/variant model.
2. Check whether you're running RGI with loose hits excluded (default) — loose hits are lower-confidence homology matches that are hidden unless you explicitly include them, and might surface a partial match you're currently not seeing.
3. Confirm you're querying the correct reading frame/strand if using the nucleotide BLAST option — easy to get a false negative from a frameshift or reverse-complement mismatch.

TL;DR: CARD and Prokka are answering different questions (curated resistance variant vs. general gene family homology), so disagreement here is expected, not a bug.