Confidence-graded, proteome-wide identification of candidate antimicrobial peptide precursors in the Ghaf tree (Prosopis cineraria) using a preregistered, decoy-calibrated pipeline

Background. Antimicrobial peptides are a candidate source of agents against resistant pathogens, and plant proteomes are routinely mined for them. Such surveys commonly report candidates without a control establishing the classifier’s background positive rate, without calibrated significance for short-peptide similarity, and without a statement of how strongly each candidate is supported. Prosopis cineraria , the Ghaf, is an arid-zone mimosoid legume with a sequenced genome, documented antimicrobial activity in crude extracts and one characterised antifungal seed protein, but its antimicrobial peptide precursor complement has not been surveyed systematically. Methods and results. We searched the annotated Ghaf proteome under a protocol registered before any sequence data were retrieved (DOI 10.17605/OSF.IO/RG9HA). Quality control and isoform reduction gave 29,222 primary proteins. Five prespecified evidence axes were applied: a machine-learning classifier, homology to curated reference peptides and precursors, antimicrobial peptide family domain signatures, secretion signals requiring agreement between two prediction tools, and an independent prediction applied to mature regions inferred under registered conditions. Short-peptide similarity was assigned empirical significance against 1,000 composition-preserving decoys per reference peptide rather than a fixed E-value threshold; 418 of 1,803 tested pairs survived calibration, with acceptance ranging from 8 of 86 pairs at 10–19 residues to 109 of 160 at 61–120 residues. Candidates were assigned to confidence tiers by an ordered decision rule fixed in advance. Of 489 shortlisted proteins, 18 reached the highest tier under both the registered rule and a stricter sensitivity analysis: six defensins with a knottin scaffold, five snakins or GASA proteins, four non-specific lipid transfer proteins, two knottins and one defensin. These candidates are short, cysteine-rich and predicted to be secreted, with a median full-precursor cysteine content of 10.1% against 3.7% in the next tier. Conclusions. Eighteen candidate antimicrobial peptide precursors in the Ghaf proteome are supported by convergent evidence from prespecified criteria with calibrated homology significance. All are computational predictions and none has been shown to possess antimicrobial activity. We also report two control results of general relevance: a composition-preserving null cannot test discrimination for a classifier built on compositional features, and a peptide classifier applied beyond its stated size range saturates in a way that is easily mistaken for a property of the tool.