iAMP-deep: an alignment-free AI framework for insect antimicrobial peptide discovery

Insect antimicrobial peptides (iAMPs) form a central component of innate immunity and represent promising templates for developing next-generation antimicrobials. However, conventional AMP identification approaches rely heavily on sequence alignment and motif similarity, which may overlook divergent, short, or noncanonical peptides. Existing insect AMP databases remain taxonomically limited, and high-throughput transcriptomic pipelines frequently generate large numbers of speculative candidates. To overcome these constraints, we developed iAMP-Deep, an alignment-free, multilayered artificial intelligence framework that integrates physicochemical descriptors, motif/domain signatures, and protein-language model (PLM) embeddings and employs ensemble learning and anomaly-aware post-processing for insect-derived AMP discovery. We compiled a curated dataset of experimentally validated insect AMPs and length-matched non-AMPs, including source accession numbers and family labels. To reduce homology-driven leakage, sequence clustering at ≥80% identity was incorporated before data partitioning. Four complementary models, viz ., LightGBM, deep neural networks (DNN), deep convolutional neural networks (DCNN), and transformer neural networks (TNN), were trained with fixed random seeds and documented hyperparameters. DCNN exhibited the strongest individual performance (AUC = 0.998), while the ensemble model achieved high precision (0.97), recall (0.96), and an F1 score of 0.96, misclassifying only six AMPs and 23 non-AMPs. Physicochemical profiling showed that AMPs are enriched in cationic, amphipathic, and moderately hydrophobic residues, forming a distinct feature space relative to non-AMPs. Post-processing via Isolation Forest and AMP-specific physicochemical constraints further refined prediction confidence. Application of iAMP-Deep to the Apis dorsata proteome identified 15 high-confidence candidates that contain 13 unique peptides. These candidates exhibited AMP-like properties, such as positive charge ( + 2.50 to +12.06), hydrophobicity ratio (0.302–0.500), amphipathicity scores (1.321–2.243), and instability indices below 45, and conserved signatures including cecropin-, defensin-, and melittin-like motifs. Functional enrichment revealed that 82% of candidates were associated with immune defense and secretory pathways. The candidate set included both motif-positive peptides and motif-poor peptides with strong physicochemical/PLM support and low global similarity to reference AMPs, supporting discovery beyond canonical motifs. On the balanced independent test set, the final anomaly-filtered pipeline achieved precision = 0.959, recall = 0.942, and F1 = 0.950, whereas conventional baselines yielded F1 values of 0.012–0.667. Applying the unchanged pipeline to 13 proteomes spanning Hemiptera, Lepidoptera, Diptera, and Coleoptera yielded 416 high-confidence entries representing 237 unique peptide sequences; together with the A. dorsata (Hymenoptera) case study, these external screens span five insect orders and consistently recovered noncanonical, lysine-rich AMP-like candidates with high ensemble confidence. iAMP-Deep provides an alignment-free, multimodal, and interpretable framework for insect AMP discovery by integrating curated insect AMP knowledge with physicochemical descriptors, motif/domain evidence, PLM embeddings, ensemble learning, and anomaly filtering. Compared with motif-only and alignment-based baselines, the framework is designed to retain sensitivity for divergent candidates while reducing false positives and prioritizing experimentally testable peptides.