Sequence and Physicochemical Associations with Antimicrobial Peptide Activity against Acinetobacter baumannii: A Homology-Aware Regression and Explainable Machine-Learning Study

Background: Carbapenem-resistant Acinetobacter baumannii is a critical-priority pathogen, but the available database records were predominantly not annotated for carbapenem resistance. We evaluated associations between interpretable peptide features and A. baumannii minimum inhibitory concentration (MIC), and tested transfer across sequence-disjoint clusters.
Methods: From 1,740 raw records, 809 exact or reported-mean MIC observations were eligible and were aggregated into 525 conservative chemical entities. The outcome was entity-level median log2 MIC in micromolar units. Twelve prespecified sequence and physicochemical features were analyzed by multiple regression with HC3 and 80%-sequence-cluster-robust uncertainty estimates. Peptides sharing at least 80% normalized global identity were kept within the same development or test partition. A median baseline, Elastic Net, Random Forest, and Extra Trees were compared by nested grouped cross-validation; the locked Extra Trees model was evaluated in 104 test entities. Descriptor, assay-metadata, structural, censoring, alignment, homology-threshold, and similarity-domain sensitivities were examined.
Results: Only 19 of 809 observations (2.3%; 14 entities) were explicitly labeled carbapenem-resistant, and 578 (71.4%) lacked a resistance label. Seven features had nominal HC3 P values below 0.05, including negative associations of MIC with charge, mean hydrophobicity, beta-strand hydrophobic moment, cysteine fraction, and C-terminal amidation. Cluster-robust and leave-one-cluster-out analyses reduced precision and showed that the cysteine and N-terminal acetylation findings were concentrated in particular sequence families. Extra Trees showed weak development-set transfer (pooled out-of-fold R²=0.054; mean outer-fold RMSE=2.174, versus 2.236 for the median baseline). In the test set, RMSE was 1.775 log2 units, R² was 0.381 (95% cluster-bootstrap CI -0.024 to 0.543), and Spearman rho was 0.579. Performance weakened when more remote families were separated and in the named-strain-only subset.
Conclusions: The data support hypothesis-generating associations rather than causal feature effects. Predictive transfer to unseen peptide clusters was weak during grouped cross-validation, and the more favorable single test estimate was uncertain. The model is not validated as a stand-alone prioritization method and does not support carbapenem-resistant-specific inference; standardized prospective MIC testing, toxicity and stability assessment, and external validation remain necessary.
Keywords: Acinetobacter baumannii; antimicrobial peptides; minimum inhibitory concentration; homology-aware validation; machine learning; sequence clustering.