Calibration-aware genomic prediction of paratuberculosis serostatus in indigenous Turkish goats

Paratuberculosis, caused by Mycobacterium avium subsp. paratuberculosis (MAP), is a chronic, incurable enteritis of ruminants in which late-onset clinical signs preclude timely intervention, making early genomic stratification a key control lever. We investigated the genetic architecture underlying the host antibody response to MAP — measured as anti-MAP ELISA serostatus — using genome-wide SNP data from 474 animals representing seven indigenous Turkish goat breeds sampled across 11 provinces (237 seropositive, 237 seronegative; estimated true seroprevalence 14.1%). Fourteen architecturally diverse machine learning frameworks — encompassing regularized regression, GBLUP, kernel machines, tree ensembles, deep neural networks, and a meta-ensemble — were evaluated under repeated stratified cross-validation using both discriminative and calibration metrics. Mutual information–based preselection served a dual function: reducing computational dimensionality while prioritizing markers with detectable phenotypic association. Six regularized and kernel-based models converged at a statistically indistinguishable discrimination ceiling (mean AUC ≈ 0.982; range 0.002), a convergence pattern consistent with an additive component captured by the available SNP data. Critically, whereas ROC-AUC varied only 1.14-fold across all models, Brier score varied 4.85-fold (0.046–0.223), with tree ensembles exhibiting systematic miscalibration despite competitive discrimination — indicating that predicted probabilities vary substantially across model architectures and cannot be reduced to discrimination metrics alone when informing breeding or biosecurity decisions. Post-hoc offset recalibration across realistic field seroprevalence conditions preserved tier-level performance rankings across the tested prevalence scenarios. These findings indicate that genome-wide SNP variation encodes sufficient signal to support accurate and calibrated classification of anti-MAP ELISA serostatus in indigenous goat populations, and provide a calibration-aware framework potentially extensible to other complex disease traits in livestock.