FireAntGA-ImmunoModel: An explainable bio-inspired hybrid framework for optimized immunodiagnostic classification

• Comprehensive UAV–FSO System Model: Developed a detailed model of a UAV-assisted Free-Space Optical (FSO) communication system that includes atmospheric turbulence, pointing errors , and angle-of-arrival (AoA) fluctuations. • Analytical Derivations: Derived closed-form and asymptotic expressions for the Probability of Outage to quantify link reliability under realistic atmospheric impairments. • Firefly Algorithm (FFA)-Based Optimization: Proposed a bio-inspired metaheuristic framework using the Firefly Algorithm to jointly optimize: • Number of source nodes (M) • UAV relay altitude/position (z) • Overall cost–performance trade-off under budget constraints • Monte Carlo Validation: Performed extensive Monte Carlo simulations confirming the analytical model accuracy and validating the optimization results. • Performance Gains: • 60% reduction in outage probability compared to non-optimized configurations • Faster convergence and better reliability than Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) • Achieved outage probability as low as 0.0021 at 20 dBm transmit power • Key Insights: • Increasing transmit power improves performance only up to ≈ 30 dBm (beyond which it saturates). • FFA effectively balances system cost, energy efficiency , and link reliability . • Demonstrated scalability and robustness for 6G aerial network applications . • Future Directions: Suggested extensions to hybrid RF/FSO, RIS-assisted UAV networks , and integration of reinforcement learning for real-time adaptive control. Drug-induced autoimmunity (DIA) remains difficult to screen reliably due to three persistent challenges: (i) severe class imbalance between DIA-positive and DIA-negative compounds, (ii) high-dimensional and highly collinear molecular descriptors, and (iii) the need for transparent, auditor-grade decision support in biomedical settings. While recent bio-inspired deep learning frameworks such as DeepAIPs-SFLA, pACPs-DNN, pNPs-CapsNet, and TargetCLP have achieved strong performance on peptide- and protein-sequence prediction tasks, their architectures assume feature modalities fundamentally different from small-molecule DIA data, often emphasize model depth over interpretability, and typically lack robustness and computational complexity reporting.To address these gaps, the proposed FireAntGA-ImmunoModel , a unified, explainable, bio-inspired pipeline is tailored for small-molecule DIA prediction. The framework integrates Firefly-guided SMOTE to generate biologically plausible minority-class samples, SHAP-informed Ant Colony Optimization to derive compact and discriminative subsets from 196 RDKit molecular descriptors, and a multi-objective Genetic Algorithm to jointly optimize Random Forest hyperparameters with respect to F1-score, ROC-AUC, and feature sparsity.Evaluated on the InterDIA benchmark, the proposed method achieves a ROC-AUC of 0.945 and an F1-score of 0.854 , outperforming classical descriptor-based feature selection approaches including PCA, mRMR, and L1-regularization, particularly in minority-class recall. In addition, the framework reports multi-seed performance statistics (mean, standard deviation, and best–worst outcomes) and stage-wise runtime analysis, ensuring methodological transparency and reproducibility.While peptide-centric deep learning models remain highly effective in sequence-based settings, FireAntGA-ImmunoModel demonstrates that a computationally efficient, interpretable, descriptor-based framework can achieve competitive performance for small-molecule DIA prediction. The proposed approach provides end-to-end explainability and is well suited for deployment in clinical and regulatory screening workflows. Code and protocols are publicly available at: https://github.com/Lig17/FireANTGAImmunoModel .