Development and Evaluation of a Standardized QSAR Workflow for Predicting Small-Molecule Bioactivity Across Antibacterial, Antifungal, and Antiviral Targets

The increasing prevalence of antimicrobial resistance and emerging viral pathogens continues to challenge conventional drug discovery approaches, necessitating scalable and generalizable computational frameworks capable of screening large chemical spaces across multiple therapeutic targets. This study presents a unified quantitative structure-activity relationship (QSAR) workflow to predict bioactivity of small molecules across antibacterial, antifungal, and antiviral systems. The workflow integrates descriptor calculation, correlation-based feature filtering, multi-algorithm machine learning modelling, internal and external validation, Y-randomisation testing, and applicability domain (AD) analysis within a standardised pipeline. Three biologically distinct targets (DNA gyrase subunit B (GyrB), lanosterol 14α-demethylase (CYP51), and SARS-CoV-2 main protease (Mpro)) were employed to evaluate model transferability. The workflow was applied to a large natural product dataset comprising 738,453 coconut-derived compounds, enabling high-throughput prediction under realistic screening conditions. Model performance varied across targets: the GyrB model demonstrated strong predictive reliability (R² = 0.834; RMSE = 0.262), CYP51 showed moderate performance, and Mpro exhibited limited predictive capability due to dataset heterogeneity and insufficient descriptor representation. Applicability domain analysis revealed high internal consistency for all models, although large-scale predictions highlighted substantial variability in chemical space coverage. Higher-confidence predictions were obtained primarily for the antibacterial target, whereas many antifungal and antiviral predictions fell outside the applicability domain and should therefore be regarded as exploratory predictions rather than reliable predictions. The antiviral model also exhibited limited validation performance, further reducing confidence in its large-scale screening results. The findings demonstrate that a standardised QSAR workflow can support transferable modelling strategies across multiple biological systems, although predictive performance remained strongly dependent on dataset quality, descriptor relevance, and target-specific structure-activity coherence. In addition, the workflow was implemented as a reproducible web-accessible software framework using the Hugging Face platform to support transparent and scalable QSAR-based virtual screening. This study provides a scalable computational framework for evaluating QSAR-driven virtual screening across multiple biological systems under a unified modelling strategy. It highlights the importance of dataset quality, descriptor relevance, and applicability-domain constraints for ensuring predictive reliability across diverse biological systems.