Navigating the Eukaryotic Paradox: Cross-Kingdom Insights for AI-Driven Clinical Repurposing Strategies Against Antifungal Resistance

1. Introduction

Global society has traversed unprecedented socio-technological and environmental transformations in recent decades. While human history has historically been characterized by gradual progression, the contemporary era, marked by technological integration, migratory flows, globalization, geopolitical instability, rising food insecurity, and climate change, has undergone a rapid acceleration. Although these developments have fostered remarkable medical advancements, they have simultaneously amplified global health vulnerabilities. In particular, these shifts have driven an increased incidence of complex chronic diseases, the re-emergence of infectious threats, widespread antimicrobial resistance (AMR), and the spillover of emerging or previously benign fungal and viral entities displaying heightened virulence and multidrug resistance [1]. Consequently, the timely identification and deployment of effective therapeutic strategies have become an imperative global health priority.

However, the traditional pipeline for de novo drug discovery is notoriously slow, prohibitively expensive, and fraught with high attrition rates. Bringing a single novel chemical entity to market typically requires over a decade and financial investments on the order of billions of dollars [2]. When confronting rapidly evolving drug-resistant pathogens, time represents a critical limiting factor. To bypass this bottleneck, Artificial Intelligence (AI) is transforming drug repurposing, a clinical and computational strategy that identifies novel therapeutic indications for previously approved or investigational molecules. Because these drugs possess well-characterized toxicological profiles, proven human tolerability, and documented pharmacokinetics, they can bypass early-stage clinical safety trials (Phase I), drastically reducing development timelines and associated financial expenditures [3]. Today, AI-driven computational methodologies are fundamentally redefining how we combat bacterial, viral, and fungal threats. While AI computational frameworks were initially benchmarked against bacterial and viral targets, applying these models to eukaryotic fungal pathogens introduces unprecedented biological and structural bottlenecks. In this review, we systematically evaluate AI-driven drug repurposing through a cross-kingdom comparative lens, explicitly contrasting bacterial and viral paradigms to highlight why medical mycology represents the ultimate challenge and frontier for computational pharmacology and clinical drug repurposing against antifungal resistance.

2. The Antibacterial Front: Combating Superbugs and Physical Permeation Barriers

Antimicrobial resistance (AMR) represents one of the most pressing public health crises of the twenty-first century. As bacteria rapidly evolve mechanisms to evade our current clinical arsenal, the pipeline for novel antibiotic classes has virtually dried up, threatening a return to a pre-antibiotic era [4]. Advanced machine learning (ML) and deep learning (DL) architectures are increasingly deployed to screen vast chemogenomic libraries of non-antibiotic drugs, synthetic compounds, and natural products to uncover hidden bactericidal or bacteriostatic properties [5]. A landmark example was the deployment of deep neural networks (Artificial Neural Networks, ANN) to identify halicin, a molecule originally investigated as an antidiabetic agent, as a potent broad-spectrum antibiotic [5,6]. ML models evaluate the molecular fingerprints of thousands of existing compounds to predict their binding affinity to conserved bacterial targets, such as those involved in cell wall biosynthesis or ribosomal translation. Furthermore, AI excels at identifying synergistic drug combinations: by pairing a conventional antibiotic with a repurposed non-antibiotic agent, AI can disarm resistance mechanisms, effectively restoring the antibiotic’s original potency.

Despite these computational successes, the transition from in silico prediction to clinical efficacy remains severely obstructed by data limitations, physical transport barriers, and translational hurdles. Machine learning models are fundamentally constrained by the fidelity of their training data. In contemporary antibiotic discovery, these datasets exhibit severe systemic limitations, chief among which is a pronounced publication bias toward positive outcomes. Because peer-reviewed literature overwhelmingly documents successful experimental results, decades of negative data, comprising molecules that demonstrated negligible or non-existent antimicrobial efficacy, remain largely unpublished or locked behind proprietary walls [7]. Deprived of these negative training exemplars, AI models suffer from elevated false-positive rates and fail to accurately map the chemical boundaries of non-efficacy. This problem is further compounded by inter-laboratory heterogeneity across historical screening data compiled over the past half-century. Widespread variations in assay protocols, bacterial strain selection, and minimum inhibitory concentration (MIC) endpoint definitions introduce profound noise into machine learning pipelines, ultimately compromising model generalizability across independent biological contexts.

The most critical clinical threats reside within Gram-negative pathogens, such as Acinetobacter baumannii and Pseudomonas aeruginosa, which present severe biophysical barriers to AI-driven drug discovery [8,9]. This challenge stems primarily from a marked disconnect between in silico predictions and in vivo efficacy. While an algorithm may successfully identify a candidate molecule that perfectly inhibits a vital bacterial target within a computational simulation, current models struggle to accurately predict outer membrane permeation kinetics across the asymmetrical, bilayered envelope characteristic of these microorganisms [8]. Furthermore, even when a compound successfully traverses this selective macromolecular barrier, computational pipelines frequently fail to anticipate whether the molecule will be immediately extruded by promiscuous bacterial efflux pumps prior to reaching its intracellular target [10].

Compounding these biophysical obstacles are regulatory, toxicological, and macroeconomic constraints that hinder real-world application. Deep learning architectures, particularly ANN, are notoriously opaque. Operating as “black boxes,” these computational models generate high-probability predictions for antimicrobial activity without elucidating the underlying molecular mechanisms of action [11]. For instance, an algorithm might predict with 95% statistical confidence that an established antihypertensive agent will exert bactericidal activity against a specific pathogen, yet remain incapable of explaining the structural or biochemical basis of this inhibition. This mechanistic opacity presents a significant hurdle for regulatory authorities such as the US FDA or EMA, which remain hesitant to grant approval, even for repurposed molecules, in the absence of a defined mechanism. This regulatory hurdle is further complicated by the safety paradox of escalated dosing: while a repurposed candidate may possess a well-documented clinical safety profile at its approved therapeutic indication, clearing a systemic, acute bacterial infection often requires serum concentrations far exceeding original physiological baselines. Current AI models frequently fail to anticipate the off-target toxicity and organ strain triggered by such high-dose regimens, highlighting a critical gap between computational predictions of efficacy and real-world clinical viability. Finally, market failures severely disincentivize antibiotic commercialization, as antimicrobial stewardship frameworks mandate that effective new agents be held in reserve, limiting sales volumes and preventing developers from recouping clinical trial expenditures [12]. While Gram-negative bacterial envelopes present complex permeation barriers, the computational challenges escalate significantly when transitioning from prokaryotic cell walls to eukaryotic fungal structures.

3. The Antiviral Front: Rapid Pandemic Response and Host Physiology

Unlike bacteria, viruses are obligate intracellular parasites with minimal autonomous metabolic machinery, replicating entirely by hijacking host cellular pathways. Consequently, AI-driven antiviral discovery encounters distinct computational and biological hurdles. Viral outbreaks necessitate rapid containment strategies that traditional, de novo drug pipelines cannot accommodate [3]. AI-based computational repurposing has thus emerged as a cornerstone of rapid-response pharmacology. Using deep learning models, algorithms can rapidly evaluate the binding affinities of thousands of small molecules against predicted 3D structures of novel viral proteins within days [13]. Moreover, because viruses depend on the host, AI leverages network pharmacology to map host-pathogen interaction networks, prioritizing host-directed therapeutics that target host factors essential for viral replication [14]. By targeting host proteins rather than hyper-mutable viral enzymes, host-directed approaches establish a higher genetic barrier to resistance, remaining effective across divergent viral strains.

The application of artificial intelligence to antiviral drug discovery faces severe biological and computational constraints, primary among which is the disconnect between dynamic mutational landscapes and static structural modeling. RNA viruses, such as Influenza, HIV, and Coronaviruses, exhibit exceptionally rapid evolutionary kinetics [15]. Because AI models are conventionally trained on static crystallographic snapshots of viral proteins, even single amino acid substitutions within a target binding pocket can rapidly invalidate an AI-predicted ligand’s efficacy [13]. Furthermore, current computational frameworks possess limited fidelity in forecasting viral evolutionary trajectories or predicting prospective escape mutations under selective drug pressure.

This challenge is further compounded by the “host-dependence barrier” inherent to obligate intracellular parasites. Because viral replication is inextricably linked with host cell physiology, targeting viral pathways carries a high intrinsic risk of cellular disruption [14]. AI architectures optimized primarily for target binding affinity frequently lack the systemic resolution required to predict off-target proteomic disruptions and multi-organ toxicities resulting from unintended interactions with homeostatic human proteins [16,17].

These predictive limitations are exacerbated during novel outbreaks by the “Disease X” data scarcity paradox. Deep learning architectures rely heavily on extensive historical datasets; consequently, upon the emergence of a novel pathogen with pandemic potential, computational pipelines encounter an acute deficit of empirical data [3]. Direct extrapolation from established viral families remains highly uncertain due to subtle yet functionally consequential genomic variations.

Even when potential candidates are identified, pharmacokinetic and pharmacodynamic disconnects frequently hinder clinical translation. Antiviral efficacy typically hinges on administration within a narrow 24-to-72-hour physiological window post-infection, yet binding-focused AI models rarely incorporate these rigid in vivo time constraints. Moreover, phenotypic screens routinely produce a high frequency of in vitro false positives, flagging diverse chemical classes (e.g., antimalarials or antipsychotics) that suppress viral replication through non-specific pathways like lysosomal pH alteration [18]. Translating these in vitro effects in vivo frequently demands cytotoxic or physiologically unachievable serum concentrations.

Finally, unlike bacteria, which share conserved targets such as peptidoglycan cell walls, viruses display extreme structural diversity. The absence of universal viral target structures prevents the development of a single, unified antiviral AI architecture, requiring tailored pipelines for distinct viral families. Although host-directed antiviral strategies share conceptual similarities with targeting host-pathogen interactions, fungi present a far more complex challenge due to their autonomous eukaryotic cell biology and shared proteomic homology with human hosts.

4. The Antifungal Front: Addressing Eukaryotic Complexity, Biofilm Shielding, and Clinical Strategies Against Antifungal Resistance

The application of artificial intelligence to antifungal drug discovery presents the most demanding computational and biological challenges in modern pharmacology. While invasive fungal infections caused by opportunistic pathogens such as Candida auris, Aspergillus fumigatus, and Cryptococcus neoformans carry alarming mortality rates exceeding 50% in immunocompromised populations, medical mycology remains historically under-funded and under-researched compared to bacteriology and virology [19,20]. The central impediment to designing therapeutic strategies is fundamental to cell biology: fungi are eukaryotic organisms that share extensive genomic conservation, metabolic pathways, and structural homology with human host cells [21,22]. Consequently, identifying repurposed small molecules that selectively exert fungicidal activity without triggering catastrophic off-target host toxicity represents an enormous computational optimization challenge, leaving clinicians reliant on a critically narrow therapeutic arsenal dominated by just three primary classes: azoles, polyenes, and echinocandins [20,21].

This eukaryotic paradox severely complicates AI-driven screening pipelines. Because essential fungal enzymes (such as Hsp90 chaperones, calcineurin, or cytochrome P450 enzymes involved in sterol synthesis) closely mirror their human orthologs, virtual screening algorithms tasked with searching non-antifungal drug libraries encounter extreme selectivity barriers [23]. Standard molecular docking and deep learning scoring functions often fail to discriminate between subtle active site variations between fungal enzymes and their human homologues. As a result, candidate repurposed molecules identified in silico frequently exhibit severe cross-reactivity with human targets, translating into severe clinical hepatotoxicity, cardiotoxicity, or nephrotoxicity [23]. Historically, this clinical bottleneck was exacerbated by a severe scarcity of experimentally resolved target structures; fewer than 2% of proteins in key fungal pathogens had known 3D structures in public repositories. While the advent of AlphaFold has fundamentally transformed this landscape by computationally mapping this ‘structural dark matter’, expanding theoretical structural coverage to over 98% of fungal proteomes, significant challenges remain [24,25].

Although machine learning architectures are increasingly trained to detect subtle genomic, transcriptomic, and proteomic divergences between host and fungal targets, these algorithms are still hindered by acute deficits in experimentally validated structural data. Deep learning prediction models rely heavily on the high-resolution empirical structures within the Protein Data Bank (PDB) for training; yet, pathogenic fungi remain profoundly underrepresented relative to bacterial and viral entities [24,25]. Consequently, predicting dynamic conformational changes and complex drug-target interactions remains exceptionally difficult.

Furthermore, historical empirical data regarding fungal replication kinetics, membrane lipid composition (such as ergosterol vs. cholesterol dynamics), and phenotypic responses remain fragmented, unstandardized, and isolated across non-interoperable databases [21].

Beyond target homology, severe biophysical and structural barriers disconnect in silico binding predictions from in vivo therapeutic clearance. The fungal cell wall is a dynamic, highly cross-linked macromolecular matrix composed of chitin, beta-(1,3)-glucans, beta-(1,6)-glucans, and outer mannoproteins [26]. This rigid structure poses an immense mechanical and steric permeation barrier to high-molecular-weight compounds. While molecular docking algorithms successfully predict ligand binding kinetics at isolated receptor sites, current AI models lack the biophysical parameters needed to simulate molecular diffusion dynamics through this complex carbohydrate mesh.

This transport bottleneck is magnified tenfold when fungal pathogens transition into refractory biofilm communities attached to mucosal surfaces or medical devices, significantly driving clinical antifungal resistance [27]. Fungal biofilms synthesize an extracellular polymeric substance (EPS) matrix rich in exopolysaccharides, extracellular DNA, lipids, and proteins. To yield clinically actionable drug repurposing strategies, AI computational models must evolve beyond raw binding-affinity predictions to integrate physical diffusion dynamics across EPS matrices.

This matrix acts as a physical shield, trapping repurposed drug molecules via charge interactions or enzymatic degradation before they can reach cell-surface targets. Furthermore, cells deep within the biofilm matrix enter a metabolically dormant “persister” state, rendering target-binding algorithms ineffective since these persister cells downregulate metabolic targets targeted by standard AI models. Standard high-throughput virtual screening models evaluate isolated unicellular states in suspension, completely failing to account for physical shielding, persister cell physiology, and altered drug accumulation within resilient biofilm architecture.

Compounding these structural challenges are dynamic morphogenetic transitions and an excessively restricted baseline chemical space. Many lethal fungal species exhibit morphological dimorphism, switching dynamically between unicellular yeast phases, pseudohyphae, and invasive, filamentous hyphal structures in response to host environmental cues such as body temperature, ambient CO₂ concentration, or local pH shifts [28]. This morphological transition fundamentally reshapes the surface proteome, cell wall architecture, and transcriptomic landscape of the fungus. An AI model trained on static yeast-phase data will predict ligand efficacy against targets that may be completely downregulated or physically masked during invasive hyphal growth in tissue. Finally, because systemic antifungal therapy relies almost entirely on azoles, polyenes, and echinocandins, the baseline training datasets for machine learning models lack chemical diversity. This narrow structural baseline deprives algorithms of broad “chemical intuition,” making non-antifungal drug repurposing predictions especially prone to false positives and biological failure during laboratory validation.

5. Methodological Pillars of AI Integration

Despite these domain-specific biological and computational constraints, the transformative potential of artificial intelligence in cross-disciplinary drug repurposing is anchored in three primary methodological paradigms. These technological pillars collectively enable researchers to bypass conventional, empirical discovery routes by systematically interrogating pathogen vulnerability, host response dynamics, and the broader biomedical literature.

The first fundamental pillar unites structural biology with deep learning architectures. Deep neural network frameworks, exemplified by milestone developments, such as AlphaFold, can predict the tertiary and quaternary 3D structures of pathogen proteins with near-atomic precision directly from primary amino acid sequences [29]. This computational breakthrough overcomes historical bottlenecks caused by the slow, labor-intensive process of crystallizing difficult-to-resolve target proteins. By generating high-fidelity structural models for uncharacterized pathogen targets, these systems facilitate high-throughput virtual screening and molecular docking simulations. Algorithms rapidly evaluate millions of existing small-molecule ligands against virtual binding pockets, scoring spatial fit, conformational adaptability, and binding free energies to prioritize candidates with high receptor affinity.

The second core paradigm focuses on transcriptomic profile matching to target systemic host-pathogen interactions. Rather than focusing strictly on direct pathogen inhibition, this approach uses machine learning to decode the complex host transcriptomic perturbations induced by infection. Algorithms map high-dimensional gene expression signatures, capturing upregulated and downregulated host pathways during disease states, and cross-reference them against extensive perturbation databases, such as the Connectivity Map (CMap) [30]. By identifying small molecules that induce anti-correlated expression profiles, AI can predict repurposed therapeutics capable of reversing pathogenic gene signatures or bolstering host cellular defenses, effectively dampening infection-induced pathology.

The third methodological pillar leverages biomedical Natural Language Processing (NLP) to mine latent unstructured knowledge across vast scientific ecosystems. Large-scale transformer-based models ingest and synthesize millions of peer-reviewed articles, clinical trial registries, chemical patent databases, and electronic health records [31]. Through advanced named entity recognition (NER) and knowledge graph construction, NLP models extract hidden phenotypic correlations, off-target interactions, and secondary clinical outcomes that have remained buried within disconnected literature. By connecting disparate empirical findings across disciplines, NLP tools surface non-obvious mechanistic links between established drugs and pathogen susceptibility that would otherwise elude manual human review.

6. Emerging Paradigms and Next-Generation Solutions

To bridge the persistent gap between in silico predictions and real-world clinical translation, the field is rapidly advancing toward next-generation paradigms that directly address the underlying physics, physiological complexity, genomic diversity, safety profiles, and ethical governance of AI-driven drug discovery. Traditional deep learning architectures rely exclusively on statistical pattern recognition from historical training sets, often neglecting fundamental physical transport constraints; to resolve this, researchers are deploying Physics-Informed Neural Networks (PINNs), which directly embed thermodynamic principles, electrostatic dynamics, and molecular diffusion equations into the network’s loss functions, thereby improving predictions of compound permeation across the Gram-negative outer membrane or fungal chitin walls while reducing black-box opacity [32,33]. Concurrently, to mitigate the high frequency of in vitro false positives originating from static 2D cell cultures, computational virtual screening is being systematically coupled with biomimetic microfluidic organ-on-a-chip technologies, which replicate dynamic fluid flow, physiological shear stress, tissue-tissue interfaces, and active biofilm formation, enabling high-throughput automated validation of repurposed drugs within biomimetic infection microenvironments [34].

Furthermore, acknowledging that historical screening pipelines heavily relied on standardized laboratory reference strains, modern computational strategies are integrating population-scale pangenomics to evaluate target conservation across the fluid accessory genome, ensuring that AI-selected molecules maintain broad efficacy against highly diverse clinical isolates and emerging resistant lineages rather than single reference variants [35]. To prevent late-stage clinical attrition driven by off-target toxicities or cytotoxic dosing requirements, contemporary pipelines are also implementing early-stage multi-objective ADMET profiling, evaluating target binding affinity in parallel with predictive models for cardiac toxicity (e.g., hERG channel inhibition), hepatic clearance, and cytochrome P450 interactions during the initial virtual screening phase [36]. Finally, as generative AI architectures and active-learning platforms gain unprecedented capabilities in predicting molecular bioactivity and toxicity, the scientific community is establishing responsible AI governance and biosecurity guardrails to proactively manage dual-use risks, ensuring these computational frameworks cannot be exploited to design immune-evading pathogens or biological toxins [37].

7. Comparative Synthesis Across Pathogen Classes

To systematically navigate this multi-kingdom landscape, computational pipelines must account for the distinct biophysical constraints, delivery bottlenecks, and computational limitations imposed by each pathogen class. A comparative overview of these structural hierarchies is summarized in Table 1.

8. Concluding Remarks

The integration of Artificial Intelligence into drug repurposing constitutes a paradigm shift in modern pharmacology and clinical strategy against infectious diseases. Nevertheless, computational algorithms represent merely an initial predictive framework. To successfully mitigate the global burden of emerging resistant fungal pathogens and overcome clinical resistance mechanisms, in silico models must be seamlessly coupled with biomimetic validation platforms and rigorous clinical stewardship, thereby reconciling digital abstractions with the fundamental complexities of fungal physiology.