Using Network Analysis to Understand How Plant Compounds Work on Multiple Targets: Combining Molecular Studies, Lab Tests, and Clinical Applications

Study selection was based on relevance to the review’s predefined conceptual domains rather than on a formal systematic review eligibility algorithm. Studies were considered particularly informative when they addressed one or more of the following: compound–target interaction, biochemical activity, cellular pathway modulation, target engagement, genetic or functional validation, pharmacokinetics or metabolism, tissue exposure, microbiome-mediated transformation, human pharmacodynamic effects, clinical outcomes, drug-delivery strategies, or computational methods relevant to phytochemical target prioritisation.

For mechanistic studies, greater weight was given to investigations that clearly specified the tested compound, experimental model, concentration or dose, and molecular or functional readout. Evidence based on biochemical interaction measurements, cellular target engagement, genetic or pharmacological perturbation, rescue experiments, resistant mutants, or other orthogonal approaches was considered stronger than evidence based solely on downstream pathway modulation. This distinction is consistent with contemporary approaches to target identification and mechanism-of-action validation [19,26].

For pharmacokinetic studies, preference was given to reports that described a defined dose and formulation, together with quantitative measurements of the parent compound and/or relevant metabolites. Human pharmacokinetic and tissue exposure studies were prioritised in assessing translational plausibility. Total, unconjugated, and unbound exposures were distinguished where the source data permitted.

For clinical evidence, randomised controlled trials, prospective intervention studies, systematic reviews, and meta-analyses were preferentially considered. Clinical efficacy, pharmacodynamic responses, surrogate or intermediate endpoints, and safety or tolerability were treated as distinct categories of evidence.

Computational and network-pharmacology studies were included when relevant to target prediction, docking, network construction, pathway enrichment, artificial intelligence, or systems-level hypothesis generation. However, computational predictions without corresponding biochemical, cellular, functional, pharmacokinetic, or target-engagement validation were treated as hypothesis-generating rather than as confirmation of the molecular mechanism [19,24].

Primary studies were preferentially used for specific quantitative statements, including binding or inhibition constants, human pharmacokinetic parameters, tissue concentrations, target-engagement findings, and clinical outcomes. Reviews and meta-analyses were used primarily to contextualise broader bodies of evidence.

Non-English publications, conference abstracts lacking sufficient methodological detail, non-peer-reviewed reports, and studies lacking sufficient experimental information to evaluate the claim of interest were generally not used as primary evidence. Articles outside the mechanistic or translational scope of the review were likewise not considered in detail.

Evidence was evaluated based on the type and strength of experimental support rather than on the number of publications reporting an association. Compound–target relationships were distinguished among computational prediction, biochemical interaction or inhibition, cellular pathway modulation, cellular target engagement, functional or genetic evidence, in vivo pharmacology, and human pharmacokinetic, pharmacodynamic, or clinical evidence [19,26].

Particular attention was given to the correspondence between experimental potency or concentration, achievable human or tissue exposure, target engagement, and functional response. Computational prediction or pathway enrichment alone was not regarded as evidence of direct molecular engagement. Similarly, convergence of several phytochemicals on the same signalling pathway was not interpreted as evidence of pharmacological synergy without appropriate quantitative interaction analysis [27].

Where findings differed among studies, discrepancies were considered in relation to experimental model, dose or concentration, compound purity, formulation, exposure duration, analytical methodology, pharmacokinetic context, and strength of target validation.

Because this review was conducted as a structured narrative synthesis, formal risk-of-bias scoring, protocol registration, quantitative meta-analysis, pooled effect estimation, and PRISMA-style study-flow accounting were not performed. This methodological limitation is acknowledged explicitly to avoid implying the completeness or reproducibility expected of a systematic review.

The objective was therefore to provide a critical, mechanistically integrated, and translationally focused synthesis of representative evidence, rather than an exhaustive systematic catalogue of all publications in the field.

Many complex disease phenotypes arise from perturbations of interconnected gene and protein subnetworks, often conceptualised as disease modules, rather than from dysfunction of a single molecular target [3,28]. This systems-level perspective provides the conceptual basis for network pharmacology, which integrates compound–target associations, protein–protein interaction (PPI) networks, and functional or pathway enrichment analyses to generate hypotheses concerning the molecular basis of complex therapeutic responses [6,7]. This framework is particularly relevant to natural products and traditional medicines, in which multiple constituents may interact with partially overlapping biological systems [6,11].

Network-pharmacology analyses of medicinal plants and phytochemicals, including Ganoderma species, have reported associations with inflammatory, metabolic, oxidative-stress, and survival-related pathways, including NF-κB, PI3K/Akt, MAPK, Nrf2, and AMPK/mTOR signalling [7,29]. Such analyses can prioritise biologically plausible targets and pathways for experimental investigation, but neither pathway enrichment nor network connectivity establishes direct target engagement, simultaneous modulation of multiple proteins, or causal pathway involvement [5]. Accordingly, network-level associations are interpreted in this review as mechanistic hypotheses rather than established pharmacological mechanisms unless supported by independent experimental evidence.

PPI networks can be subjected to topological analysis to identify highly connected nodes, or “hubs”, within disease-associated networks [5]. Such nodes may provide useful candidates for experimental investigation. However, network centrality should not be equated with pharmacological importance because network representations depend on database coverage, interaction evidence, biological context, and analytical assumptions [5]. A highly connected node in a static network is therefore not necessarily an accessible drug target or a functionally dominant mediator of a phytochemical response.

Functional enrichment analyses provide an additional layer of interpretation by assessing whether predicted or experimentally identified target sets are statistically associated with specific biological processes or pathways, as represented by Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) annotations [5]. Their interpretation depends on the input target set, the database, the annotation coverage, and the analytical parameters. Overlap among related biological pathways may consequently contribute to apparent network convergence without demonstrating independent pharmacological modulation. Enrichment analyses should therefore be used primarily to prioritise biological hypotheses rather than establish pathway-level mechanisms [5].

The apparent complexity of a predicted network should likewise not be equated with the pharmacological complexity of a compound. Broad pathway modulation may arise from a smaller number of primary molecular events, whereas current databases or prediction algorithms may incompletely represent genuine multi-target effects. Network pharmacology is therefore most appropriately used to organise and prioritise hypotheses for subsequent experimental validation [5,19].

A central premise of this review is that network convergence is not equivalent to demonstrated polypharmacology. Shared enrichment of NF-κB, AMPK, Nrf2, PI3K/Akt, or related pathways may reflect engagement of multiple primary targets, downstream convergence following perturbation of fewer upstream targets, adaptive cellular responses, or biases arising from prediction algorithms, database coverage, network topology, and pathway annotation. These possibilities cannot be reliably distinguished from network topology or enrichment analysis alone [5].

For this review, heterogeneous mechanistic evidence is classified using the following operational six-tier framework.

Tier 0—Computational/predicted evidence. Molecular docking, network pharmacology, quantitative structure–activity relationship modelling, molecular dynamics, knowledge-graph analysis, or AI-based target prediction without independent biochemical or cellular confirmation. Such evidence is considered hypothesis-generating.

Tier 1—Biochemical/direct molecular evidence. Cell-free binding, enzymatic inhibition or activation, and covalent-modification studies performed under defined experimental conditions, including surface plasmon resonance, isothermal titration calorimetry, and validated biochemical activity assays. Depending on the assay, these approaches provide evidence of molecular interaction or biochemical activity, but do not by themselves establish cellular or physiological relevance [19].

Tier 2—Cellular/pathway evidence. Cell-based reporter assays, phosphorylation or transcriptional measurements, target-expression studies, and phenotypic experiments demonstrating modulation of a molecular pathway or cellular process. Such studies establish cellular activity but do not necessarily identify the molecular event responsible for the observed response [19].

Tier 3—Functional/causal and target-engagement evidence. Genetic perturbation approaches, including CRISPR knockout/knock-in, knockdown, rescue, or resistant-mutant studies, together with orthogonal target-engagement methods such as cellular thermal shift assays (CETSA) or chemoproteomic approaches [26,30,31]. Genetic dependency and physical target engagement provide complementary rather than interchangeable evidence. Confidence in target attribution is greatest when independent functional and target-engagement approaches converge on the same proposed mechanism [19,26].

Tier 4—In vivo/animal evidence—demonstration of pharmacological, biomarker, or phenotypic effects in an intact organism. Interpretation requires consideration of species differences, dose, formulation, route of administration, metabolism, tissue distribution, and exposure.

Tier 5—Clinical evidence. Evidence obtained in humans, including pharmacodynamic responses, validated biomarkers, clinical endpoints, or therapeutic outcomes. Its strength depends on study design, sample size, formulation, exposure, endpoint validity, reproducibility, and the distinction between pharmacodynamic activity and demonstrated clinical benefit. These tiers distinguish types of evidence rather than impose a rigid linear ranking. A biochemical interaction may provide stronger evidence of direct molecular interaction than a downstream clinical biomarker. In contrast, a human pharmacodynamic response may demonstrate biological activity without identifying the responsible molecular target. The framework, therefore, separates what has been directly demonstrated from what remains predicted, inferred, or causally unresolved.

Throughoutand, mechanistic claims are interpreted according to this framework. Where evidence of target-specific perturbation or direct target engagement is unavailable, pathway-level associations are described as putative or hypothesis-generating rather than as established molecular mechanisms. Section 4 Section 5

The concept of “selective non-selectivity”, developed in the context of multi-target drug design, describes an approach in which activity across several therapeutically relevant targets is balanced against avoidance of pharmacologically undesirable interactions [32]. It therefore provides a useful conceptual distinction between purposeful multi-target pharmacology and indiscriminate molecular promiscuity. Importantly, however, the presence of several measurable molecular interactions does not by itself establish therapeutically useful selectivity [10,32].

Affinity or potency alone is insufficient to determine whether a multi-target interaction profile is pharmacologically meaningful. For this review, reported IC₅₀, K_i, Kd, or EC₅₀ values are interpreted in relation to unbound drug concentration, tissue distribution, target abundance, binding kinetics (where known), cellular permeability, subcellular localisation, active metabolites, and the pharmacodynamic consequences of target modulation. Potency measured in a purified biochemical system may differ substantially from activity within cellular or tissue environments.

This issue is particularly relevant to phytochemicals because individual compounds are frequently associated with the modulation of several molecular targets and signalling pathways [7]. Such breadth should not automatically be interpreted as pharmacologically advantageous. For chemically reactive or assay-interfering compounds, apparent broad activity may also arise from nonspecific reactivity, aggregation, redox-related assay effects, or other assay-interference mechanisms [12,33,34]. The pharmacological value of a multi-target profile, therefore, depends on which targets are engaged, at what exposure, in which tissues, for what duration, and with what functional consequences.

For this critical review, therapeutically credible polypharmacology is operationally evaluated using five criteria:

Based on the evidence evaluated in this review, none of the four representative phytochemicals consistently satisfies all five criteria across their broad reported target profile. Berberine provides comparatively strong evidence for an energy-sensing, AMPK-associated pharmacological axis, including mechanisms involving mitochondrial complex I and downstream metabolic signalling. However, its low systemic parent-drug exposure and tissue-specific pharmacokinetics remain important translational considerations. By contrast, many reported cellular effects of curcumin and EGCG occur at concentrations that may exceed clinically achievable free systemic exposure, although local exposure, formulation, metabolism, and metabolite activity complicate direct comparison [16,20,21,35]. Accordingly, therapeutically useful selectivity remains a compound-, target-, tissue-, and exposure-dependent empirical question, rather than a conclusion that can be inferred from the breadth of a reported target list.

Hormetic dose–response relationships have been extensively reported for resveratrol, with lower and higher exposures producing qualitatively different biological effects in multiple experimental systems [36]. More generally, hormesis is defined as an adaptive response characterised by a biphasic dose–response relationship [37]. However, not every non-monotonic or biphasic dose–response relationship should automatically be classified as hormesis. Apparent non-monotonic responses may also be influenced by cytotoxicity, compound aggregation, nonspecific chemical reactivity, redox-related assay effects, or other forms of assay interference [33,34].

For this review, a response is classified as hormetic when it demonstrates: (i) a reproducible biphasic dose–response relationship; (ii) a low-dose stimulatory response compatible with an adaptive biological process; (iii) evidence that the stimulatory phase is not simply a consequence of compromised viability or nonspecific toxicity; and (iv) appropriate experimental controls and replication supporting a genuine adaptive response rather than an assay-specific artefact [36,37].

For example, activation of Keap1–Nrf2–antioxidant response element signalling can participate in adaptive cellular responses to electrophilic or oxidative stress. However, increased Nrf2-associated transcription alone does not establish hormesis; an appropriate dose–response relationship and separation of adaptive signalling from toxicity are also required [38,39].

This distinction is particularly important for chemically reactive phytochemicals. At increasing experimental concentrations, apparent changes in activity may be influenced by redox chemistry, nonspecific chemical reactivity, aggregation, assay interference, or overt cytotoxicity [12,33,34]. Such phenomena can produce concentration-dependent or non-monotonic responses without demonstrating a therapeutically beneficial hormetic mechanism. The terms hormesis, non-monotonic dose–response, and dose-dependent toxicity should therefore not be used interchangeably [37].

From a translational perspective, this review further considers a hormetic mechanism clinically relevant only when the adaptive response occurs within a pharmacologically achievable exposure range, produces a reproducible pharmacodynamic effect, and is not outweighed by toxicity or nonspecific molecular effects. A low concentration that activates a stress-response pathway and a higher concentration that reduces cellular viability are not, in themselves, sufficient to establish a therapeutically meaningful hormetic mechanism.

Taken together, these principles establish the conceptual framework used throughout this review. Network pharmacology prioritises candidate targets and pathways; biochemical and cellular studies characterise molecular and pathway activity; functional perturbation and target-engagement studies strengthen causal attribution; PK/PD analyses assess whether proposed mechanisms are compatible with in vivo exposure and response; and clinical studies evaluate whether such pharmacology translates into meaningful human benefit. The resulting translational workflow, incorporating iterative computational prediction, experimental validation, pharmacokinetic/pharmacodynamic assessment, formulation optimisation, biomarker evaluation, and clinical testing, is summarised in Figure 2.

Curcumin, berberine, epigallocatechin-3-gallate (EGCG), and resveratrol are widely investigated phytochemicals associated with inflammatory, metabolic, redox-sensitive, and cell-survival signalling processes [7]. However, their chemical structures, physicochemical properties, pharmacokinetic behaviour, and levels of mechanistic evidence differ substantially [12,14,21,40]. Accordingly, the target profiles considered below distinguish direct or putative biochemical interactions (Tier 1) from cellular pathway effects (Tier 2) and, where available, functional, causal, or target-engagement evidence (Tier 3). Reported target associations are therefore not assumed to represent equivalent pharmacological mechanisms.

Berberine also affects broader metabolic signalling and undergoes substantial interaction with intestinal, metabolic, and microbial processes [14,44,45]. These effects require separate consideration because their interpretation depends on intestinal and tissue exposure, metabolism, microbial transformation, and evidence of causal mediation. Berberine, therefore, illustrates how a comparatively well-characterised pharmacological pathway can coexist with a broader set of reported biological effects whose directness, exposure compatibility, and translational relevance remain variable.

EGCG has also been reported to inhibit DNA methyltransferase (DNMT) activity in a cell-free nuclear-extract assay, with competitive inhibition observed at micromolar concentrations [17]. Because the assay used nuclear extracts rather than purified DNMT1, these data support biochemical inhibition of DNMT activity but do not by themselves establish selective direct engagement of DNMT1. EGCG-associated modulation of PI3K/Akt, MAPK, NF-κB, Nrf2, and receptor-associated signalling has predominantly been demonstrated at the cellular level [7,21]. Such observations therefore support pathway modulation rather than direct engagement of every molecular component within these pathways. EGCG auto-oxidation and formation of reactive products further complicate the interpretation of experiments performed at relatively high concentrations [16,21].

Several effects previously attributed to direct SIRT1 activation may instead involve indirect mechanisms. Experimental studies have implicated inhibition of cyclic-AMP phosphodiesterases, altered cAMP signalling, CaMKKβ-dependent AMPK activation, and downstream NAD⁺/SIRT1-associated signalling [47]. Resveratrol is also associated with broader metabolic, inflammatory, and stress-response pathways [7,47]. These findings illustrate why downstream signalling responses should not be regarded as equivalent to validated direct target engagement.

Taken together, the four phytochemicals illustrate distinct forms of mechanistic complexity. Curcumin combines electrophilic reactivity with broad cellular signalling effects; berberine possesses a comparatively well-characterised energy-sensing pharmacological axis alongside additional metabolic and intestinal mechanisms; EGCG combines experimentally supported molecular interactions with chemical instability and exposure constraints; and resveratrol has a historically prominent but disputed SIRT1 mechanism alongside indirect signalling routes. Their inclusion as representative case studies does not imply equivalent or fully validated polypharmacological profiles. Rather, they illustrate how predicted or reported target networks should be evaluated against biochemical, cellular, functional, pharmacokinetic, in vivo, and clinical evidence.

A compound-centred translational comparison is provided in Table 1.

Curcumin and EGCG exhibit chemical and redox reactivity under experimental conditions [12,16,41], whereas berberine can alter cellular redox and metabolic states through effects on mitochondrial energy metabolism [44]. Resveratrol has also been associated with dose-dependent adaptive and stress-related responses [36]. However, describing all four compounds as exhibiting a uniform physiologically meaningful “redox switch” would overstate the available evidence.

Curcumin can undergo oxidation and electrophilic reactions associated with its conjugated carbonyl system, whereas EGCG is particularly susceptible to auto-oxidation [12,16,41]. Under appropriate experimental conditions, such chemistry can generate reactive intermediates and influence redox-sensitive signalling. Electrophilic or oxidative stress may activate cytoprotective signalling through the Keap1–Nrf2 pathway [38,39]. These observations are more appropriately interpreted as context-dependent biochemical and cellular effects (Tier 1–2) rather than as evidence of a universal physiological antioxidant mechanism.

Activation of Nrf2-associated transcription alone should not be interpreted as evidence of direct Keap1 modification or hormesis. Demonstration of direct target modification requires appropriate biochemical or target-engagement evidence, whereas classification as hormesis additionally requires a reproducible biphasic dose–response relationship and separation of adaptive responses from nonspecific toxicity [37].

Berberine lacks the polyphenolic architecture characteristic of EGCG and curcumin. Its effects on cellular stress and redox balance are more appropriately considered in relation to mitochondrial respiratory complex I inhibition, altered cellular energy metabolism, and downstream signalling [44]. Resveratrol-associated metabolic and stress responses may likewise involve indirect cAMP–CaMKKβ–AMPK and NAD⁺/SIRT1-associated signalling [47].

The apparent antioxidant or pro-oxidant behaviour of phytochemicals should therefore be interpreted as exposure- and context-dependent pharmacology rather than an intrinsic binary property. For compounds with known chemical reactivity or assay interference liabilities, high experimental concentrations may increase the contribution of nonspecific chemical reactivity, aggregation, assay interference, or cytotoxicity to the observed phenotypes [12,33,34]. Redox-associated pathway activation should consequently be assigned mechanistic significance only when supported by appropriate biochemical, cellular, functional, or target-engagement evidence.

Epigenetic effects have been proposed as an additional component of phytochemical activity. Still, direct enzyme inhibition should be distinguished from downstream changes in DNA methylation, histone modification, chromatin regulation, and non-coding RNA expression.

EGCG has been reported to inhibit DNMT activity in a cell-free nuclear-extract assay at micromolar concentrations [17]. Because this assay used nuclear extracts rather than purified DNMT1, the results support biochemical inhibition of DNMT activity but do not by themselves establish selective direct engagement of DNMT1. The same study reported reactivation of several methylation-silenced genes in cellular models [17]. These findings constitute cellular epigenetic evidence (Tier 2) rather than proof of clinically relevant DNMT1 engagement.

Curcumin has been shown to inhibit p300/CBP-associated histone acetyltransferase activity in biochemical systems [54]. Separately, curcumin has been reported to alter microRNA expression profiles in human pancreatic cancer cells [55]. These observations represent different levels of evidence: biochemical enzyme inhibition constitutes direct molecular evidence under defined experimental conditions, whereas changes in microRNA expression reflect cellular regulatory responses. Altered microRNA expression should therefore not be interpreted as evidence that individual microRNAs are direct molecular targets of curcumin.

Resveratrol has been associated with SIRT1-related deacetylation and mitochondrial regulation, but interpretation remains constrained by uncertainty regarding direct SIRT1 activation [18,47]. These effects are therefore better described as SIRT1- or AMPK-associated downstream responses than as unequivocal evidence of direct resveratrol–SIRT1 engagement.

Overall, epigenetic observations should not be incorporated into a polypharmacological target network solely because changes in DNA methylation, histone modification, chromatin state, or gene expression are observed. Stronger target attribution requires evidence connecting the phytochemical to a defined molecular target, demonstrating target engagement where feasible, and establishing that target modulation contributes functionally to the observed phenotype [19,26].

Curcumin, EGCG, and berberine exhibit limited systemic exposure to their unmetabolized parent forms after conventional oral administration, while gastrointestinal exposure, metabolism, and microbial transformation may contribute substantially to their disposition and biological effects [14,16,45,56]. These characteristics have stimulated interest in the gut microbiome as a potential contributor to the pharmacology of phytochemicals. However, microbiome-associated changes should not automatically be interpreted as evidence that microbial modulation mediates a therapeutic phenotype.

The evidence for berberine illustrates this distinction. In the PREMOTE study, berberine treatment was associated with changes in gut microbial composition, bile-acid metabolism, and metabolic outcomes, with mechanistic analyses implicating altered microbial bile-acid biotransformation involving Ruminococcus bromii [57]. Nevertheless, concurrent changes in microbial composition and host metabolic endpoints do not establish that every altered microbial population is necessary or sufficient for the therapeutic effect.

Causal mediation requires experimental approaches that can separate microbial effects from parallel host- or drug-mediated mechanisms. Relevant strategies include germ-free or gnotobiotic models, defined microbial communities, faecal microbiota transplantation, targeted manipulation of candidate organisms, and administration of defined microbial metabolites [58]. Convergent evidence from such approaches provides stronger support for microbiome-mediated pharmacology than changes in microbial abundance alone.

Comparable caution applies to curcumin. A small randomised pilot study reported changes in gut microbial diversity and composition following turmeric or curcumin supplementation, but responses were highly individualised [59]. These findings demonstrate microbiome responsiveness to treatment but do not establish that particular microbial taxa causally mediate therapeutic effects.

An additional complication is that the pharmacologically relevant molecular species may differ from the administered parent compound. Gut bacteria can convert berberine into a more readily absorbable form, demonstrating that microbial metabolism can alter its intestinal disposition [45]. Parent compounds, host-generated metabolites, and microbiota-derived metabolites should therefore be considered separately when interpreting phytochemical pharmacology [45,60].

Accordingly, microbiome-mediated effects are considered here as an indirect layer of pharmacology rather than as evidence of polypharmacology per se. Host–microbiome–metabolite interactions may contribute to biological responses, but demonstrating changes in multiple microbes or hosts does not establish simultaneous, therapeutically relevant engagement of multiple molecular targets. Limited systemic exposure of a parent compound does not necessarily exclude intestinal or metabolite-mediated pharmacology; conversely, therapeutic effects should not be attributed to “microbiome modulation” solely because microbial composition changes.

The microbiome is therefore best incorporated into phytochemical network models as a candidate intermediate biological layer between oral exposure and host molecular effects rather than automatically as an established therapeutic target. Demonstration of causal mediation requires integrating microbiome perturbation, metabolite measurements, pharmacokinetics, target engagement (where applicable), and phenotypic outcomes.

NF-κB signalling is a frequently reported point of pathway-level convergence among curcumin, berberine, EGCG, and resveratrol, although their molecular entry points differ substantially [7].

Curcumin has been reported to suppress IKK/NF-κB-associated signalling in cellular systems [42,43]. However, these observations should not automatically be interpreted as evidence that IKKβ is the dominant intracellular target responsible for NF-κB suppression. Curcumin’s electrophilic reactivity, chemical instability, and assay-interference liability complicate attribution of broad cellular effects to a single molecular interaction [12]. Accordingly, cellular NF-κB suppression is classified primarily as Tier 2 evidence unless supported by direct target-engagement and functional studies.

EGCG has been shown to bind the 67 kDa laminin receptor (67LR) and has independently been reported to suppress NF-κB-associated signalling in cellular systems [21,46]. However, the presence of a validated EGCG–67LR interaction does not by itself establish that 67LR is the dominant mediator of NF-κB suppression. The NF-κB phenotype is therefore best interpreted primarily as modulation of cellular pathways (Tier 2), based on the evidence in the present bibliography.

Berberine inhibits mitochondrial respiratory complex I and subsequently activates AMPK-associated energy-sensing pathways [44]. NF-κB-associated modulation has also been reported for berberine [7], but the available references do not establish a single causal complex I–AMPK–NF-κB pathway. The relationship is therefore more appropriately regarded as indirect pathway-level convergence rather than direct NF-κB engagement.

SIRT1 can repress NF-κB transcriptional activity by deacetylating RelA/p65 at Lys310, and resveratrol treatment was associated with reduced NF-κB-dependent transcription in the original cellular study [61]. However, direct activation of SIRT1 by resveratrol is mechanistically disputed, whereas alternative studies support indirect PDE–cAMP–AMPK–NAD⁺/SIRT1-associated signalling. Resveratrol-mediated NF-κB suppression should therefore be described as a SIRT1-associated cellular mechanism rather than evidence of direct resveratrol–SIRT1 engagement [18,47].

Collectively, these observations demonstrate pathway-level convergence without equivalent molecular mechanisms. Curcumin, berberine, EGCG, and resveratrol can influence NF-κB-associated signalling through different biochemical or cellular routes [7]. Shared suppression of the same downstream pathway does not, by itself, establish simultaneous multi-target engagement or an equivalent causal mechanism.

Accordingly, NF-κB suppression is not treated in this review as intrinsically beneficial; its pharmacological significance is considered to depend on the biological and disease context.

The Nrf2/Keap1 system is another frequently reported signalling axis associated with phytochemical responses. However, the mechanisms leading to Nrf2 activation differ among compounds and should not be reduced to a generalised antioxidant effect [7,39].

Curcumin contains electrophilic structural features capable of reacting with nucleophilic residues [41], whereas EGCG is susceptible to oxidation and formation of reactive products under experimental conditions [16]. More generally, electrophilic or oxidative modification of redox-sensitive Keap1 cysteines can disrupt Keap1-dependent repression of Nrf2 and promote antioxidant-response-element-dependent transcription [38,39]. These chemical- and pathway-level observations, however, do not demonstrate that curcumin or EGCG directly engages Keap1 under pharmacologically relevant intracellular conditions.

Cellular Nrf2 activation should therefore be distinguished from direct Keap1 engagement. Demonstrating a specific phytochemical–Keap1 interaction requires biochemical or target-engagement evidence, whereas demonstrating that such an interaction causes a biological phenotype requires additional functional validation.

Nrf2 activation should likewise not automatically be described as hormetic. As defined in Section 3.4, hormesis requires an appropriately characterised biphasic dose–response relationship rather than pathway activation alone [37].

Nrf2 is functionally interconnected with inflammatory, metabolic, and stress-response networks, including NF-κB and kinase-associated signalling [39]. It is therefore appropriately considered a regulatory node within a broader signalling network, while the therapeutic relevance of its modulation remains dependent on exposure, tissue context, and causal evidence.

AMPK is a central cellular energy-sensing kinase that interacts directly with mTOR-regulated autophagy mechanisms, including regulation of ULK1 [62].

Among the four phytochemicals considered here, berberine has comparatively strong mechanistic evidence for mitochondrial respiratory complex I inhibition, followed by altered cellular energy status and AMPK activation [44]. However, systemic exposure to unchanged berberine in humans is low, and intestinal, tissue-specific, and metabolite-mediated exposure must be considered when extrapolating these mechanisms to clinical pharmacology [14].

Resveratrol can activate AMPK indirectly through PDE inhibition, increased cAMP signalling, CaMKKβ activation, and downstream metabolic signalling [47]. Curcumin has been reported to influence Akt/mTOR and autophagy-associated processes in cellular models [42], while EGCG has also been associated with cellular autophagy-related responses [21]. These observations predominantly represent Tier 2 pathway evidence and do not establish AMPK, mTOR, or the autophagy machinery as direct molecular targets of each compound.

AMPK can suppress mTORC1 activity and regulate autophagy through phosphorylation of ULK1 [62]. Nevertheless, observation of AMPK-, mTOR-, or autophagy-associated changes across several phytochemicals does not establish a shared causal mechanism. Similar downstream pathway changes can arise from distinct upstream molecular perturbations.

AMPK convergence should therefore be regarded as a plausible organising framework rather than a universal mechanism of phytochemical action. Stronger causal attribution requires pathway-specific or target-specific perturbation together with complementary molecular evidence [19,26].

Persistent STAT3 activation occurs in multiple malignancies and can contribute to tumour-cell proliferation, survival, inflammatory signalling, and immune regulation [63].

Curcumin has been reported to suppress JAK/STAT3-associated signalling in cellular cancer models [43]. Based on the references available here, this is best regarded as cellular pathway modulation (Tier 2) rather than as evidence of direct curcumin–STAT3 engagement.

EGCG has been shown experimentally to suppress STAT3 signalling in human pancreatic cancer models [64]. Resveratrol has likewise been shown to inhibit Src–STAT3 signalling and reduce STAT3-dependent transcriptional activity in malignant cells [65]. These studies support cellular inhibition of STAT3-associated signalling but do not demonstrate that EGCG and resveratrol engage STAT3 through identical molecular mechanisms.

The recurring observation across these compounds is therefore more accurately described as convergent suppression of STAT3-associated signalling rather than equivalent STAT3-targeted pharmacology. Reduced STAT3 phosphorylation or transcriptional activity does not by itself identify the initiating molecular target; direct target attribution requires orthogonal biochemical or target-engagement evidence [19,26].

The signalling pathways discussed above are interconnected. AMPK regulates mTOR and autophagy-associated signalling [62]. Nrf2 interacts with inflammatory and stress-response networks [39]; and SIRT1 can regulate NF-κB-dependent transcription through deacetylation of RelA/p65 [61]. These relationships provide a biological basis for the propagation of perturbations across signalling modules.

However, pathway crosstalk should not be equated with pharmacological synergy. Simultaneous modulation of multiple pathways can yield additive, synergistic, antagonistic, or compensatory responses, depending on the biological system and experimental conditions.

Formal synergy requires quantitative comparison with an explicit reference model. Approaches, including Bliss independence and Loewe additivity, can be selected according to the experimental design and underlying assumptions [27]. Consequently, simultaneous changes in multiple signalling pathways alone are insufficient to establish pharmacological synergy.

Similarly, this review does not use the term “emergent effect” merely because several pathways change simultaneously. The term is reserved for situations in which a network-level phenotype cannot be adequately accounted for by experimentally established individual effects or an appropriate additive framework.

Where formal interaction analysis is unavailable, the terms “network-level convergence” or “potentially cooperative effects” are used instead of “synergy”.

A pathway-centred comparison of the evidence supporting modulation of NF-κB, Nrf2/Keap1, AMPK/mTOR/autophagy, and STAT3 by curcumin, berberine, EGCG, and resveratrol is summarised in Table 2. The table distinguishes direct or functional molecular evidence from cellular pathway modulation, human pharmacodynamic observations, and conceptual network-level convergence.

Structurally distinct phytochemicals can produce overlapping pathway-level responses despite differing molecular entry points. Apparent convergence may arise from distinct primary molecular interactions, downstream signalling propagation, adaptive responses, or limitations of network and pathway analyses [7,13,62,63]. Accordingly, shared pathway annotations across curcumin, berberine, EGCG, and resveratrol should be regarded as candidate mechanistic intersections rather than evidence of equivalent polypharmacology.

Within the evidentiary framework defined in Section 3.2, these intersections should be prioritised for experimental testing, with translational relevance determined by appropriate molecular or functional support and compatibility with pharmacologically relevant exposure [19,26].

A central barrier to the clinical translation of phytochemical polypharmacology is the potential mismatch between the concentrations that produce biological effects in experimental systems and the exposures achievable in humans. This issue is particularly important for network pharmacology because identification of a predicted molecular target does not establish that a pharmacologically active species reaches that target at sufficient concentration in vivo or that target engagement occurs. Curcumin, berberine, EGCG, and resveratrol all exhibit substantial pharmacokinetic constraints involving absorption, distribution, metabolism, and/or formulation dependence that complicate direct extrapolation of mechanistic in vitro observations to human pharmacology [14,20,21,22].

Accordingly, Table 3 emphasises measured human parent-compound exposure, metabolites, tissue concentrations where available, and representative target-specific activity or affinity data. Unbound target-site concentrations are not inferred from total or unconjugated plasma concentrations when such measurements are unavailable.

The quantitative comparisons in Table 3 demonstrate why plasma concentration, tissue concentration, mechanistic activity concentration, and target affinity should not be treated as interchangeable quantities. For curcumin, a recent independent crossover study found very low unconjugated plasma concentrations even with formulations designed to enhance systemic exposure [67], whereas direct sampling in patients demonstrated measurable parent compound in colorectal tissue following oral administration [48]. Similarly, direct human tissue studies have detected EGCG and its methylated metabolite in prostate tissue and resveratrol together with multiple conjugated metabolites in colorectal tissue [51,52]. Low systemic parent-compound exposure, therefore, does not exclude local pharmacology, but tissue detection alone does not establish engagement of a particular molecular target.

EGCG illustrates why translational assessment should be target-specific rather than compound-wide. The reported affinity of EGCG for 67LR is in the nanomolar range, whereas inhibition of nuclear DNMT activity was observed at micromolar concentrations in a nuclear-extract assay. The latter study did not selectively establish purified DNMT1 engagement [17,46]. Thus, a human exposure that may be more consistent with one proposed molecular interaction cannot automatically be used to support another mechanism that requires substantially higher experimental concentrations.

A further limitation is the scarcity of quantitative unbound target-site concentrations. Total, unchanged, or unconjugated plasma measurements should not automatically be equated with the freely available concentration at an intracellular molecular target. Protein binding, membrane permeability, subcellular distribution, metabolite formation, exposure duration, and tissue-specific disposition can all affect the relationship between measured systemic concentration and molecular target exposure. Consequently, comparisons between human exposure and experimental activity should use target-specific potency or affinity together with unbound target-site exposure whenever these quantities are available.

For network pharmacology, this distinction is fundamental. A computationally plausible interaction may remain pharmacologically inaccessible if the relevant active molecular species does not reach the proposed target at a sufficient exposure level. Conversely, local gastrointestinal or tissue exposure may make selected mechanisms more plausible than would be inferred from plasma measurements alone. Target prioritisation should therefore incorporate measured human exposure, metabolite identity, tissue distribution where available, target-specific potency or affinity, and direct evidence of target engagement, rather than network position or docking affinity alone.

Tissue exposure can differ substantially from circulating exposure. This distinction is particularly relevant to orally administered phytochemicals because gastrointestinal tissues may encounter the parent compound before extensive first-pass metabolism. Direct human measurements support this principle for curcumin, EGCG, and resveratrol [48,51,52].

For curcumin, oral administration of 3.6 g/day resulted in measurable parent compound in both normal and malignant colorectal tissue despite only trace peripheral exposure [48]. These observations support the plausibility of local gastrointestinal pharmacology even when systemic parent concentrations are low. However, colorectal tissue exposure should not be extrapolated to distant organs without corresponding tissue-specific measurements.

Likewise, EGCG and its methylated derivative have been detected directly in human prostate tissue following repeated green-tea consumption [51]. These data establish tissue access but do not demonstrate that unbound intracellular concentrations at a particular molecular target are sufficient to reproduce mechanisms identified in higher-concentration cellular experiments.

Resveratrol provides a further example of tissue-specific exposure. Parent resveratrol and multiple glucuronide and sulfate metabolites have been quantified in human colorectal tissue after repeated oral administration, with substantial variation according to anatomical location and individual patient [52]. Such findings emphasise that plasma exposure alone may not accurately describe local gastrointestinal exposure.

For neurological indications, the blood–brain barrier (BBB) introduces an additional requirement. Robust quantitative human brain target-site exposure data remain limited for curcumin, berberine, and EGCG in the current reference set. Systemic dosing, peripheral pharmacodynamic activity, or preclinical CNS effects should therefore not be interpreted as evidence of clinically relevant human brain target engagement.

For resveratrol, human evidence is somewhat stronger. In a randomised Phase II study in patients with Alzheimer’s disease, parent resveratrol and several metabolites were measurable in cerebrospinal fluid, demonstrating CNS access of resveratrol-derived molecular species [69]. Nevertheless, detection in cerebrospinal fluid does not establish adequate unbound concentration at a specific neural target, nor does it demonstrate target engagement or causal contribution to a clinical outcome.

Thus, tissue detection and BBB penetration are necessary but insufficient evidence for target-specific pharmacology. For CNS mechanisms in particular, translational confidence requires quantitative brain or CSF exposure where feasible, identification of the pharmacologically active molecular species, and corresponding pharmacodynamic or target-engagement evidence.

The divergence between experimental activity and achievable human exposure represents a major translational challenge for phytochemical pharmacology. This relationship cannot be assessed using plasma parent-compound concentration alone because local gastrointestinal exposure, tissue distribution, metabolite formation, formulation, and interindividual pharmacokinetic variability can substantially modify exposure [48,49,50,51,52,53,67,68].

These factors should not, however, be invoked automatically to explain an apparent exposure gap. Higher tissue concentration does not establish target engagement or therapeutic efficacy, and detection of a metabolite does not establish that the metabolite is pharmacologically active. Likewise, downstream modulation of a signalling pathway after dosing does not demonstrate that a predicted molecular target was directly engaged.

A stronger translational chain is:

administered dose → systemic and/or local exposure → active molecular species → target-site exposure → target engagement → proximal pharmacodynamic response → functional or clinical outcome.

Each step addresses a distinct pharmacological question. Pharmacokinetic measurements establish exposure; biochemical or biophysical studies establish the capacity for molecular interaction under defined conditions; target-engagement approaches determine whether the proposed interaction occurs in the relevant biological environment; and pharmacodynamic measurements assess whether the expected biological response accompanies molecular engagement [19,26]. Clinical benefit represents a further evidentiary step and should not be inferred from pharmacodynamic activity alone.

For network pharmacology, this distinction is fundamental because computational target prediction, PPI topology, molecular docking, and pathway enrichment can generate biologically plausible associations without establishing whether sufficient active compound reaches the proposed targets. Pharmacokinetic filtering should therefore be incorporated into target prioritisation before broad network convergence is interpreted as therapeutically relevant polypharmacology.

Where sufficient quantitative information is available, physiologically based pharmacokinetic (PBPK) modelling can integrate dose, absorption, metabolism, physiological parameters, and tissue distribution to estimate exposure at anatomical sites that are difficult to sample directly [70]. Such predictions remain dependent on model structure, parameter quality, assumptions, and validation and should therefore complement rather than replace direct measurements of tissue exposure, target engagement, and pharmacodynamic response [70].

For translational interpretation, proposed mechanisms can be considered as computationally plausible but pharmacokinetically unsupported, exposure-compatible but causally unresolved, or exposure-compatible and experimentally supported. These categories distinguish mechanisms lacking evidence of sufficient target-site exposure from those for which exposure is plausible but direct engagement or functional necessity remains uncertain, and from those supported by convergent evidence from exposure, target engagement, pharmacodynamics, function, and, where available, in vivo or clinical studies.

Certain phytochemicals can exhibit substantial assay-interference liabilities, with curcumin as a prominent example due to its chemical reactivity, instability, aggregation-related behaviour, and potential interference across multiple assay formats [12,33,34]. Consequently, activity observed in a single biochemical or cellular assay should not, by itself, be interpreted as evidence of a specific molecular interaction.

PAINS alerts should be regarded as methodological warnings rather than evidence that a compound is biologically inactive [33,34]. Curcumin illustrates this distinction: its electrophilic and chemically reactive properties can generate nonspecific effects under some experimental conditions, while the same chemistry may also permit biologically relevant reactions with susceptible nucleophilic residues in proteins [12,41]. Thus, reported activity should neither be dismissed solely because of assay-interference liability nor accepted as target-specific without independent validation.

The distinction among chemical reactivity, biochemical interaction, cellular pathway modulation, physical target engagement, and causal target dependence is therefore essential. A biochemical interaction with an isolated protein provides evidence of molecular activity under defined conditions but does not establish cellular or physiological relevance. Conversely, loss of a phenotype following genetic perturbation supports functional involvement but does not independently demonstrate direct physical binding [19].

For compounds with substantial assay-interference liability, target validation should therefore rely on complementary and mechanistically independent approaches. These may include orthogonal biochemical assays, cellular target-engagement methods such as the cellular thermal shift assay (CETSA), chemoproteomic approaches, genetic perturbation, rescue or resistant-mutant experiments, and structure–activity analysis [19,26]. Confidence in a proposed molecular target is greatest when direct or orthogonal target-engagement evidence converges with functional perturbation and a phenotype compatible with pharmacologically achievable exposure.

Berberine illustrates the complementary principle that absence of a prominent assay-interference concern does not itself establish mechanistic validity. Clinical studies and meta-analyses support the effects of berberine on selected glycaemic and lipid outcomes, while multiple metabolic and microbiome-associated mechanisms have been proposed [14,60,71]. Thus, assay-interference liability and pharmacological validity are distinct questions; mechanistic validity additionally requires exposure compatibility, target engagement, functional contribution, and biological relevance.

Publication and selective-reporting bias are established concerns in biomedical research and can distort the apparent consistency of a literature by preferentially representing positive findings [72]. This issue is relevant to phytochemical pharmacology because repeatedly reported molecular or cellular associations may subsequently be incorporated into mechanistic reviews, databases, and computational target networks.

Reproducibility is also affected by the quality of the experimental system. Cell-line misidentification and contamination are documented sources of irreproducibility in biomedical research [73]. For chemically challenging phytochemicals, compound identity, purity, stability, and assay conditions can further influence experimental outcomes [12]. Botanical preparations introduce additional variability because their composition may depend on source material, cultivation, processing, extraction, purification, and storage [11].

A predominance of positive reports should therefore not be interpreted automatically as evidence of a large or reproducible biological effect. Publication selection, experimental heterogeneity, and assay-specific limitations may contribute to apparent consistency [34,72].

This issue is particularly important in network pharmacology because computational networks frequently incorporate previously reported compound–target relationships. Bias or error in the underlying literature and databases can therefore propagate into target networks and pathway-enrichment analyses [5]. Accordingly, network density, repeated database assignment, or high node connectivity should not be regarded as independent confirmation of pharmacological promiscuity or polypharmacology. Explicit evidence grading and independent experimental validation are required to reduce this risk.

Molecular docking and related structure-based approaches are useful for generating and prioritising candidate phytochemical–target interactions, but computational scores are model-derived quantities rather than experimentally measured binding affinities or evidence of target engagement [19,25].

An important limitation is that proteins are conformationally dynamic. Docking against a single static structure may sample only one of several biologically relevant receptor states, and predicted ligand interactions can therefore depend on the selected structural conformation. Ensemble docking addresses this limitation by evaluating ligands across multiple receptor conformations and thereby incorporating receptor flexibility into structure-based prediction [25].

Nevertheless, greater computational sophistication does not convert a prediction into experimental evidence. A favourable docking score, a stable molecular-dynamics trajectory, or consistency across several computational models can strengthen the plausibility of a hypothesis, but none independently establishes direct binding, intracellular target engagement, or pharmacological relevance [19,25].

A stronger strategy is progressive orthogonal validation. Computationally prioritised interactions can be followed by biochemical interaction measurements, cellular target-engagement assays, and functional perturbation studies [19,26]. A relevant phytochemical example is provided by Leo et al., who combined docking and molecular-dynamics analysis of oleacein and oleocanthal with pharmacological inhibition of PPARγ and PPARG gene silencing [74]. Such convergence between computational prediction and functional perturbation substantially strengthens mechanistic attribution. However, pharmacological inhibition and gene silencing should still not be interpreted as direct biophysical evidence of ligand–receptor binding.

The same distinction applies to genetic validation more generally. Knockdown, knockout, or CRISPR dependency can demonstrate that a protein contributes functionally to a phenotype, but genetic dependency alone does not establish that a phytochemical physically binds that protein. Confidence in direct target attribution is strongest when functional perturbation converges with independent biochemical or cellular target-engagement evidence [19,26].

Large network-pharmacology target sets require similar caution. Predicted target lists can arise from chemical similarity, ligand-based inference, database mining, the literature-derived associations, or computational prediction and may substantially exceed the number of proteins actually engaged at pharmacologically relevant exposure [5,6]. PPI topology and pathway enrichment are likewise influenced by database coverage and annotation structure and should therefore be interpreted as prioritisation tools rather than direct measures of pharmacological importance [5].

Agreement among computational approaches can increase confidence in target prioritisation, but computational convergence remains distinct from experimental convergence. Docking, molecular dynamics, AI-based prediction, and network analysis are therefore best regarded as complementary hypothesis-generating tools rather than substitutes for biochemical interaction measurements, target engagement, functional perturbation, pharmacokinetic plausibility, or clinical evidence [19,24].

Standardisation represents an additional challenge in phytochemical research. Natural-product preparations can vary according to botanical source, cultivation and processing conditions, extraction, purification, storage, and formulation, resulting in differences in chemical composition and biological activity [11].

Curcuminoid preparations may contain differing proportions of curcumin and structurally related curcuminoids [13]. Green-tea preparations likewise contain mixtures of catechins whose abundance and stability can vary according to preparation and processing conditions [16,21]. Such compositional differences can influence pharmacokinetics, metabolism, assay behaviour, and interpretation of biological effects.

Standardisation is particularly important in network pharmacology because computational analyses depend on the identity of the compounds entered into the analytical model [5]. If an experimental preparation contains multiple constituents but the computational analysis represents only a single purified compound, the resulting target network may not accurately reflect the material experimentally tested. Conversely, the biological activity of a multicomponent preparation should not automatically be attributed to its most abundant constituent.

Mechanistic studies should therefore document the identity and composition of the tested material and provide appropriate chemical characterisation, particularly when extracts or multicomponent botanical preparations are investigated [11]. When network pharmacology is applied to such preparations, compounds included in the computational analysis should, where possible, correspond to chemically verified constituents of the material being tested.

Overall, the complexity of a computational network should not be equated with the complexity of pharmacological action. Confidence in a multi-target interpretation increases when chemical characterisation, computational prediction, biochemical interaction, cellular target engagement, functional validation, pharmacokinetic plausibility, and, where relevant, in vivo or clinical evidence are mutually consistent. This evidence chain helps distinguish credible polypharmacology from assay interference, database-derived association, downstream pathway convergence, and exposure-incompatible target predictions.

The clinical studies considered in this review do not establish stand-alone anticancer efficacy or clinically validated multi-target anticancer pharmacology for the representative phytochemicals discussed below.

In a Phase II study of patients with advanced pancreatic cancer, oral curcumin at 8 g/day was generally well tolerated and produced clinical and biological activity in a small number of patients, but evidence of objective antitumour efficacy was limited [66]. 25 patients were enrolled, and 21 were evaluable for response. The study also demonstrated low circulating curcumin exposure despite the high oral dose [66]. These findings support tolerability and limited human biological activity but do not establish curcumin as an effective anticancer therapy or demonstrate engagement of the multiple molecular targets proposed from preclinical studies.

A Phase IIa study of curcumin in colorectal neoplasia provides a different type of evidence. In this nonrandomized, open-label study, 44 eligible participants received 2 or 4 g/day curcumin for 30 days, with 41 completing treatment [75]. The 4 g/day dose was associated with a significant reduction in the number of aberrant crypt foci (ACFs), whereas the 2 g/day dose was not. However, the prespecified mucosal measures PGE₂ and 5-HETE and the proliferation marker K_i-67 were not significantly altered [75]. The reduction in ACF should therefore be interpreted as an intermediate finding and not as a validated surrogate for the prevention of colorectal cancer. Direct human tissue measurements from a separate study demonstrated measurable curcumin and metabolites in colorectal tissue after oral administration, supporting the plausibility of local gastrointestinal exposure despite low systemic parent-compound concentrations [48].

Green tea extract provides a complementary example. In a randomised Phase II study of patients with high-risk oral premalignant lesions, the overall clinical lesion-response rate was numerically higher with green-tea extract than with placebo but did not reach statistical significance [76]. Higher-dose treatment groups showed evidence of a dose–response relationship, whereas histological improvement was not statistically significant [76]. Because the intervention was a green tea extract rather than purified EGCG, these findings should not be attributed specifically to EGCG or interpreted as evidence that multiple proposed EGCG targets were engaged simultaneously. The results instead indicate a short-term lesion-response signal that requires confirmation in studies designed to evaluate cancer-preventive efficacy.

These oncology studies illustrate why extensive preclinical pathway modulation should not be equated with clinical polypharmacology. Demonstration of NF-κB, STAT3, Nrf2, apoptotic, or other signalling effects in experimental models does not establish that the corresponding molecular targets are simultaneously engaged at achievable human exposure. Clinical mechanistic attribution, therefore, requires an exposure-compatible link among molecular engagement, pharmacodynamic response, and a clinically meaningful outcome.

Berberine has a comparatively substantial human evidence base for effects on metabolic biomarkers and intermediate endpoints. Meta-analyses of randomised studies have reported reductions in measures of glycaemia, including fasting and postprandial glucose and HbA1c, together with improvements in selected lipid parameters [71,77]. However, the earlier meta-analysis identified limitations in the methodological quality of the available studies [71]. The evidence, therefore, supports effects on selected metabolic endpoints but should not be interpreted as equivalent to definitive evidence of improvement in long-term cardiovascular, microvascular, or mortality outcomes.

Mechanistic studies implicate mitochondrial energy metabolism and downstream AMPK activation in the pharmacology of berberine [44]. In humans, the PREMOTE randomised clinical study demonstrated treatment-associated changes in gut microbial composition and bile acid metabolism and provided mechanistic evidence implicating altered Ruminococcus bromii-associated deoxycholic acid biotransformation in the metabolic response [57]. This evidence supports a microbiome-mediated contribution more strongly than microbial association alone. However, it does not establish that this mechanism accounts for all of the observed metabolic benefit, nor does it establish AMPK activation as the causal mechanism in humans.

Berberine, therefore, illustrates an important distinction between evidence of human metabolic activity and evidence of its molecular cause. Improvement in HbA1c, glucose, or lipid parameters does not demonstrate simultaneous engagement of multiple predicted targets or establish network-level polypharmacology. Prospective studies integrating parent-compound and metabolite exposure, mechanistically proximal pharmacodynamic biomarkers, and, where feasible, direct target-engagement measurements are needed to link clinical responses to specific molecular mechanisms.

Clinical translation has been less convincing for curcumin and resveratrol in established Alzheimer’s disease.

In a randomised, double-blind, placebo-controlled study of 36 participants with mild-to-moderate Alzheimer’s disease, curcumin at 2 or 4 g/day did not produce significant differences in cognitive outcomes or the measured plasma and cerebrospinal fluid Alzheimer-related biomarkers compared with placebo [78]. Plasma concentrations of native curcumin were low. Although treatment was generally tolerated, several participants receiving curcumin withdrew because of gastrointestinal adverse effects [78]. Thus, the trial provided neither convincing clinical efficacy nor biomarker evidence of therapeutic activity under the conditions studied.

A randomised, double-blind, placebo-controlled Phase II study of resveratrol in 119 participants with mild-to-moderate Alzheimer’s disease demonstrated a different pattern [69]. Resveratrol and its major metabolites were measurable in plasma and cerebrospinal fluid, and treatment altered selected biomarker trajectories. The trial found no significant treatment effects on major cognitive and global measures, including ADAS-cog, MMSE, CDR-SOB, and NPI, although the decline in the secondary functional measure ADCS-ADL was smaller in the resveratrol group than in the placebo group [69]. Because the trial was primarily designed to evaluate safety, tolerability, pharmacokinetics, and biomarker outcomes, and was not powered to establish definitive clinical efficacy, this isolated secondary-outcome signal should be interpreted with caution. The investigators likewise emphasised that the observed biomarker changes did not establish clinical benefit [69].

These studies illustrate the distinction among CNS exposure, pharmacodynamic activity, intermediate biomarker responses, and clinical efficacy. Detection of a parent compound or metabolite in plasma or cerebrospinal fluid does not demonstrate that a specific neural target is engaged, and alteration of a biomarker does not necessarily translate into improvement in cognition, function, or disease progression.

For CNS indications, the exposure and target-accessibility considerations discussed in Section 6.2 are therefore central to interpretation. The stronger human CNS-exposure evidence available for resveratrol demonstrates access of resveratrol-derived molecular species to cerebrospinal fluid, but does not establish exposure-compatible engagement of any particular molecular target [69]. Corresponding quantitative human CNS target-site evidence remains substantially more limited for curcumin, berberine, and EGCG within the literature evaluated in this review.

The principal clinical studies discussed above, along with their study designs, major outcomes, human PK/PD or mechanistic information, and translational interpretations, are summarised in Table 4.

The studies summarised in Table 4 also illustrate why comparisons across diseases require attention to trial design, sample size, formulation and dose, exposure, endpoint type, and evidentiary purpose. A change in a biomarker, intermediate endpoint, lesion response, or metabolic measure should therefore be interpreted according to the clinical question the study was designed to answer, rather than being placed on the same evidentiary level as improvement in a validated clinical outcome.

A recurrent translational problem is the disconnect between experimentally demonstrated biological activity and pharmacologically relevant human exposure. This constraint is well documented for curcumin, berberine, EGCG, and resveratrol [14,20,21,22]. As discussed in Section 6, local tissue exposure, metabolites, formulation, and interindividual variability can modify this relationship, but none should be invoked automatically to explain an apparent exposure mismatch. Improved plasma or tissue exposure likewise represents a pharmacokinetic observation and does not by itself establish molecular target engagement, pharmacodynamic benefit, or improvement in a clinically meaningful endpoint.

Clinical translation should therefore be interpreted in accordance with the exposure–target engagement–pharmacodynamic–outcome framework described in. Future clinical studies should incorporate mechanistically informative exposure and response measurements wherever feasible. Pharmacokinetic analyses should distinguish parent compounds from relevant metabolites; pharmacodynamic analyses should prioritise validated target- or pathway-proximal biomarkers; and tissue sampling, molecular imaging, or other target-accessibility measurements should be incorporated when scientifically justified and feasible. Trial design should additionally prespecify clinically meaningful primary outcomes, distinguish exploratory biomarkers and secondary endpoints from efficacy endpoints, and account appropriately for formulation, dose, disease stage, sample size, multiplicity, and statistical power. Section 6.3

The available clinical evidence supports different levels of human biological and pharmacological activity rather than a uniform demonstration of clinically meaningful polypharmacology. Berberine has comparatively stronger evidence of improvements in selected glycaemic and lipid endpoints, although the molecular mechanisms underlying these effects remain incompletely elucidated in humans [57,71,77]. Curcumin shows limited biological activity in advanced pancreatic cancer and an intermediate colorectal neoplasia signal, without established anticancer efficacy [66,75]. Green tea extract has shown a short-term lesion-response signal in oral premalignant lesions, without establishing cancer-preventive efficacy or EGCG-specific polypharmacology [76]. In Alzheimer’s disease, curcumin did not demonstrate convincing clinical or biomarker efficacy in the cited trial. In contrast, resveratrol showed CNS exposure and selected biomarker effects, along with an isolated secondary functional signal, but no consistent evidence of cognitive or global clinical benefit [69,78].

Overall, current clinical evidence does not establish that the therapeutic effects of these phytochemicals result from simultaneous engagement of the broad target networks proposed by computational or preclinical studies. More definitive mechanistic attribution will require the prospective integration of human exposure, target engagement, mechanistically proximal pharmacodynamic evidence, causal validation, and appropriately powered clinical outcome data.

Phytosomes are phospholipid-based complexes developed to improve the absorption of poorly bioavailable phytochemicals. For curcumin, the lecithin-based formulation Meriva^® has been evaluated directly in humans [23].

In a randomised crossover pharmacokinetic study, total curcuminoid absorption was approximately 29-fold greater with Meriva than with the corresponding unformulated curcuminoid mixture [23]. Importantly, this increase represented total curcuminoid absorption rather than a 29-fold increase in circulating unchanged parent curcumin. Only phase II metabolites were detectable in plasma, and the formulation altered the relative circulating profile of curcuminoid-derived species [23].

This study, therefore, demonstrates improvement at the formulation-to-pharmacokinetic level but does not demonstrate increased engagement of specific molecular targets. Improved absorption should not automatically be interpreted as evidence that the additional exposure reaches a relevant tissue in a pharmacologically active form, engages predicted targets, produces the expected pharmacodynamic response, or improves a clinically meaningful outcome.

Phytosome technology, therefore, illustrates an important translational distinction: formulation can reduce an exposure barrier without establishing the subsequent mechanistic sequence from tissue exposure to target engagement and therapeutic benefit.

Polymeric nanoparticle formulations can protect chemically unstable compounds, modify drug release, and improve oral or tissue exposure under experimental conditions [80,81]. Curcumin nanoformulations have been extensively investigated as strategies to improve delivery and pharmacological performance [80].

For EGCG, a pH-sensitive polymeric nanoparticle formulation increased oral bioavailability by more than 2.4-fold relative to EGCG powder in a rat model and was associated with reduced proteinuria and improved renal histopathological scores in nephrotic-syndrome rats [81]. These findings provide preclinical formulation, pharmacokinetic, and disease-related phenotypic evidence, but they do not establish equivalent benefit in humans, direct exposure of a defined molecular target, or target engagement.

Thus, increased nanoparticle-associated systemic exposure, tissue accumulation, or intracellular uptake should primarily be interpreted as evidence of delivery or pharmacokinetics unless accompanied by direct target-engagement measurements and mechanistically proximal pharmacodynamic validation.

Polymeric nanocarriers also face translational challenges related to formulation reproducibility, physical and chemical stability, scale-up, manufacturing control, biodistribution, safety, and regulatory translation [79,80]. These issues become particularly important when therapeutic performance depends on reproducible particle characteristics and tissue-specific disposition.

Self-microemulsifying drug delivery systems (SMEDDS) are designed to improve the dissolution and gastrointestinal absorption of poorly water-soluble compounds. A curcumin-loaded SMEDDS markedly increased curcumin solubility and dissolution and enhanced oral absorption relative to a curcumin suspension in preclinical studies [82].

These findings establish formulation and preclinical pharmacokinetic improvements, but they do not demonstrate that the increased exposure results in greater molecular target engagement, a corresponding pharmacodynamic response, or clinical efficacy. For network-derived mechanisms, the critical next question is whether formulation-induced increases in exposure translate into exposure of the relevant active molecular species at the proposed tissue and intracellular targets.

Accordingly, SMEDDS should presently be regarded as an exposure-enhancing delivery strategy, whose mechanistic and clinical significance remain dependent on downstream tissue distribution, target engagement, pharmacodynamic, safety, and efficacy studies.

Plant-derived extracellular vesicle-like nanoparticles are described using several partially overlapping terms in the literature, including plant-derived exosome-like nanoparticles and plant-derived extracellular vesicles [83,84]. These membrane-associated nanoscale structures can contain lipids, proteins, nucleic acids, and other molecular cargo and are increasingly being investigated as biogenic delivery platforms because of their potential for cellular uptake, cargo protection, and tissue delivery [83,84].

Their application to central nervous system delivery has attracted particular interest. Reviews of plant-derived vesicle-like systems describe their potential for blood–brain barrier transport and CNS drug delivery, although the supporting evidence remains predominantly preclinical [83].

More directly, a recent preclinical study evaluated plant-derived extracellular vesicles loaded with itraconazole using glioblastoma-cell assays and an in vitro blood–brain barrier permeability model [85]. Ginger- and aloe-derived vesicles demonstrated permeability in that experimental system. Because itraconazole is not itself a phytochemical and the BBB evidence was obtained using an in vitro model, this study supports the feasibility of the plant-derived vesicle carrier platform rather than demonstrating CNS delivery of a phytochemical, in vivo brain exposure, or molecular target engagement [85].

These findings should therefore remain classified as predominantly preclinical evidence of delivery. Demonstration of cellular uptake or barrier permeability does not establish that an encapsulated compound reaches an intracellular molecular target at an exposure sufficient to produce a pharmacological effect. In CNS applications, BBB passage, brain exposure, target engagement, pharmacodynamic response, and therapeutic efficacy must be regarded as separate evidentiary stages.

Important translational challenges include standardisation of isolation and characterisation procedures, particle heterogeneity, reproducibility, cargo definition, stability, scalable production, biodistribution, immunological effects, and long-term safety [83,84]. Comparative superiority over established delivery systems, therefore, requires direct evaluation of pharmacokinetic, pharmacodynamic, safety, and efficacy rather than inference from cellular uptake or barrier-crossing capability alone.

Within the evidence evaluated in this review, none of the delivery platforms discussed above has been shown to complete the full progression from formulation optimisation to a clinically meaningful outcome for phytochemical polypharmacology. Human pharmacokinetic evidence is available for the curcumin phospholipid formulation Meriva^®, which substantially increased total curcuminoid absorption relative to an unformulated preparation [23]. By contrast, the polymeric EGCG nanoparticle and curcumin SMEDDS studies included in the present bibliography provide primarily preclinical evidence for improved oral exposure and related biological effects [81,82]. At the same time, plant-derived extracellular vesicle-like systems remain emerging and predominantly preclinical approaches [83,84,85].

These platforms should therefore not be compared solely according to fold increases in apparent bioavailability, intracellular uptake, or theoretical targeting potential. Their interpretation should instead reflect the furthest experimentally demonstrated stage of translation and consider biodistribution, metabolite exposure, selectivity, and tolerability, particularly because increased exposure may also increase off-target pharmacology or safety liabilities [79].

From a network-pharmacology perspective, advanced delivery systems are best regarded as exposure-modifying technologies rather than evidence of a polypharmacological mechanism. Their translational value depends on whether improved exposure ultimately supports pharmacologically relevant tissue concentrations, target engagement, appropriate pharmacodynamic activity, acceptable safety, and clinically meaningful benefit.

Artificial intelligence and machine-learning approaches are increasingly being applied to drug–target prediction and systems-level target prioritisation [86,87]. Different approaches use distinct molecular representations. DeepDTA encodes ligand SMILES strings and protein amino-acid sequences using convolutional neural networks to predict drug–target binding affinity, whereas GraphDTA introduced graph-based molecular representations to capture ligand topology [88,89]. Predicted affinities from either approach remain computational estimates rather than experimentally measured binding affinities or evidence of cellular target engagement.

Graph neural networks, multimodal approaches, and knowledge graphs can further integrate molecular and biological information for target identification or drug repurposing applications [86,87,90]. Such integration can improve candidate prioritisation, but prediction quality remains dependent on the underlying data, molecular representation, model assumptions, and biological coverage [24,86].

Particular caution is required when interpreting model performance mechanistically. Training–test overlap, biased datasets, limited applicability domains, and other validation problems can lead to apparently strong predictive performance that may not generalise prospectively [24]. AI predictions should therefore be evaluated using appropriately separated validation data and, ultimately, prospective experimental testing. Model explanations or feature-attribution methods may help identify factors contributing to a prediction, but they do not establish biological causality.

In phytochemical pharmacology, AI-derived targets are therefore best regarded as candidates for prioritisation. Their translational relevance should subsequently be assessed in relation to compound identity, structural plausibility, pharmacologically achievable exposure, existing experimental evidence, and the feasibility of direct validation.

Predicted protein structures provide an additional resource for evaluating computationally proposed compound–protein interactions. The AlphaFold Protein Structure Database substantially expands access to predicted protein structures, including proteins for which experimentally determined structures may be unavailable [91]. Such structures can support computational hypothesis generation, but they do not demonstrate that a phytochemical binds the predicted protein.

Structure-based predictions are also sensitive to protein conformation. Proteins are dynamic rather than rigid structures, and docking against a single conformation may not represent the full range of ligand-accessible states. Ensemble docking provides one strategy for incorporating receptor flexibility by evaluating ligands across multiple protein conformations [25]. Accordingly, a favourable docking pose or a stable molecular dynamics trajectory should be interpreted as increased structural plausibility rather than as proof of molecular binding or target engagement, and should be strengthened by orthogonal experimental evidence [19].

CRISPR-based genetic screening and perturbation provide complementary approaches for identifying genes that influence cellular phenotypes or drug responses [30,31]. However, a gene identified through a CRISPR screen should not automatically be classified as the direct molecular target of a phytochemical. Genetic dependency may arise through upstream or downstream signalling, transport, metabolism, stress responses, or other cellular processes [31].

Genetic dependency and physical target engagement, therefore, address different mechanistic questions. Confidence in direct target attribution increases when genetic evidence converges with biochemical or biophysical interaction data and cellular target-engagement measurements [19,31]. CETSA, for example, assesses compound-induced changes in protein thermal stability in cells or tissues and can provide evidence of target engagement in a biological environment [26]. Rescue experiments, resistant mutants, structure–activity relationships, and complementary chemical-biology approaches can provide additional functional support where appropriate [19].

These approaches provide complementary forms of evidence within the hierarchy defined in: biochemical and biophysical studies assess molecular interactions; target-engagement methods assess interactions in biological systems; and genetic perturbation evaluates functional dependence. Genetic dependency should therefore not be considered equivalent to direct molecular binding. Section 3.2

Single-cell RNA sequencing enables transcriptome profiling at cellular resolution and can resolve molecular heterogeneity that is obscured in bulk-tissue measurements [92,93]. When applied to pharmacological experiments, this resolution can be used to identify cell populations exhibiting distinct transcriptional responses to treatment.

Spatial transcriptomics extends molecular profiling by retaining anatomical information within tissues, while newer spatial multi-omics approaches can integrate multiple molecular layers and characterise spatially organised biological states [94,95,96]. These technologies can therefore help determine where and in which cellular populations treatment-associated molecular responses occur.

Their outputs should not, however, be interpreted as direct evidence of target engagement. Changes in transcript abundance, pathway scores, or spatial molecular patterns are generally downstream observations and do not independently identify the molecular event responsible for the response. Likewise, computationally inferred interactions derived from expression data remain hypotheses until experimentally tested.

The principal contribution of single-cell and spatial technologies to phytochemical pharmacology is therefore improved biological and anatomical resolution. Candidate mechanisms identified through these approaches can subsequently be examined using cell-type-specific perturbation, protein-level measurements, biochemical validation, or target-engagement approaches. Thus, single-cell and spatial analyses can refine mechanistic hypotheses but do not independently establish direct molecular engagement or therapeutic efficacy.

Physiologically based pharmacokinetic (PBPK) modelling provides a quantitative framework for integrating compound-specific pharmacokinetic properties with physiological parameters to predict systemic and tissue exposure. Its application to dietary phytochemicals is particularly relevant because experimentally proposed mechanisms may occur at concentrations that differ substantially from achievable systemic or tissue exposure [70].

PBPK modelling can complement network pharmacology by testing whether predicted compound–target relationships are pharmacokinetically plausible. Where sufficient data are available, candidate mechanisms can be evaluated in relation to dose, absorption, metabolism, tissue distribution, metabolite formation, and estimated exposure at relevant anatomical sites [70].

However, model predictions are dependent on the quality and completeness of the underlying physiological and compound-specific parameters. Uncertainty in absorption, metabolism, protein binding, tissue partitioning, or other inputs can propagate into predicted exposure. PBPK predictions should therefore be compared with measured pharmacokinetic data whenever possible and should complement rather than replace experimental measurements [70].

Integration of exposure modelling with network, structural, and omics information may ultimately help prioritise mechanisms that are both biologically plausible and compatible with achievable exposure. At present, however, the appropriate objective is not automated prediction of therapeutic efficacy but quantitative assessment of whether a proposed mechanism is pharmacokinetically plausible and experimentally testable.

The complementary roles, principal limitations, recommended validation strategies, and evidentiary positions of these computational, experimental, pharmacokinetic, and clinical approaches are summarised in Table 5.

Collectively, these technologies should be regarded as complementary components of an evidence-building framework rather than interchangeable forms of mechanistic validation. AI and network approaches prioritise hypotheses; structural methods refine molecular plausibility; single-cell and spatial technologies improve cellular and anatomical resolution; genetic perturbation strengthens functional attribution; target-engagement approaches test molecular interaction in biological systems; and PBPK modelling imposes quantitative exposure constraints [19,26,70,86].

The translational objective is to connect these approaches into an evidence chain extending from chemical identity and computational prediction to molecular engagement, functional contribution, exposure compatibility, pharmacodynamic response, and clinical outcome. Such integration may reduce unsupported target assignments and improve discrimination between credible, exposure-compatible polypharmacology and computationally inferred pathway convergence.

Medicinal chemistry optimisation offers a potential strategy for addressing chemical and pharmacokinetic limitations of natural products while retaining desirable pharmacological properties [11]. Curcumin provides a relevant example because its chemical instability, reactivity, and pharmacokinetic limitations have stimulated extensive investigation of structural analogues and modified derivatives [12].

For berberine, structural modification can likewise alter pharmacokinetic and pharmacological behaviour. Dihydroberberine, for example, has been investigated as a more bioavailable berberine derivative and retains effects on mitochondrial complex I and downstream AMPK-associated signalling [44]. Such findings illustrate how chemical modification can alter exposure while preserving selected elements of the parent compound’s pharmacology.

More broadly, natural-product scaffolds provide starting points for medicinal-chemistry optimisation, while designed multiple-ligand strategies provide a conceptual framework for engineering activity across more than one therapeutically relevant target [2,8,11]. The objective should not, however, be to maximise biochemical potency or preserve the largest possible number of predicted targets. A more potent derivative may remain translationally limited if metabolic stability, tissue exposure, or target accessibility is inadequate, whereas reducing chemically reactive or therapeutically irrelevant interactions may improve pharmacological selectivity.

Network pharmacology may assist this process by prioritising disease-associated targets and molecular modules for experimental evaluation [1,5]. A more rigorous strategy would combine computational prioritisation with structure–activity relationships, orthogonal biochemical assays, cellular target engagement, functional perturbation, and PK/PD assessment. The goal is therefore evidence-guided multi-target optimisation rather than preservation of broad molecular promiscuity.

Inter-individual pharmacokinetic and biological variability may contribute to heterogeneous responses to phytochemicals. Human studies of berberine, for example, demonstrate substantial inter-individual pharmacokinetic variability, while both clinical and mechanistic studies indicate that gut-microbial composition and metabolism can alter berberine disposition and metabolic responses [45,57,68].

Biological variability alone, however, does not establish a clinically useful predictive biomarker. Future studies should therefore move beyond descriptive molecular signatures toward prospectively validated biomarkers linked to exposure, pharmacodynamic response, and clinically meaningful outcomes. Candidate biomarkers may include metabolite profiles, pathway-associated molecular signatures, genetic features, microbiome-derived variables, or tissue-specific molecular patterns, but their predictive value should be assessed prospectively and independently.

The microbiome is particularly relevant because microbial metabolism can alter phytochemical disposition and generate molecular species with absorption or biological properties that differ from those of the administered parent compound [45,58]. Associations between microbial taxa and treatment response should therefore be distinguished from causal mediation. Stronger evidence can be generated using longitudinal metabolite measurements together with experimental approaches such as gnotobiotic models, defined microbial communities, microbial-transfer experiments, or targeted manipulation of candidate organisms [58].

Precision approaches should also distinguish total systemic concentration from pharmacologically relevant target-site exposure. Tissue distribution, intracellular disposition, metabolite formation, protein binding, and target accessibility can all modify the relationship between administered dose and molecular response. Accordingly, a biomarker-defined subgroup should not be considered pharmacologically responsive solely because it expresses a pathway-associated signature; stronger evidence would demonstrate that the biomarker predicts exposure-compatible target engagement, a proximal pharmacodynamic response, or a clinically meaningful outcome.

The future role of network pharmacology should increasingly shift from broad target enumeration to the integrated prioritisation of exposure-compatible and causally testable targets or combinations thereof. AI and multi-omics approaches can contribute molecular and disease-context information, while PBPK modelling can impose quantitative constraints by estimating whether relevant exposure is achievable within specific physiological compartments [70,86,87].

These approaches should be regarded as components of an evidence-building framework rather than automated mechanism-discovery systems. Network enrichment and topology are influenced by database and annotation coverage, whereas AI predictions depend on training data, model assumptions, applicability domains, and validation strategy [5,24]. Pharmacokinetic constraints should therefore be incorporated during target prioritisation by considering systemic and tissue exposure, metabolite formation, intracellular accessibility, target abundance, biochemical potency or affinity, exposure duration, and the relationship between exposure and pharmacodynamic response [70].

Consistent with the translational categories defined in Section 6.3, targets that are computationally plausible but incompatible with available exposure data should receive lower translational priority unless local exposure or active metabolites provide quantitative support. Conversely, convergence of exposure compatibility with molecular, functional, or target-engagement evidence should increase confidence in a proposed mechanism. A smaller network of experimentally supported, exposure-compatible, and causally relevant targets may therefore provide greater translational information than a larger network dominated by computational associations. Candidate mechanisms can subsequently be advanced through the integrated validation framework summarised in Figure 2.

When combinations of compounds or targets are investigated, additive, synergistic, antagonistic, and compensatory effects should be experimentally distinguished. Formal interaction frameworks such as Bliss independence or Loewe additivity should be used where appropriate, rather than inferring synergy from pathway convergence alone [27].

Clinical development should remain proportional to the level of mechanistic evidence available. Early-phase studies need not establish every component of a proposed molecular mechanism, but hypotheses carried into clinical development should, where feasible, be linked to measurable exposure and pharmacodynamic endpoints. Formulation, dose, parent-compound and metabolite exposure, tissue accessibility, biomarker status, and clinical outcomes should be evaluated together according to the translational framework described in. Section 6.3

Similarly, computationally prioritised targets should not be promoted to clinical biomarkers solely based on network centrality, docking scores, AI confidence, pathway enrichment, or altered gene expression. Target attribution is strengthened when evidence of biochemical or cellular engagement converges with functional perturbation [5,19,26]. Future clinical studies should therefore use evidence-tiered designs in which mechanistic claims remain proportional to supporting evidence and predefined pharmacological hypotheses are evaluated alongside clinically meaningful endpoints.

The principal opportunity for next-generation phytochemical pharmacology is not to generate increasingly complex interaction networks, but to determine which proposed mechanisms are experimentally supported, pharmacologically accessible, causally relevant, and therapeutically meaningful.

AI and network methods can prioritise hypotheses [86]; structural and biochemical approaches can refine molecular plausibility; single-cell and spatial technologies can improve cellular and anatomical resolution [95]; genetic perturbation can strengthen functional attribution; target-engagement approaches can test molecular interaction in biological systems [19,26]; advanced delivery systems can modify exposure; and PBPK modelling can assess whether proposed mechanisms are compatible with achievable systemic or tissue concentrations [70]. These approaches are complementary and should be interpreted according to the evidentiary questions each can address.

Future progress, therefore, requires a transition from network-rich descriptions to evidence-rich mechanisms. Precision phytopharmacology should not be defined by the number of targets assigned to a compound, but by the identification of targets or target combinations that are pharmacologically accessible and supported by convergent molecular, functional, pharmacokinetic, pharmacodynamic, and clinical evidence. Such an exposure-aware and evidence-graded strategy provides a more rigorous basis for translating systems-level phytochemical pharmacology into reproducible and clinically relevant mechanisms.

No new data were created or analyzed in this study. Data sharing is not applicable to this article.