Advances in Herbal Research

1. Introduction

Cancer, taken as a whole, is not slowing down. GLOBOCAN’s most recent estimate puts the 2024 burden at roughly 20.6 million new cases and 9.8 million deaths worldwide, with 34.4 million cases projected by 2050 (Sung et al., 2026) — and, tellingly, the steepest proportional rise is expected in lower-HDI countries, where diagnostic infrastructure, treatment affordability, and registry completeness are often already stretched thin (Sung et al., 2026; Pramesh et al., 2022). It is worth sitting with that asymmetry for a moment: the places least equipped to absorb a growing cancer burden are, on present trends, the places where it will grow fastest. That alone would be reason enough to keep searching for therapeutics that are not only effective but scalable and affordable — natural products and repurposed approved drugs being two obvious candidates on that front.

But there is a deeper reason the search has moved toward networks rather than single targets. Cancer is not really one disease with one broken part; it is closer to an evolving ecosystem, shaped simultaneously by genomic alteration, transcriptional rewiring, shifting metabolism, immune surveillance (or evasion), treatment pressure, and the idiosyncrasies of the host. A one-drug/one-target logic can still work, of course — it has produced real drugs — but it struggles to capture how a compound might act on several connected nodes of that ecosystem at once. Network pharmacology, and its close relative network medicine, were built partly to address exactly this gap, representing diseases and therapies as modules embedded within protein-interaction, regulatory, metabolic, drug-target, or heterogeneous knowledge networks (Hopkins, 2008; Barabási et al., 2011; Nogales et al., 2022).

Two candidate classes sit at the center of this review, and they are not, on reflection, equally straightforward. Approved drugs arrive with an enormous head start: known human pharmacokinetics, established manufacturing, and — perhaps most valuable of all — a track record of safety in some population. What they lack is any guarantee that this safety profile survives translation into oncology doses, novel combinations, or patient groups quite different from the ones the drug was originally tested in. Natural products sit at almost the opposite pole. Their appeal lies in chemical diversity and, often, genuine polypharmacology — a single extract may contain dozens of bioactive constituents acting on overlapping pathways. Their trouble is that this same diversity makes composition, standardization, target annotation, and bioavailability considerably harder to pin down (Fang et al., 2018; Noor et al., 2022; Panossian, 2025).

The methodological toolkit for connecting either candidate class to a network has also grown substantially, and unevenly. Interactome-proximity methods, disease-module localization, expression-reversal signatures, constraint-based metabolic models, co-expression networks, multilayer graphs, graph neural networks, and knowledge graphs have all, in one study or another, been used to prioritize candidates (Guney et al., 2016; Cheng et al., 2018; Cheng et al., 2019; Zeng et al., 2019; Zitnik et al., 2018). What tends to get less attention — understandably, perhaps, since it is less exciting to report — is how sensitive these outputs are to the completeness of the underlying interactome, to literature bias toward well-studied proteins, to which database version happened to be used, and to subtle leakage between the data a model is trained on and the data it is tested against. Reproducibility, in turn, depends on something almost mundane: whether the code, parameters, identifiers, and software environment were actually reported in enough detail for someone else to rebuild the analysis (Schaduangrat et al., 2020).

So the question this review keeps returning to is not really whether a network can produce a plausible-looking candidate list — by now, it clearly can, repeatedly, across many cancers and many methods. The harder and more useful question is which of these workflows produce candidates that are stable across databases, supported by data the model never saw, causally validated rather than merely correlated, and realistic enough pharmacologically to matter in a clinic. That distinction, more than any single algorithm, is what this review is built to make explicit.

2. Network Pharmacology in Cancer Drug Repurposing: Comparing Approved Drugs and Natural Products

2.1 From single targets to disease modules

It has become something of a truism in systems biology that disease-associated genes rarely scatter randomly across a molecular network; instead, they tend to cluster in identifiable neighborhoods, a pattern that underlies both the disease-module concept and the idea of network proximity (Barabási et al., 2011; Guney et al., 2016). If that clustering is real — and the evidence, on balance, suggests it is — then a drug has several ways of reaching a disease module, not just one. It might act directly on a disease protein, sit near the module without touching it directly, intervene on a compensatory pathway that the tumor has come to depend on, or exploit a vulnerability that only becomes essential within a particular molecular subtype. These are genuinely different mechanisms, and treating them as interchangeable — which the literature sometimes does, if only implicitly — risks flattening distinctions that matter for how much confidence a prediction deserves.

2.2 Two candidate classes, two very different evidentiary burdens

For approved drugs, the case for repurposing rests substantially on what is already known: human exposure data, manufacturing pathways, and a regulatory history. That is a real advantage. It is not, however, a free pass. The risks cluster around indication-specific dosing, pharmacokinetic mismatch when the drug is pushed into a new context, interaction toxicity in combination regimens, and — a less scientific but still practical concern — weak commercial incentive to pursue a repurposed indication once the original patent has expired. Natural products invert this picture almost entirely. Their opportunity is chemical and mechanistic diversity, sometimes genuine polypharmacology across several oncogenic pathways at once. Their risk, though, is that this very diversity can obscure what is actually being tested: uncertain composition, target predictions that are more promiscuous than causal, unmeasured or unrealistic exposure, batch-to-batch variation, and — a subtle but recurring problem — the conflation of a single isolated molecule with the crude, multi-component formulation it was extracted from (Fang et al., 2018; Noor et al., 2022; Panossian, 2025).

2.3 Where the existing reviews stop short

Several prior reviews have already mapped pieces of this landscape — computational oncology repurposing broadly, comparisons of network-based methods against one another, medicinal-plant network pharmacology specifically, and natural-product polypharmacology as its own subfield (Turanli et al., 2021; Palve et al., 2021; Cavalcante et al., 2024; Fiscon et al., 2022). A 2024 scoping review of in silico oncology repurposing, covering 238 studies through 2021, found that most were computational-only and that molecular modeling, if anything, was used more often than explicitly network-based approaches (Cavalcante et al., 2024). What none of these reviews quite does, though, is put natural products and approved drugs side by side under one shared appraisal standard — network construction quality, external validity, experimental causality, reproducibility, translational maturity — so that a reader can judge a colorectal-cancer herbal formula and a repositioned kinase inhibitor by the same yardstick. That gap is the one this review tries to close.

2.4 A taxonomy of network-driven computational strategies

The methods used to connect a network to a candidate are not one method but a small family of related ones, and it helps — perhaps more than the literature usually acknowledges — to keep them conceptually separate rather than lumping everything under “network pharmacology.” Interactome-proximity approaches ask whether a drug’s targets sit near disease proteins in the human interactome; Guney and colleagues (2016) showed that known therapeutic effects do tend to localize near disease modules, a finding that Cheng and colleagues (2018) later extended to large-scale, population-validated repurposing, and then refined further with a genome-wide positioning algorithm (Cheng et al., 2019). These methods are appealingly interpretable, but they inherit whatever gaps exist in the interactome itself, along with target-annotation bias. Expression and co-expression networks take a different route, searching for compounds that reverse a disease signature or perturb an influential transcriptional module — used, for instance, to prioritize hepatocellular-carcinoma candidates through gene co-expression modules (Yuan et al., 2022) and cervical-cancer candidates through HPV-host interaction networks (Ahmed et al., 2023). Metabolic and constraint-based models go after a different kind of selectivity altogether, looking for vulnerabilities that are essential in tumor tissue specifically but dispensable in matched normal tissue (Pacheco et al., 2019). Heterogeneous networks, knowledge graphs, and graph neural networks integrate drugs, targets, diseases, and phenotypes simultaneously — deepDR, for example, learned drug representations across ten linked networks (Zeng et al., 2019), while graph-convolutional models have captured multi-relational polypharmacy effects (Zitnik et al., 2018) — at the cost of interpretability and, more worryingly, a real vulnerability to information leakage when related drugs or diseases appear on both sides of a train/test split. Finally, combination-discovery methods try to cover complementary disease modules or exploit synthetic lethality while keeping overlapping toxicity in check, as in a multi-cancer framework that integrated disease

Figure 1. Taxonomy of network-driven computational strategies used to prioritize anticancer candidates. This figure groups the methods encountered in this review into five families — interactome proximity, expression/co-expression networks, metabolic and constraint-based models, heterogeneous networks/knowledge graphs/deep learning, and combination discovery — alongside a representative citation, a characteristic strength, and a characteristic risk for each family. The shared question beneath all five families, shown in the lower panel, is the appraisal test this review applies uniformly: did the network genuinely reshape which candidates were prioritized, or was it introduced afterward to illustrate a decision already made on other grounds? Readers should use this figure as an orientation map before interpreting Tables 8 and 9, which classify individual studies by exactly these method families.

Figure 2. The network-to-translation conceptual framework guiding this review’s appraisal logic. Candidate credibility is modeled here as a chain of five linked stages — data inputs, network model construction, prioritization, validation, and translation — with a cross-cutting appraisal band beneath covering data provenance, leakage, biological context, code transparency, comparator calibration, and equity/feasibility. The framework’s central claim, elaborated in Section 2.5 and Table 1, is that a weak link at any stage cannot be compensated for by strength elsewhere in the chain, however visually elaborate the network diagram in a given paper may be. This figure underlies the eligibility and extraction decisions described in Sections 3.5 through 3.7.

perturbation with drug mechanism and chemistry (Federico et al., 2022). Figure 1 summarizes this taxonomy alongside the strength and principal risk of each family; what unites all five, and what this review keeps asking of every study it encounters, is a single appraisal question — did the network genuinely change which candidates rose to the top, or was it added afterward to explain a choice that had already been made on other grounds?

2.5 A conceptual framework linking data to translation

Because “network-driven” can mean almost anything if left undefined, this review adopts an explicit chain-of-evidence framework (Figure 2): data provenance shapes how a network is built; network construction shapes which candidates get prioritized; prioritization, on its own, is not evidence — it requires independent and, ideally, causal validation; and translation on top of that requires realistic exposure, demonstrated safety, clinical utility, and a plausible path to access and implementation. The framework is deliberately unforgiving on one point: a weak link anywhere in that chain cannot be rescued by a visually elaborate network diagram elsewhere. A beautiful figure showing hundreds of nodes and edges says very little about whether the underlying data were current, whether the prediction was tested on data the model never saw, or whether anyone measured the compound in the tissue that mattered. Table 1 operationalizes the framework’s key constructs — network-driven, natural product, approved drug, targeted therapy, translational readiness — into definitions this review applies consistently during appraisal, so that, for example, a study that builds a network only after candidates have already been chosen is excluded on principle rather than judged case by case.

2.6 Evidentiary pathways illustrated: approved drugs and natural products

It is one thing to describe these method families in the abstract; it is another to see how far a given study actually travels along the validation chain once it starts. Figure 3 traces this for approved-drug repurposing, using published examples that move — with varying success — from network signal, through candidate ranking, to orthogonal biochemical or cellular signals, and in the stronger cases through causal or in vivo testing. Ceritinib’s noncanonical polypharmacology, identified through integrated functional proteomics, was followed by phenotypic screening, proteomic confirmation, and combination testing (Kuenzi et al., 2017) — a genuinely strong chain by the standards of this literature, though still short of clinical evidence. Maprotiline’s route to PD-L1 modulation via SPOP, by contrast, reached mechanistic and in vivo combination evidence with anti-CTLA4 therapy but stops there for now, with dose, safety, and efficacy in humans still unresolved (Tian et al., 2025). Natural products travel a parallel but distinct pathway (Figure 4): constituent identification, target prediction, network-disease overlap, and then — this is the step that seems to make or break credibility — a chemical and exposure check confirming that the predicted constituent is actually present, and at a plausible concentration, in the material that was tested. When that check happens, as with Tricin from Weijing decoction (Li et al., 2021) or triptolide’s transcriptional targets (Seo et al., 2021), the resulting evidence chain looks considerably more trustworthy than when a promising compound is simply asserted to be present in an extract on the strength of a database entry alone (Fang et al., 2018; Panossian, 2025).

3. Methods

3.1 Design and reporting standards

This review will be conducted as a protocol-led systematic review, reported in line with the PRISMA 2020 statement (Page et al., 2021), with search reporting following the PRISMA-S extension (Rethlefsen et al., 2021) and — a step that is easy to skip but shouldn’t be — peer review of the MEDLINE strategy using the PRESS checklist before a single record is retrieved (McGowan et al., 2016). Where quantitative pooling is not appropriate, which we expect for a meaningful share of this literature, findings will be reported using Synthesis Without Meta-analysis (SWiM) principles (Campbell et al., 2020) rather than forced into a summary statistic that would misrepresent heterogeneous evidence. The protocol will be posted on the Open Science Framework before screening begins and, if the eligibility scope meets current registration criteria, submitted to PROSPERO as well. We flag this order deliberately: registration precedes screening, not the other way around.

3.2 Review framework (PICOS)

Table 2 sets out the review’s PICOS framework in full. Briefly, the population comprises human cancers, cancer-derived experimental systems, patient-derived material, or computational cancer datasets; the index approach is

Figure 3. Evidentiary pathway traced by illustrative approved-drug repurposing studies. This figure follows eight published approved-drug candidates through five sequential evidentiary stages — network signal, candidate ranking, orthogonal biochemical or cellular signal, causal or in vivo testing, and clinical readiness — using ceritinib, maprotiline, sitagliptin, clindamycin, and related examples discussed in Section 4.1 and tabulated fully in Table 8. Every example reviewed reaches at least the orthogonal-signal stage, and several reach causal or in vivo testing, but none yet reaches confirmed clinical readiness, which is the pathway’s intended message: further rightward movement, not the existence of a network prediction itself, is what should drive translational confidence.

Figure 4. Natural-product evidence chain from constituent identification to causal validation. This figure decomposes the standard natural-product network-pharmacology workflow into five sequential steps and then splits the outcome into two contrasting patterns observed across the literature reviewed in Section 4.2 and Table S1. The left-hand branch (“confirmed chain”) lists examples — Tricin/Weijing decoction, triptolide, nitidine chloride, and Xianlian Jiedu decoction — where the predicted constituent was chemically confirmed in the administered material and carried through to causal or in vivo validation. The right-hand branch (“stalled at docking/enrichment”) describes the more common pattern in the wider literature, where a constituent is computationally predicted but never measured in what was actually tested, leaving the evidence chain incomplete at an early stage.

Table 1. Operational definitions applied to the review’s key constructs. This table defines, for consistent use throughout screening and extraction, what this review means by a network-driven study, a natural product, an approved drug, a targeted therapy, and translational readiness, together with the practical implication each definition carries for inclusion or interpretation. The definitions are designed to prevent common ambiguities — for example, distinguishing a study in which the network materially shaped candidate selection from one in which a network diagram was added only after the candidate had already been chosen.

Table 2. PICOS review framework. This table specifies the population, index approach, comparators, outcomes, timing, and setting that define this review’s scope, following the standard PICOS structure recommended for systematic-review protocols. It is intended to be read alongside the eligibility criteria in Table 3, which operationalize these same elements into explicit include/exclude rules.

network-driven prioritization of natural products or approved drugs for treatment, sensitization, resistance reversal, immunomodulation, or rational combination therapy; comparators include known indications, standard therapies, negative controls, random or network null models, alternative algorithms, or — where justified and stated explicitly — no comparator at all; outcomes span candidate rank, target/pathway identification, predictive performance, external validation, target engagement, and preclinical through clinical response measures; and the timing window runs from 1 January 2015 through the final search date in September 2026, chosen to capture the modern network-medicine era without reaching so far back that database versions become impossible to reconstruct.

3.3 Information sources

Six information-source categories will be searched, each serving a slightly different purpose. MEDLINE via PubMed provides biomedical indexing, MeSH vocabulary, and the bulk of primary studies; Embase adds broader pharmacology coverage, conference abstracts, and Emtree indexing with a stronger European footprint; Scopus and Web of Science Core Collection cover the interdisciplinary computational literature and enable citation tracking; the Cochrane Library is searched for relevant intervention reviews and controlled-trial records, though we expect its yield here to be modest given how young this specific literature is; ClinicalTrials.gov, the WHO ICTRP, EU CTIS, and regulatory labels are consulted for translational follow-up on prioritized candidates — understood as a supplement to peer-reviewed evidence, never a substitute for it; and Google Scholar’s first 200 relevance-ranked records will be used for citation chasing and for surfacing hard-to-index computational work, with the date, browser state, and inherent ranking limitations of that search documented explicitly. Backward and forward citation searching of included studies and key reviews rounds out the source list, alongside retraction checks through PubMed and Retraction Watch wherever available.

3.4 A fully reproducible PubMed search strategy

Because reproducibility is as much a methods question as a results question, the complete PubMed strategy is reported here verbatim, exactly as it would be entered, rather than paraphrased:

(“Neoplasms” OR neoplasm* OR cancer*[tiab] OR oncolog*  OR tumor* OR tumour* OR carcinoma* OR leukemia* OR leukaemia* OR lymphoma* OR melanoma* OR sarcoma* AND (“network pharmacology” OR “network medicine” OR “systems pharmacology” OR “network-based” OR “network driven” OR interactom* OR “protein-protein interaction network* OR “drug-target network* OR “disease-gene network* OR “gene co-expression network* OR “knowledge graph* OR “network proximity” OR “disease module* OR “graph neural network* AND (“Drug Repositioning” OR repurpos* OR reposition* OR “approved drug* OR “existing drug* OR “Natural Products” OR “Plants, Medicinal” OR “Phytotherapy” OR “natural product*” OR phytochemical* OR botanical* OR herbal OR “medicinal plant* OR “traditional medicine” OR “bioactive compound*” AND (“2015/01/01” to “2026/09/01)

No human, animal, clinical-trial, or language filter will be applied at the point of retrieval, since PubMed’s built-in filters can inadvertently strip out exactly the computational and preclinical records this review needs most. The final search date must, of course, replace “2026/09/01” if execution slips later into the month, and database-specific controlled vocabulary and proximity operators will be translated for Embase, Scopus, and Web of Science rather than copied across verbatim — a shortcut that looks efficient but reliably degrades sensitivity. Appendix A reports the exact pilot query already executed for feasibility testing, together with its unscreened, undeduplicated PubMed count, so that any discrepancy between pilot and final searches is auditable rather than hidden.

3.5 Eligibility criteria

Table 3 lists eligibility criteria across six domains — publication type, cancer context, network role, candidate type, study design, and outcome — and Appendix B provides the full-text exclusion-reason hierarchy that will be applied, in a fixed order, to every excluded record. The domain most likely to require reviewer judgment, and the one we flag for extra calibration during pilot screening, is network role: a study earns inclusion only if the network materially informs candidate, target, or combination ranking, not if a PPI diagram or Cytoscape figure is added after the candidate has already been selected on other grounds.

3.6 Study selection and record handling

Every record will be exported with full citation metadata, abstract, DOI, and PMID or Embase ID into both a reference manager and a review platform, after which

Table 3. Eligibility criteria across six domains. This table sets out what will be included and excluded across publication type, cancer context, network role, candidate type, study design, and outcome, forming the operational rulebook applied during title/abstract and full-text screening. The network-role domain is expected to require the most reviewer calibration, since it distinguishes a study in which the network drove candidate selection from one in which a network figure was added afterward.

Table 4. Data-extraction domains and fields. This table lists the ten structured domains — bibliographic, cancer context, candidate, input data, network, algorithm, output, validation, reproducibility, and translation — that two independent reviewers will populate for every included study, as described in Section 3.7. The domains are designed to capture not only what a study found, but how reproducibly and how causally it arrived there.

Table 5. Risk-of-bias tool assignment by study component. This table matches each type of study component encountered in this literature — computational network models, natural-product characterization, in vitro/organoid work, animal studies, non-randomized human data, and randomized trials — to the appraisal tool this review will apply, together with the key domains each tool covers. The computational-network and natural-product rows use the purpose-built NDDA instrument (detailed further in Table 6), which is explicitly informed by, but not a validated equivalent of, PROBAST.

deduplication proceeds in three passes — first by PMID/DOI, then by exact title match, and finally by fuzzy title combined with author and year — with a source log retained throughout so that each database’s contribution to the final pool stays auditable. Two reviewers will independently pilot-test the eligibility criteria on at least 50 records, refine the screening instructions if needed (without altering the registered question itself), and then proceed to screen all titles, abstracts, and full texts independently. Disagreements will be resolved by consensus or, failing that, by a third reviewer, and every full-text exclusion will be logged against a single primary reason. Reports that cannot be retrieved will go through institutional access channels, document delivery, and two author-contact attempts spaced at least seven days apart before being parked in an “awaiting classification” category — deliberately not silently excluded. Companion papers describing the same underlying study will be linked and analyzed as one unit, and corrections, expressions of concern, and retractions will be checked before both data extraction and manuscript submission. Figure 5 presents the PRISMA 2020 flow-diagram template that will record these counts once screening is complete; the counts are intentionally blank at this protocol stage, and we say so plainly rather than let a populated-looking diagram imply results that do not yet exist.

3.7 Data extraction

A structured extraction form (Table 4) captures ten domains for every included study: bibliographic details, cancer context, candidate characteristics, input data provenance, network construction parameters, algorithmic details, output and ranking information, validation level, reproducibility indicators, and translational plausibility. Two reviewers will extract independently, cross-check discrepancies against the source text, and resolve disagreements by consensus.

3.8 Risk-of-bias and quality appraisal

No single validated instrument currently covers the full span of network-driven drug-discovery studies, which is precisely why a modular strategy is needed rather than a single off-the-shelf checklist. For computational network models, this review applies a purpose-built instrument — provisionally named the Network-Drug Discovery Appraisal, or NDDA — that draws its structural logic from PROBAST (Wolff et al., 2019) without pretending to be a validated substitute for it. Table 5 assigns the appropriate tool to each study component (computational network model, natural-product characterization, in vitro/organoid work, animal work, non-randomized human data, and randomized trials), and Table 6 defines explicit low- and high-risk signals across the NDDA’s seven domains — data provenance, network construction, analysis, validation, exposure and mechanism, reproducibility, and reporting/conflicts — so that two reviewers scoring the same study should, in principle, converge on the same judgment. Animal components will be appraised with the SYRCLE risk-of-bias tool (Hooijmans et al., 2014), using ARRIVE 2.0 (Percie du Sert et al., 2020) to assess reporting completeness specifically, since reporting quality and risk of bias are related but distinct questions that this review is careful not to conflate. Non-randomized human studies will use ROBINS-I (Sterne et al., 2016), and any randomized evidence will use RoB 2 (Sterne et al., 2019).

3.9 Evidence synthesis and, where appropriate, meta-analysis

The primary synthesis stratifies studies by candidate stream (approved drug, isolated natural product, or extract/formula), cancer type and subtype, network method family, and the highest validation level reached — deliberately avoiding vote-counting by statistical significance, which tends to flatten exactly the nuance this review is trying to preserve. Table 7 sets out, outcome type by outcome type, which effect measures are appropriate and under what conditions pooling is defensible: continuous preclinical outcomes as Hedges’ g or a log response ratio, but only for the same candidate, cancer context, model class, comparator, and a compatible time point, and only with at least three independent studies; dichotomous outcomes as risk or odds ratios, taking care not to mix response, recurrence, and mortality endpoints under one pooled estimate; time-to-event outcomes as log hazard ratios; prediction-performance metrics analyzed within, never across, differing candidate universes or label definitions; and a substantial residual category — docking energies, enrichment p-values, uncalibrated ranks — that will be synthesized narratively and in tables rather than pooled, because pooling them would imply a level of comparability the underlying methods simply do not have. Where meta-analysis is feasible, random-effects models with REML estimation will be used, with Hartung-Knapp confidence intervals, τ², I², and a prediction interval reported alongside the pooled estimate; multiple effects from a single study

Figure 5. PRISMA 2020 study-selection flow diagram. This figure reproduces the standard PRISMA 2020 flow-diagram template that will record the number of records identified, deduplicated, screened, retrieved, assessed for eligibility, excluded (by reason, per Appendix B), and finally included once formal screening is complete. The counts are intentionally left blank in this protocol-stage document, consistent with the statement in Section 3.6 and Section 5.8 that formal dual screening has not yet occurred; populating this figure with real counts is a required step before this manuscript can be considered a completed systematic review.

Figure 6. Seven-level validation ladder for classifying translational maturity. This figure orders the possible evidentiary states for any candidate-cancer pair from Level 0 (internal computational ranking only, including docking and pathway enrichment on their own) through Level 6 (prospective clinical efficacy or regulatory evidence), with intermediate levels covering independent computational replication, biochemical target engagement, functional cellular or organoid perturbation, in vivo tumor models, and human observational or patient-derived validation. Introduced in Section 4.3 and applied throughout Tables 8 and 9, the ladder’s purpose is to prevent a study from being credited with a higher validation level than it actually demonstrated simply because a stronger method is mentioned elsewhere in its discussion section.

Table 6. NDDA appraisal domains with low-risk and high-risk signals. This table operationalizes the seven domains of the proposed Network-Drug Discovery Appraisal (NDDA) instrument — data provenance, network construction, analysis, validation, exposure and mechanism, reproducibility, and reporting/conflicts — by defining a concrete low-risk and high-risk signal for each, so that two independent reviewers scoring the same paper are more likely to converge on the same judgment. As noted in Section 3.8 and Section 5.8, this instrument still requires formal piloting and inter-rater calibration before it can be treated as validated.

will be handled through a prespecified outcome hierarchy or robust variance estimation; funnel plots and small-study tests will only be attempted once at least ten sufficiently comparable studies are available; and sensitivity analyses will exclude high-risk studies, unstandardized natural-product mixtures, non-independent datasets, and implausible exposure conditions.

3.10 Certainty of evidence

GRADE (Guyatt et al., 2008) is well suited to bodies of clinical intervention-outcome evidence, but mechanically applying it to computational target rankings or exploratory mechanistic claims would lend those claims a false air of clinical certainty. Instead, for preclinical and computational evidence, this review will present a transparent confidence profile spanning risk of bias, consistency, directness, precision, external validation, and publication bias — an approach that is admittedly more effortful to report than a single GRADE rating, but one that seems, on balance, more honest about what this evidence base can and cannot support.

4. Mapping the Evidence: A Thematic Synthesis of Network-Guided Oncology Candidates

A word of orientation before this section proceeds: what follows is an evidence-informed thematic synthesis built from verified, representative studies, not the final results of a completed systematic review — the formal screening described in Section 3 has not yet been carried out, and Table 8’s included-study column will remain unpopulated until it has. We think it is more useful, and considerably more honest, to show what this kind of evidence actually looks like using real, checked examples than to leave the section abstract. Readers should treat the patterns below as illustrative of what a completed synthesis will need to weigh, not as a prevalence estimate of anything.

4.1 Approved-drug repurposing: what the stronger examples share

Approved-drug studies in this space range across target-centric polypharmacology, subtype-specific modeling, immune-state networks, metabolic reconstruction, and cross-cancer screening, and their shared advantage — access to a compound whose human pharmacology is already partly known — comes bundled with a shared limitation: the distance between a new mechanistic hypothesis and a regimen a clinician could actually prescribe remains, in nearly every case, considerable (Turanli et al., 2021; Palve et al., 2021; Cavalcante et al., 2024; Fiscon et al., 2022). Table 8 tabulates eight illustrative examples, and a pattern emerges from reading them side by side that is easy to miss reading them one at a time: the studies that feel most convincing are not necessarily the ones with the most elaborate network, but the ones that treated the network prediction as a starting point for a validation chain rather than its endpoint. Kuenzi and colleagues (2017), for instance, used integrated functional proteomics to surface a noncanonical ceritinib target in ALK-negative lung cancer, then backed that finding with phenotypic screening, proteomic confirmation, inhibitor and RNAi work, and combination testing — a genuinely thorough chain, even though the authors themselves stop short of claiming clinical efficacy. Pacheco and colleagues (2019) took a different route through patient-derived metabolic models to prioritize naftifine, ketoconazole, and mimosine in colorectal cancer, with experimental evaluation following the computational step; the main open question here is less about the biology and more about how much the model’s assumptions need independent replication before the finding travels further. Lee and colleagues (2022) combined a transcription-factor network with the Connectivity Map to recover ATRA for leukemia differentiation therapy and propose new combinations, using genetic and pharmacologic perturbation in primary patient-derived cells — arguably one of the stronger translational chains in the table, if only because differentiation therapy already has clinical precedent to anchor against. Firoozbakht and colleagues (2022), by contrast, remain at the computational stage for breast-cancer subtypes, and — to their credit — say so explicitly, calling for clinical confirmation rather than implying it. Ahmed and colleagues (2023) prioritized drugs and combinations for HPV-associated cervical cancer through interactome proximity, with gene-set and literature support standing in, for now, where prospective functional validation is still needed. Ng and colleagues (2024) took a less conventional path, screening dendritic-cell antigen-presentation states to identify sitagliptin, and followed up with in vitro, mouse, and an observational human recurrence association — hypothesis-strengthening, certainly, but not causal proof, as the authors themselves note. Tian and colleagues (2025) pushed further into in vivo territory, combining maprotiline with anti-CTLA4 therapy in a PD-L1/SPOP mechanism, leaving dose, safety, and human efficacy as the remaining open

Table 7. Effect measures and pooling rules by outcome type. This table specifies, for five outcome categories, which effect measure this review will extract and under what conditions — same candidate, cancer context, model class, comparator, and time point, with a minimum study count — pooling is defensible. The final row, covering docking energies, enrichment p-values, and uncalibrated ranks, is explicitly reserved for narrative and tabular synthesis only, since pooling such measures would imply a comparability the underlying methods do not support.

Table 8. Illustrative approved-drug repurposing studies mapped to the validation ladder. This table presents eight verified, representative approved-drug repurposing studies discussed in Section 4.1, describing the cancer context, the network’s specific contribution, and how far each study’s validation and interpretation extend. It is illustrative of methodological and validation diversity rather than a complete or randomly sampled inventory of the eligible literature, consistent with the limitations stated in Section 5.8; the final systematic review will replace this table with the full set of included studies

questions. And Weich and colleagues (2025) grounded a clindamycin-caspase-1 hypothesis in single-cell transcriptomics and regulatory-core network analysis, supported by macrophage functional assays in patient- derived tumor-associated macrophages, with tumor-level and clinical validation still pending. Across all eight, the throughline — and perhaps the single most transferable lesson for anyone appraising this literature — is that computational-only findings remain genuinely useful for generating hypotheses, but describing them as evidence of anticancer efficacy would be, at minimum, premature.

4.2 Natural products and botanical formulations: where the chain tends to hold, and where it tends to break

Natural-product network pharmacology usually follows a fairly standard workflow — constituent identification, target prediction, disease-target intersection, protein-interaction network construction, enrichment analysis, and docking — and that workflow, useful as it is, has a known failure mode: it can repeatedly elevate the same well-annotated, high-degree proteins (AKT1, EGFR, assorted inflammatory regulators) almost regardless of which natural product is being studied, simply because those proteins are so densely represented in public databases (Fang et al., 2018; Noor et al., 2022; Panossian, 2025). Table S1 collects eight examples that, in our judgment, illustrate both ends of this spectrum. At the stronger end, Li and colleagues (2021) combined metabolomics with network pharmacology to link Tricin, from Weijing decoction, to PRKCA/SPHK/S1P signaling in non-small-cell lung cancer, with cell and animal work following the computational prediction — though formula-to-component attribution and human exposure remain, understandably, open questions for any multi-herb decoction. Seo and colleagues (2021) took triptolide through transcriptomic clustering, promoter-motif analysis, reporter assays, EMSA, and docking to examine NF-κB and GATA1 hypotheses across several cancer cell lines, and the value here lay specifically in using falsifiable assays rather than resting on network topology alone. Yu and colleagues (2022) traced nitidine chloride through cell-death inhibitor panels, microscopy, target-engagement assays, pathway rescue, and a xenograft model to support a PI3K/AKT-related pyroptosis mechanism in lung cancer — clinical exposure and selectivity still need working out, but the mechanistic chain itself is unusually complete. Lin and colleagues (2023) followed network analysis of a Wei-Tong-Xin extract with cell assays, pathway modulation, and a mouse xenograft in colorectal cancer, though extract standardization and independent replication remain, as with most multi-component formulations, the sticking points. Hu and colleagues (2024) moved from cell phenotyping through pathway perturbation to a zebrafish xenograft for Triphala in oral squamous-cell carcinoma, again leaving multi-component composition and human pharmacology unresolved. Yang and colleagues (2024) converged network, docking, cell, and mouse xenograft evidence on EGFR/PI3K/AKT/mTOR-related targets for cepharanthine in nasopharyngeal carcinoma, with the caveat — worth repeating, since it recurs across this entire literature — that docking affinity is not the same thing as demonstrated target engagement. Fan and colleagues (2025) arguably went furthest on the exposure question, measuring plasma-absorbed compounds from Xianlian Jiedu decoction by UPLC-Q-TOF-MS before network analysis, docking, and cell, mouse, and protein-level validation in colorectal cancer. And Zhang and colleagues (2025) combined target prediction, docking, ADME/molecular-dynamics modeling, cell assays, and an AML xenograft to validate compounds from Yinchen Wuling San. Taken together (Figure 4), what separates the more persuasive studies from the less persuasive ones is not method sophistication so much as a willingness to check whether the predicted constituent was actually present, at a plausible concentration, in whatever was administered — a step that sounds almost too basic to matter, and yet is precisely where a great deal of this literature quietly stalls (Fang et al., 2018; Noor et al., 2022; Panossian, 2025).

4.3 A validation ladder for judging translational maturity

Because “validated” gets used rather loosely across this literature — sometimes meaning a docking score, sometimes meaning a phase II trial — this review proposes a seven-level validation ladder (Figure 6) that assigns each candidate-cancer pair the highest level actually demonstrated, never the highest method merely mentioned in a paper’s discussion. Docking, on this ladder, is a computational plausibility check and stays at Level 0 until something external or experimental follows it; cell viability alone climbs only slightly higher, since viability changes do not by themselves establish mechanism; and the ladder’s upper rungs — orthogonal target engagement, genetic or pharmacologic rescue, multiple biologically appropriate models, blinded animal work, patient-derived systems, and prospective clinical testing — are reached by a distinctly smaller subset of the studies reviewed here (Table 8, Table S1). Applied consistently, the ladder does something we think matters more than it might first appear: it stops a well-produced network diagram from being mistaken for strong evidence, simply by asking, level by level, what was actually shown.

4.4 Why studies disagree

Readers new to this literature sometimes assume that disagreement between studies reflects genuine biological complexity alone, and while that is certainly part of it, Table S2 suggests at least six further, more mundane sources of disagreement worth separating out. Different input databases — with different disease-gene sets, evidence codes, species coverage, and version histories — can alter network topology before any biology enters the picture at all. Well-studied proteins accumulate more recorded edges simply because they have been studied more, producing hub inflation that looks like biological importance but may be substantially an artifact of literature and database bias. Cancer heterogeneity means that bulk tissue, cell lines, molecular subtypes, resistant states, and microenvironmental cell populations are, in an important sense, different biological systems, and pan-cancer generalization across them is often less justified than it appears. Different algorithmic objectives — proximity, signature reversal, network controllability, metabolic essentiality, link prediction — optimize genuinely different mathematical constructs, so disagreement between methods is not necessarily a sign that one is right and the other wrong. The candidate universe itself varies enormously between approved-only libraries, investigational-compound libraries, natural-product databases, and database-specific compound sets, making raw rank comparisons across studies close to meaningless without normalization. And, finally, experimental design choices — dose, exposure duration, assay type, controls, model identity — shift observed efficacy in ways that a network diagram alone cannot reveal. This review’s response to each source of disagreement, summarized in the right-hand column of Table S2, is essentially the same in spirit each time: extract enough detail to make the disagreement legible, rather than treating it as noise to be averaged away.

4.5 What can be said with confidence, and what cannot

Table S3 sorts the review’s working conclusions into four tiers, and it is worth being explicit about why the tiers matter. It seems reasonably well established, at this point, that networks can integrate heterogeneous biological and pharmacological data to prioritize candidates and mechanisms with genuine hypothesis-generating value (Guney et al., 2016; Cheng et al., 2018; Cheng et al., 2019; Zeng et al., 2019; Zitnik et al., 2018; Kuenzi et al., 2017 through Weich et al., 2025). It also seems well established, on methodological grounds this time rather than biological ones, that external validation, target engagement, plausible exposure, and reproducible reporting are prerequisites for any translational claim, not optional extras (Cavalcante et al., 2024; Panossian, 2025; Schaduangrat et al., 2020). More tentatively, single-cell, spatial, patient-specific, causal, and multimodal knowledge-graph approaches look like promising directions, though the prospective benchmarking needed to confirm that promise is still thin on the ground. And several beliefs that circulate informally in this space remain, on the evidence available, genuinely uncertain rather than established — that a high centrality score or favorable docking energy predicts therapeutic efficacy; that approved-drug status guarantees acceptable oncology safety or dosing; and, perhaps most persistently, that a natural origin implies safety, affordability, or clinical effectiveness on its own. None of these three beliefs is unreasonable to hold provisionally, but none of them, as things currently stand, has earned the confidence with which it is sometimes asserted.

5. From Prediction to Proof: Clinical, Equity, and Research Implications

5.1 What this means for patients

It bears saying plainly, even though it should be obvious: a computational prioritization result is not permission to self-medicate, whether with an approved drug used off-label or with a supplement or herbal formulation. A repurposing candidate identified through network analysis may need a substantially different dose, a specific combination partner, or a biomarker-defined population before it does anything useful, and natural products in particular can vary in composition or interact unpredictably with ongoing anticancer therapy. Anyone communicating these findings to patients — clinicians, researchers, or the media covering their work — has a responsibility to keep a computational hypothesis, preclinical activity, and demonstrated clinical benefit visibly separate from one another, rather than letting the excitement of a promising target blur those distinctions.

5.2 What this means for clinicians

The most clinically useful output from this literature is rarely a long, unranked list of candidates; it is a small number of testable drug-biomarker-context hypotheses, accompanied by honest exposure and safety information. Clinicians are, in our view, underused as collaborators at this stage — they are well placed to help define clinically meaningful endpoints, characterize resistance states and comorbidities, specify an acceptable therapeutic window, and flag where a retrospective real-world analysis might plausibly precede, and de-risk, a prospective trial.

5.3 What this means for researchers building the next generation of studies

Several practices recur, almost as a checklist, across the stronger studies reviewed here, and are worth stating as explicit recommendations rather than leaving implicit: preregister candidate-selection logic before looking at results; version every data source used; benchmark against simple baselines rather than only against other complex methods; use degree-preserving and temporal null models to guard against the hub-inflation problem described above; actively prevent train/test leakage; report negative and sensitivity analyses rather than only the headline finding; and release executable code alongside the paper. On the experimental side, target engagement, genetic or pharmacologic rescue, pharmacologically plausible exposure, multiple appropriate models, and toxicity controls should be treated as standard follow-up, not optional extensions reserved for a later, better-funded study.

5.4 What this means for policymakers and funders

Public investment in cancer registries, biobanks, pharmacovigilance systems, interoperable omics infrastructure, and open computational tooling would, we think, do more for this field’s translational output than funding additional prediction algorithms in isolation. Funding calls that reward prediction volume alone risk incentivizing exactly the computational-only pattern this review repeatedly flags as insufficient; calls that require a prospective validation plan, and an explicit access and affordability plan, would push the field toward the kind of evidence it actually needs.

5.5 Relevance to low- and middle-income settings, Asia, and Bangladesh specifically

LMIC cancer-research priorities tend to center on earlier diagnosis, affordable treatment, value-based care, and technology that can actually survive transfer into a constrained health system (Pramesh et al., 2022), and network-driven repurposing has an obvious appeal here: approved drugs and locally studied natural products can, in principle, generate lower-cost hypotheses than de novo drug discovery. That promise, though, is conditional rather than automatic — laboratory validation, quality-controlled manufacturing, pharmacokinetic characterization, toxicity monitoring, regulatory review, and trials all still require resources that a lower-cost starting hypothesis does not, by itself, provide. Bangladesh sits in an interesting position on this front, with real opportunities in bioinformatics capacity, biodiversity-informed natural-product research, generic-drug manufacturing, and regional scientific collaboration, but also with fragmented cancer surveillance and limited nationally representative outcome data that constrain both patient-specific modeling and real-world validation (World Health Organization, 2020; Iqbal, 2025). Table S4 sets out five specific opportunity-risk-safeguard triads relevant to this setting — generic approved-drug libraries, regional medicinal plants, public multi-omics data, cloud and open-source computation, and rapid candidate prioritization — and a Bangladesh-focused research agenda building from this review would, we think, do well to prioritize registry strengthening, standardized biospecimen and clinical metadata, common data models, and partnerships structured to retain local governance and analytical capacity rather than simply exporting samples and data elsewhere.

5.6 Controversies and open gaps the field has not resolved

Several tensions recur across this literature without a settled answer, and it seems more honest to name them as open rather than paper over them. There is a persistent question of whether networks are being used predictively or merely descriptively — many studies, on close reading, infer a plausible mechanism after a candidate has already been selected on other grounds, rather than genuinely predicting new candidates in advance. Hub-gene inflation (AKT1, TP53, EGFR, MYC, and a handful of inflammatory pathways appear again and again) may reflect real biological centrality, annotation density, or, most likely, some mixture of both, and current methods rarely disentangle which. Validation inflation is common — docking, enrichment, and reuse of the original discovery dataset are frequently labeled “validation” despite lacking independence or causal structure. Natural-product identity remains a recurring soft spot, since a computationally predicted compound may never have been measured in the actual administered extract, and formulation-level activity is not automatically attributable to any single molecule within it. Clinical transportability is another open question, since tumor type, prior therapy, ancestry, comorbidity, microbiome composition, and concomitant medications can all shift response and toxicity in ways a network model rarely captures. Reproducibility itself is fragile, since database drift and inaccessible code can prevent exact reconstruction even when a written workflow looks, on paper, complete. Negative results are almost certainly underreported, inflating the apparent validation success rate of this literature as a whole. And equity remains under-modeled: available datasets and candidate libraries disproportionately reflect research-rich regions, while implementation constraints in resource-limited settings are rarely built into the models themselves.

5.7 Where the field should go next

Table S5 sets out eight priorities we think deserve particular attention going forward, and it may help to group them conceptually rather than list them mechanically. Several concern rigor at the prediction stage — prospective, time-stamped benchmarking where predictions are evaluated against later, independently generated data rather than selected post hoc; context-specific networks built around subtype, treatment state, tissue, and single-cell or spatial information; and causal validation through target engagement plus genetic or pharmacologic perturbation, not correlation or docking alone. Others concern translation readiness specifically — pharmacology-aware ranking that accounts for free-drug exposure, tissue penetration, metabolites, and drug interactions; and natural-product standards covering taxonomy, voucher specimens, quantitative chemistry, and batch reproducibility. And a final cluster concerns the scientific ecosystem around the work itself — open and executable science with persistent data, code, and complete parameters; equity-aware translation that explicitly models local epidemiology, cost, and implementation; and a living-review structure with annual search updates and versioned evidence tables, since a field moving this quickly will otherwise be out of date before the ink on any single review has dried.

5.8 Limitations of this synthesis

A few limitations deserve stating directly rather than left implicit. This document remains, at the time of writing, a protocol and evidence-informed draft rather than a completed systematic review; the pilot PubMed retrieval reported in Appendix A was neither deduplicated nor screened, and no inclusion proportions have been calculated from it. The thematic examples in Section 4 were selected to illustrate methodological and validation diversity, not to represent a comprehensive or randomly sampled set of the eligible literature. Subscription databases recommended in Section 3 were not queried during this drafting stage, and no quantitative meta-analysis has been performed, since no homogeneous, independently extracted outcome set yet exists. The NDDA instrument proposed in Section 3.8 is purpose-built and requires piloting, inter-rater calibration, and transparent reporting before it can be treated as more than a working scaffold; it does not, and should not, replace PROBAST or design-specific risk-of-bias tools. This review should not, in short, be submitted or cited as a completed systematic review until protocol registration, full-database searching, dual screening and extraction, calibrated risk-of-bias judgments, and a populated Results section and PRISMA diagram are all in place.

6. Conclusion

Network medicine and systems pharmacology offer a genuinely coherent way of prioritizing multi-target natural products, repurposed approved drugs, and rational combinations in cancer — and the verified studies surveyed here show that network predictions can, when followed through, lead to mechanistically informative cell, animal, patient-derived, and observational validation. What the evidence does not support, however, is treating network centrality, docking, enrichment, or an in silico rank as proof of clinical efficacy; that distinction, more than any single method, is what separates a credible hypothesis from an overstated one. A high-quality review of this literature, this one included, should therefore judge each candidate not only by what was predicted, but by how the underlying network was built, whether the prediction outperformed a credible baseline, whether validation was independent and causal rather than circular, whether the tested exposure was pharmacologically realistic, and whether the eventual translation pathway is genuinely feasible and equitable across the settings that most need better cancer therapeutics.

References