A Clinical Evidence-Guided Bioinformatics Framework for Expanding Cisplatin-Based Combination Therapies in Cervical Cancer

Technical Summary: A Clinical Evidence-Guided Bioinformatics Framework for Expanding Cisplatin-Based Combination Therapies in Cervical Cancer

Problem Statement
Cisplatin-based combination therapy remains the cornerstone of systemic treatment for advanced cervical cancer. However, identifying effective therapeutic partners to expand these regimens is hindered by significant tumor heterogeneity and the exponential number of potential drug combinations. Existing computational platforms primarily focus on retrospective analysis of experimentally generated drug pairs or the prediction of arbitrary synergies. These approaches often lack a direct link to clinically established treatment backbones and provide limited mechanistic context for why specific combinations should be prioritized. Consequently, there is a need for a framework that begins with clinically investigated regimens, systematically evaluates whether additional agents improve their pharmacological performance within specific biological models, and integrates multi-omics evidence to provide explainable therapeutic hypotheses.

Methodology
The authors developed a clinical evidence-guided computational decision-support platform implemented as an interactive R Shiny application. The workflow proceeds through the following sequential layers:

1. Clinical Evidence Definition: Clinically investigated cisplatin-containing regimens were curated from ClinicalTrials.gov. These regimens served as reference backbones. Only combinations with corresponding pharmacological profiles in the Genomics of Drug Sensitivity in Cancer (GDSC2) database were retained for computational analysis.
2. Pharmacological Reconstruction and Simulation: Model-specific dose-response profiles were reconstructed from GDSC2 fitted parameters (EC50, Hill slope, asymptotes). Reference regimens were simulated across eligible cervical cancer cell lines. Candidate third-drug additions were evaluated against these references using four interaction models: Bliss Independence, Highest Single Agent (HSA), Loewe Additivity, and Zero Interaction Potency (ZIP).
3. Candidate Prioritization Strategy: Rather than applying a universal synergy threshold, candidates were prioritized based on whether they improved the pharmacological performance (e.g., Bliss score) of the specific clinical reference regimen within the same biological model. This allowed for the distinction between intrinsic monotherapy potency and combination-specific benefit.
4. Mechanistic Evidence Integration: Pharmacological prioritization was contextualized by integrating three complementary evidence layers:
   • Functional Genomics: CRISPR gene dependency scores from the Cancer Dependency Map (DepMap) to assess target essentiality.
   • Transcriptomics: Gene expression levels (TPM) to evaluate target availability.
   • Somatic Mutations: Mutation status of target genes to provide genomic context.
5. Graph-Learning Framework: A graph convolutional network (GCN) was constructed where candidate drugs were nodes and edges represented weighted similarities derived from pharmacological responses, molecular targets, mechanisms of action, and genomic evidence. This allowed for systems-level prioritization, propagating information across biologically related compounds.
6. Patient-Level Contextualization: Prioritized molecular targets were projected into an independent patient transcriptomic cohort (GSE299125, comprising HIV-positive and HIV-negative Kenyan women) to identify patient subgroups where the therapeutic vulnerabilities were transcriptionally represented.
7. Validation: The graph-learning framework was internally validated using cross-validation, ROC analysis, confusion matrices, Decision Curve Analysis (DCA), and graph-clustering metrics, benchmarked against pharmacology-only logistic regression and prevalence baselines.

Key Results

• Model-Specific Heterogeneity: Clinically established cisplatin regimens (e.g., cisplatin-paclitaxel) exhibited substantial heterogeneity in pharmacological interactions across different cervical cancer models. Some models showed synergistic interactions, while others showed antagonism, reinforcing the necessity of model-specific benchmarking over global thresholds.
• Third-Drug Prioritization: In an unrestricted screen of 286 candidates against the cisplatin-paclitaxel backbone, the top-ranked third drugs varied significantly by model. For instance, Glutathione was the top candidate for C-33 A, while Navitoclax was top for HeLa. The framework successfully identified candidates that improved upon the reference regimen, distinguishing them from those that were merely potent monotherapies.
• Mechanistic Discordance: Integration of CRISPR dependency revealed that pharmacological improvement did not always correlate with strong dependency on the drug's canonical target. For anti-angiogenic candidates, dependencies were often found in cell-cycle regulators (PLK4, CDK2) rather than canonical vascular targets (KDR, FLT1). Furthermore, high target expression did not consistently predict functional dependency, and somatic mutations were rare (3.8% of associations) and not predictive of pharmacological benefit.
• Graph-Learning Performance: The graph convolutional network achieved a pooled ROC-AUC of 0.907 in cross-validation, outperforming the pharmacology-only logistic regression model (AUC 0.840) and the prevalence baseline (AUC 0.671). Decision Curve Analysis demonstrated that the graph model provided higher net benefit in specific models (e.g., Ca Ski, SiHa, SISO) across various probability thresholds.
• Patient Projection: Projection into the GSE299125 cohort revealed marked inter-patient heterogeneity. Specific patient subsets were identified with elevated expression of targets linked to prioritized hypotheses (e.g., CDK2 for bosulif, TOP1 for topotecan, MTOR for mTOR inhibitors), while others showed distinct multi-target profiles. This confirmed that the platform-derived vulnerabilities were transcriptionally represented in patient cohorts, though not uniformly across all patients.

Significance and Claims
The paper claims to present a clinically grounded, explainable computational framework that shifts the paradigm of drug combination analysis from retrospective synergy assessment toward evidence-guided therapeutic optimization. By anchoring the analysis in clinically investigated regimens rather than arbitrary drug pairs, the platform provides a reproducible foundation for generating biologically interpretable therapeutic hypotheses.

The authors emphasize that their approach integrates clinical trial evidence, reconstructed pharmacology, functional genomics, graph learning, and patient-level molecular contextualization within a unified workflow. This allows for the rational prioritization of personalized cisplatin-based combination therapies. The study demonstrates that therapeutic response is fundamentally tumor-state dependent, and that combining pharmacological modeling with systems-level biological interpretation (via graph learning) offers a more robust basis for hypothesis generation than pharmacological scores alone.

The platform is explicitly positioned as a research tool for computational therapeutic prioritization and hypothesis generation, not for clinical diagnosis or treatment recommendation. The authors acknowledge limitations, including the reliance on reconstructed rather than prospective experimental combination data and the exclusion of proteomic and microenvironmental factors, but argue that the modular architecture allows for future integration of these data types. The work illustrates a transition toward "intelligent therapeutic reasoning," where clinical evidence and systems biology are integrated to support precision oncology research in cervical cancer.