Climbing the causal ladder: a new framework sorts real cancer drug targets from statistical noise

Genomic studies of cervical cancer routinely produce long lists of genes whose activity is altered in tumours, but a dysregulated gene is not automatically a drug target. A team of researchers from Manipal University Jaipur, Cincinnati Children’s Hospital Medical Center, Marshall University and the University of Lucknow has now proposed a solution to this chronic ambiguity: the Causal Integration Ladder, or CIL, an evidence-accounting framework published in Frontiers in Systems Biology that explicitly separates statistical association from causal, therapeutically tractable mechanism. The framework arrives at a moment when multi-omic catalogues of cancer are expanding faster than the mechanistic evidence needed to act on them.

The core insight behind CIL is that three different claims are routinely conflated in cancer genomics: that a gene is associated with disease, that it is causally relevant to the disease process, and that it is a suitable target for therapeutic intervention. Differential expression may reflect reverse causation or shifts in the cellular composition of a tumour sample. A genome-wide association study signal may merely tag a nearby causal locus. An expression quantitative trait locus may operate in a proxy tissue rather than the tumour. And a gene dependency observed in a cell line may reflect general essentiality rather than a cancer-selective vulnerability. CIL forces researchers to state which type of claim each data modality actually supports.

The ladder itself has five rungs. Level I records disease association from expression, clinical, genomic or other omics data — a nomination tier that confers no causal status. Level II evaluates genetically constrained inference, split into inherited Level II-G evidence from germline regulatory variation and acquired Level II-S evidence from somatic mutation and copy-number alteration. Level III is divided into Level III-A, scalable computational functional evidence such as CRISPR dependency screens and pharmacogenomic correlations, and Level III-B, direct experimental perturbation with appropriate controls and rescue. Level IV reserves translational validation for higher-fidelity efficacy models, therapeutic-window assessment and clinical testing. Crucially, viral etiology — in cervical cancer, persistent high-risk HPV infection — is treated as an explicit context modifier rather than folded into the genetic evidence.

To demonstrate the framework, the team applied it to The Cancer Genome Atlas cervical cancer cohort, TCGA-CESC, comprising 306 tumours and only 3 normal samples. Because the control group was so small, the resulting 25-gene panel — 15 upregulated and 10 downregulated genes ranked by effect size — is explicitly exploratory rather than confirmatory. The upregulated genes told a coherent story of epithelial transformation: keratins KRT5, KRT6A and KRT15; cell-cycle drivers SFN, MYBL2, CDKN2A and CDC20; and the inflammatory alarmin pair S100A8 and S100A9. The downregulated genes were dominated by smooth-muscle and stromal cytoskeletal programs, including DES, CNN1, ACTG2, MYH11 and ACTA2, reflecting the loss of normal tissue architecture as tumour invades.

HPV status proved to be a powerful sensitivity axis. Sample-level annotations were available for 291 primary tumours, of which 280 were HPV-positive, 9 HPV-negative and 2 indeterminate. When the analysis was restricted to HPV-positive cases, all 25 Level I candidates retained their original direction of effect — a robustness result that matters in a cancer where the virus is the necessary etiological event. Exploratory comparisons between HPV-positive and HPV-negative tumours identified six genes with panel-level FDR-significant differences, including higher expression of CDKN2A, S100A9 and SERPINB5 in HPV-positive cases, though the authors caution that the tiny, histologically heterogeneous negative group prevents any claim of HPV-specific effects.

The somatic analysis delivered the study’s most concrete molecular findings. Among 295 copy-number-evaluable tumours, ten candidates showed false-discovery-rate-significant associations between copy number and expression, with CDKN2A standing out: one deep deletion, 58 shallow deletions, 52 gains and a single high-level amplification, coupled with a copy-number–expression correlation of ρ = 0.372 (FDR = 6.5 × 10⁻¹⁰). Because CDKN2A encodes the RB-pathway regulator p16, its somatic dosage must be interpreted jointly with the viral E7 oncoprotein’s inactivation of the same pathway. By contrast, recurrent protein-altering mutation was uncommon among the eight prioritized genes, and KRT5 and KRT6A showed no copy-number–expression association at all, marking them as epithelial and HPV-context candidates rather than CNA-driven ones.

The genetic audit was deliberately harsh. Each of the eight upregulated genes carried forward for Level II had documented cis-eQTL records from GTEx v8 and contextual GWAS annotations from the NHGRI-EBI catalog, yet no candidate possessed a complete, source-auditable cervical-cancer colocalization record at PP.H4 ≥ 0.80 plus a valid harmonized causal estimate. Mendelian randomization was deemed ineligible for this implementation because the retained records lacked aligned variant-level effect estimates. All eight genes therefore remain Partial Level II-G — a result the authors present not as failure but as an honest map of what is missing. Absence of germline support, they stress, is not evidence against somatic, viral or functional relevance.

Computational functional evidence fared similarly. CDC20 posted the strongest cervical-line CRISPR dependency, a Chronos score of −2.74, but pan-essentiality and the absence of model-level distributions kept it at Moderate Level III-A, prioritized for therapeutic-window testing rather than labelled a selective vulnerability. KRT6A’s weaker −0.55 score also remained Moderate. The remaining six candidates were classified as Context-only Level III-A: nominal drug-response correlations from GDSC and single-cell compartment annotations from CZ CELLxGENE guide experiment selection but do not constitute functional support. Notably, the myeloid localization of S100A8 and S100A9 reframes them as microenvironmental rather than tumour-cell-autonomous hypotheses, pointing toward tumour–myeloid co-culture and cytokine-blockade experiments instead of tumour-cell viability assays.

The framework’s most provocative demonstration is its reverse-entry workflow. Three clinically established cervical-cancer therapies — bevacizumab targeting VEGFA, validated in the randomized phase III GOG-240 trial; pembrolizumab targeting PD-1, validated in KEYNOTE-826; and tisotumab vedotin directed at tissue factor, validated in innovaTV 301 — were each retained at Level IV for their specific target-treatment relationships, with lower levels reported as not assessed rather than imputed. This preserves a boundary that clinical success often obscures: the efficacy of a checkpoint inhibitor does not establish that PD-1 is a tumour-cell dependency, and an antiangiogenic survival benefit does not prove that VEGFA initiated carcinogenesis. Therapeutic validation and disease causality are different claims, and CIL refuses to let one stand in for the other.

The practical payoff is a prioritized experimental agenda rather than a list of validated targets. For KRT6A, the most informative next study is inducible knockdown across several HPV-positive cervical models with cDNA rescue, epithelial-state markers and non-malignant cervical keratinocyte comparators. For CDC20, partial time-controlled suppression with rescue should probe the therapeutic window rather than whether complete loss impairs proliferation. For CDKN2A, perturbation should be stratified by RB-pathway state and evaluated alongside CDK4/6-inhibitor response. The authors also outline how a machine-readable CIL implementation — structured evidence records, explicit provenance and fixed decision rules exposed through an MCP-compatible service — could let AI clients query evidence and propose the next uncertainty-reducing analysis, while reserving tier promotion for expert review. The design limitations are stated plainly: three normal samples, proxy-tissue eQTLs, precomputed colocalization and no direct perturbation. But the ladder’s logic — that missing evidence is more informative than an assigned label — may prove the framework’s most transferable idea, from HPV-driven cervical cancer to EBV-related lymphomas and beyond.

Subject of Research: A multilevel evidence framework for evaluating therapeutic targets in cervical cancer

Article Title: The causal integration ladder: a multilevel evidence framework for therapeutic target evaluation in cervical cancer

Article References: Singh, S., Srivastava, M., Uppada, P., Ameta, D., Singh, A., Banerjee, M., & Kushwah, A. S. (2026). The causal integration ladder: a multilevel evidence framework for therapeutic target evaluation in cervical cancer. Frontiers in Systems Biology, 6, Article 1874355. https://doi.org/10.3389/fsysb.2026.1874355

Image Credits: AI Generated

DOI: 10.3389/fsysb.2026.1874355

Keywords: cervical cancer, causal inference, HPV, therapeutic targets, genomics, TCGA, systems biology, drug discovery, causal, integration, ladder, multilevel