CancerPPIr is an exploratory network-prioritization workflow. It supports hypothesis generation and structured evidence review; it is not a diagnostic device, treatment recommendation system, or substitute for qualified clinical interpretation.
CancerPPIr now enforces raw differential-expression pvalue and base-2
tumor specimen / reference condition logFC semantics. Positional column
fallback, incomplete numeric rows and duplicate input symbols are rejected.
The selected headers and contract policies are written to output provenance.
This validation does not recover the biological composition of the reference condition, the upstream normalization, design formula, covariates, filtering or multiple-testing method. Those study-level details must remain with the source differential-expression analysis. Comparisons across runs require the same upstream model, contrast and selection procedure.
Raw differential-expression p-values are not adjusted within CancerPPIr. Appropriate multiple-testing control and input-gene selection must therefore be performed and documented upstream; enrichment FDR does not correct the input differential-expression p-values.
HGNC normalization, alias correction and STRING mapping can still change row counts. Input rows, mapped input rows, unique mapped proteins and final graph nodes are distinct quantities.
Bulk tumor RNA-seq combines malignant cells with immune, stromal, endothelial, and other specimen components. CancerPPIr does not deconvolve cell types, infer cell fractions, or prove tumor-cell-intrinsic origin. Module compartment and lineage fields are evidence-based contexts, not abundance estimates.
Histology, tumor purity, sampling site, and—when available—single-cell, spatial, immunohistochemical, or flow-cytometry data remain necessary for cell-origin claims.
STRING edges are curated, experimental, predicted, co-expression, text-mined, or otherwise integrated associations. They are not patient-specific physical interaction measurements. Network structure is influenced by STRING coverage, literature bias, organism annotation depth, the input gene set, mapping, and the selected score threshold.
Highly studied proteins may appear more connected than poorly characterized proteins. A node absent from the final network may still be biologically important.
The pinned human STRING v12 files are acquired from the official distribution
at https://stringdb-downloads.org/download/ when missing or invalid and are
then reused locally. The network uses the full STRING association graph with
combined_only evidence and a default score threshold of 400/1000. Individual
evidence-channel contributions are not exported, and an included edge is not
required to have direct experimental support. Louvain communities are
calculated without edge weights; interaction scores determine edge inclusion
but do not weight module detection. A deterministic seed improves
reproducibility, not biological validity.
The candidate score combines five normalized base components across three equally weighted evidence domains. Degree, betweenness, and log-transformed stress centrality form the topology domain; absolute logFC and transformed statistical evidence form the other two domains. It is a relative ranking inside one reconstructed network.
It is not:
- a probability of response;
- an essentiality score;
- a druggability or ligandability score;
- a toxicity or therapeutic-index estimate;
- evidence of causal oncogenic dependency;
- evidence from clinical trials.
Cross-case numerical comparison requires additional normalization and a separate study design; within-case ranks are the primary intended use.
The predefined candidate score is not externally calibrated or benchmarked against independent reference targets, functional-dependency experiments, treatment response, or patient outcomes. Its domain weights are heuristic, and no sensitivity, specificity, predictive value, uncertainty interval, or clinical accuracy is established. Absolute logFC prioritizes both increased and decreased expression; the signed logFC must be reviewed separately.
Canonical module interpretations are computational inferences from statistically significant, non-generic local STRING terms. The primary interpretation is the qualifying term with the lowest FDR; technical or covariate signatures override biological priority. Curated marker-rule evaluations remain auxiliary audit information and do not determine canonical interpretation or automatic priority.
Supported biological modules currently receive moderate confidence. Marker,
lineage, state, process, and conflict fields retained for schema compatibility
must not be mistaken for independently resolved or validated classifications.
Technical/covariate and unresolved modules remain visible but are not
automatically promoted. An unresolved module is not a pipeline error and does
not imply absence of biological function.
Functional enrichment is evaluated only for the five largest Louvain modules containing at least five proteins. Smaller or lower-ranked modules may remain unresolved because they were not tested, not because their biological function is absent. A qualifying term can be supported by as few as two proteins; supporting genes, module size, term specificity, and pathology must therefore be reviewed together.
Offline enrichment depends on the content and release of locally cached STRING v12 resources. Database categories may be redundant, unevenly annotated, or biased toward well-studied processes. Generic terms are excluded from the primary analytical evidence but retained in raw technical tables.
Statistical enrichment does not establish causal relevance, therapeutic vulnerability, or sample-specific pathway activity.
Module enrichment uses a hypergeometric test against the case-network background restricted to proteins represented in the STRING term map. Terms require at least two supporting proteins and between three and 500 annotated background proteins for reporting. Benjamini-Hochberg adjustment is calculated within each enrichment query across all terms meeting the three-to- 500 background-size bounds, before the two-supporting-protein reporting threshold is applied. Adjustment is not performed across modules or analyses. The resulting FDR is not the probability that a module assignment is correct. Whole-network enrichment instead uses the annotated human STRING background.
STRING network associations and STRING functional annotations are not independent sources of evidence. Local enrichment is computed by CancerPPIr from cached STRING annotations and its selected background; it is not an independent external validation or a live STRING web-service result.
Entity classification uses gene-symbol rules to prevent automatic promotion of special entities such as immunoglobulin/T-cell receptor loci, predicted loci, pseudogene-like, non-coding, mitochondrial, ribosomal, or Y-associated genes. These entities remain in network evidence where appropriate. Classification is conservative and may require manual review.
SHA-256 verifies exact bytes, not biological or semantic equivalence. XLSX files may receive different byte-level hashes across independent runs because of ZIP metadata or workbook timestamps even when visible tables are equivalent. Semantic comparison must inspect workbook tables and network attributes.
The standard manifest records STRING cache basenames and sizes but does not re-read multi-gigabyte cache files solely to hash them. Git metadata may be unavailable when CancerPPIr is run from a source archive rather than a Git working tree.
CancerPPIr is appropriate for:
- prioritizing proteins for follow-up analysis;
- summarizing network-associated biological programs;
- generating auditable candidate rationales;
- comparing hypotheses under a predefined study design;
- selecting candidates for experimental, pharmacological, or literature review.
Any translational claim requires independent pathology, molecular, protein- level, model-system, pharmacological, safety, and clinical evidence.