Working thesis chapter · 19 September 2026 · computational research
CONFLUENCE: An Evidence-Gated Dynamical Framework for Integrating Oncology Knowledge, Molecular Observations and Adaptive Cancer-State Modeling
Abstract
Cancer research now produces knowledge graphs, multi-omic assays and dynamical simulators in parallel. The failure mode is collapsing those layers. This study asks whether a provenance-controlled pipeline can keep the layers apart while still allowing testable predictions. Scientific success is a staged chain — traceability, mathematical validity, identifiability, out-of-sample prediction, experimental falsification — not disease eradication. PR #9 is architectural evidence for the gates, not therapeutic efficacy.
Why the study is necessary
GLOBOCAN 2024 estimates, published 2026: about 20.6 million diagnoses and 9.8 million deaths [1,2] (EVID-001). Breast cancer about 2.43 million new cases [1,2] (EVID-002). WHO: many cancers can be cured if found early and treated well; access is uneven [3] (EVID-003). Nigeria GLOBOCAN 2022 profile: 127,763 new cases, 79,542 deaths, breast 32,278 — 2022 estimates, not 2026 incidence [5] (EVID-004).
Tumours evolve and adapt under therapy [7] (EVID-006). Multi-omics increase resolution without automatically producing a causal model. Mathematical oncology supplies in-silico laboratories [8] (EVID-007). Adaptive therapy treats treatment as a process under selection [9] (EVID-008). The gap is connecting knowledge to dynamical hypotheses without dropping provenance.
Evidence for the problem
TNBC is a test case because NCI describes it as roughly 15% of breast cancers, typically faster-growing and more recurrent, and heterogeneous [6] (EVID-005). That is not a CONFLUENCE parameter. This branch adds a read-only OnCo adapter with nine tests that refuse knowledge-as-parameter and do not edit CancerODE [11] (EVID-009). OnCo is a cited knowledge graph, not Theta [10].
Knowledge ≠ Evidence ≠ Causal mechanism ≠ Parameter ≠ Prediction
legacy gene→parameter map ≠ identified parameter · structured representation ≠ biological observation
Problem statement
How can heterogeneous oncology knowledge and molecular observations be incorporated into a cancer dynamical model without collapsing evidence, mechanism, parameterisation and prediction into unsupported assumptions?
Knowledge → Evidence → Hypothesis → Mechanism → Parameter → Prediction → Experiment → New evidence
Every arrow is a failure point. A page may motivate a hypothesis; it must not become Theta.
Significance of the study
Scientific
Treat cancer as dynamic and adaptive rather than as independent targets [7] (EVID-006).
Methodological
Evidence gates. OnCo confidence is not P(H). Idea maturity is not evidence level. Failed hypotheses stay visible [10,11].
Computational
Auditable objects. PR #9 is the current artifact [11] (EVID-009).
Translational potential — not demonstrated benefit
Models can compare hypotheses before a wet experiment. This page does not claim patient benefit. Translation waits for Gates 6–8.
Future research
Attach genomic, single-cell, spatial, PK/PD and controllers around a frozen biological model.
CONFLUENCE evidence architecture
Oncology knowledge (OnCo, literature)
↓ read-only adapter
Evidence objects → genomic / molecular observation → cellular state
↓
Population dynamics — frozen confluence_v2_15d → controller → prediction → experiment → evidence
Lineages frozen: tnbc_mod_3s (ROS audit pending), confluence_report_6s (paper), confluence_v2_15d (live ODE), confluence_v1_calibrator (legacy map, assumed / unidentified).
Success protocol
Not a universal cure. Sequential and falsifiable:
Three-arm programme — hypotheses, not results
validation/validation_protocol.md states H0 versus H1. Thresholds such as r > 0.5 are protocol criteria, not findings [12] (EVID-010).
Separate validation
Cancer and metabolic systems independently.
Conjoined validation
Coupled versus independent models on comorbidity data.
Universality test
Whether ΔΦ distributions match after labels are stripped — against nulls.
Evidence ledger
Grok Evidence Auditor
Grok output is an evidence-audit aid. Not a truth score, not a parameter estimator, not expert review. Citations still load from the ledger when the auditor is offline.
Limitations and falsifiers
- No new wet-lab measurement on this page.
- PR #9 does not identify p_lactate or any other Theta.
- The 3-state ROS notebook remains unaudited.
- The three-arm protocol can fail; that would still be a result.
- OnCo counts move with buildDate.
- Grok may miss sources; insufficient is preferred to invented citations.
- Nigeria 2022 estimates are setting context only.
References
- Sung H et al. Global cancer statistics 2024. CA Cancer J Clin. 2026. doi:10.3322/caac.70090
- IARC. GLOBOCAN 2024 news. 8 July 2026. iarc.who.int
- WHO. Cancer fact sheet. 2026. who.int
- WHO. Global status report on cancer 2026. 9789240123977
- IARC. Nigeria fact sheet (GLOBOCAN 2022). gco.iarc.who.int
- NCI. Triple-negative breast cancer. cancer.gov
- NCI. Tumour heterogeneity. cancer.gov
- Altrock PM, Liu LL, Michor F. The mathematics of cancer. Nat Rev Cancer. 2015. PMC5663316
- Gatenby RA et al. Adaptive therapy. Cancer Res. 2009. PMC2676449
- Gomila J et al. OnCo. onco.cc
- Ogbonna K. Project Confluence PR #9. pull/9
- Validation protocol. validation_protocol.md