Let’s connect

Curetanics Cellular Response Model

CCRM: individual analysis, scientific methods and evidence

The Curetanics Cellular Response Model is in operational use. Its workflow begins with individual analysis, then connects the resulting profile with structured evidence, mechanistic modeling and traceable outputs for responsible expert review.

01 / Individual analysis and profile

Individual analysis: the starting point of CCRM

Individual analysis comes first. The profile brings together the relevant, available and lawfully usable information for the person and question being investigated. The categories below explain the information considered and its scientific role; they are not a mandatory test panel for every person. Source, timestamp, units, uncertainty and missingness are retained, because a missing measurement is not a normal value.

Tumor, pathology and molecular findings

Relevant information fields
Cancer site, histology, grade and stage; sample site and date; assay method and laboratory report; clinically available biomarkers. For sequencing: somatic versus germline context, variant notation, genome build, transcript, variant allele fraction and relevant quality information.
Scientific purpose
Define the biological system and what the sample can actually represent. Molecular findings may change the relevance of a hypothesis.
Interpretation and limits
Include only relevant existing results. Tumor sequencing does not automatically establish inherited status; one sample does not capture every tumor clone. No genomic test is ordered by this page.
[12]

Blood and organ-function measurements

Relevant information fields
When relevant: complete blood count with differential, creatinine and reported renal-function estimate, ALT/AST, bilirubin, alkaline phosphatase, albumin and other justified markers. Preserve numerical value, unit, laboratory range, method, sampling time and relation to treatment or acute illness.
Scientific purpose
Describe measured systemic context and potential exposure or safety constraints; retain trends rather than treating different sampling dates as simultaneous.
Interpretation and limits
These are candidate fields, not a required test panel. Blood values do not directly reveal all intracellular states or establish a compound’s safety. Missing and outdated data must remain visible.

Treatment and complete medication exposure

Relevant information fields
Exact anticancer regimen and cycle dates; current and previous drugs; active ingredients, formulation, route, prescribed amount and schedule, start/stop dates and as-needed use. Include relevant supplements, extracts and non-prescription products with product identity and timing.
Scientific purpose
Support a defined interaction question involving exposure, treatment phase, co-medication and relevant metabolites, rather than a simple drug-name match.
Interpretation and limits
Recording prescribed amounts is not dose calculation or advice. Reconciliation and clinical interpretation belong to the treating team; incomplete lists cannot establish compatibility.
[4]

Clinical history and patient-relevant outcomes

Relevant information fields
Relevant age, sex-related biological variables, body measurements, comorbidities, allergies, prior adverse reactions and treatment history. Record defined symptoms and validated patient-reported or functional measures where appropriate, including instrument, baseline and observation time.
Scientific purpose
Specify the intended population, important competing explanations, outcome definition and factors that may affect generalizability.
Interpretation and limits
Select fields for an explicit research question, not a maximal personal dossier. A global “health score” must not replace clinical assessment or validated endpoints.

Longitudinal, lifestyle and exposure context

Relevant information fields
Treatment phase, time since diagnosis or treatment, repeat measurements and symptom trajectories. If relevant to the question, document smoking, alcohol, dietary or activity observations and exposure history, with their source and measurement limitations.
Scientific purpose
Distinguish an enduring characteristic from a transient state and align observations with the period being investigated.
Interpretation and limits
Self-reports and proxies have uncertainty. These fields do not reconstruct a whole-person cellular state or prove that a behavior caused an individual cancer.

02 / Graph sources and contribution

Knowledge sources and their scientific roles

After individual analysis, evidence mapping relates the person’s context to molecular and clinical knowledge. The resources below explain the roles and limits of reference databases used in this field. The relevant source selection, version and evidence context must be interpreted through the CCRM documentation. Listing a source does not imply partnership, endorsement or unrestricted reuse; applicable licenses and access conditions remain relevant.

Reference resource · interpret in context

UniProt

Scientific role of the source
Protein accession mapping, sequence and functional annotations, with organism and isoform context. It can anchor protein nodes to stable identifiers.
Evidence limits
An accession, gene symbol and isoform are not interchangeable. Retain annotation provenance and version; a functional annotation does not demonstrate a patient-specific response.
[13]

Reference resource · interpret in context

Reactome

Scientific role of the source
Curated reaction and pathway relationships and pathway membership, with links to supporting biological knowledge. It can supply an explicit reaction/pathway vocabulary.
Evidence limits
Preserve species and inferred-versus-experimental context. Membership or enrichment does not prove that a pathway is active, causal or therapeutically actionable in the individual.
[2]

Reference resource · interpret in context

ChEMBL

Scientific role of the source
Compound–assay–target bioactivity records, assay types, measured activity values, units and target-assignment information. It can supply experimentally contextualized candidate edges.
Evidence limits
Assay formats and potency measures differ. Target-assignment confidence is not clinical certainty; IC50, Ki and Kd must not be silently treated as equivalent.
[14]

Reference resource · interpret in context

PubChem

Scientific role of the source
Chemical identity, structures, identifiers and links to deposited substances and bioassay information, retaining the contributing data source.
Evidence limits
A compound record can aggregate submissions from multiple providers. Resolve stereochemistry, formulation and source provenance; database presence is not evidence of anticancer efficacy.
[15]

Reference resource · interpret in context

IntAct

Scientific role of the source
Experimentally derived molecular interaction records curated from literature or depositions, with interaction participants and experimental context.
Evidence limits
An interaction depends on its assay and biological conditions. A reported physical association does not alone establish regulation, pathway direction or clinical relevance.
[16]

Reference resource · interpret in context

NCI PDQ interaction summaries

Scientific role of the source
A curated clinical-literature layer for documented or suspected food, supplement and cancer-drug interactions, separating study types and supporting reports.
Evidence limits
PDQ is not an exhaustive compatibility engine or a formal prescribing guideline. Missing entries are not negative evidence; primary reports and the treating team’s assessment remain necessary.
[4]

03 / Retrospective validation

Retrospective validation: methods and interpretation

Existing CCRM validation documentation addresses model evaluation and its scope. This section explains the scientific principles of retrospective evaluation against known interactions and measured outcomes; it does not reproduce the underlying CCRM reports or numerical results.

01

Define the claim before selecting a benchmark

Evidence extraction and mechanistic prediction are different tasks. An extraction test asks whether a system identifies a documented interaction and cites the correct evidence. A quantitative prediction test asks whether a specified model reproduces a measured exposure change under defined conditions. Freeze the task, intended population, input availability and success criteria before evaluating either. Correctly retrieving a known interaction does not validate a cellular simulator.

[17][5]
02

Construct an independently reviewed reference set

Reference records would need exact compound and drug identity, formulation, regimen, study population, timing, experimental conditions and the measured endpoint. Clinical pharmacokinetic evidence must remain separate from in vitro signals or case reports. Reviewers should label documented effects, adequately tested non-effects and inconclusive evidence explicitly. A pair missing from a database is not a negative example. Curated summaries can identify candidate records but do not replace the underlying study.

[4][17]
03

Separate development knowledge from the test

Freeze source versions, model versions and the analysis protocol. Split appropriately by source document, compound family, regimen and time; remove duplicates and keep preprocessing inside development data. Use an independently assembled external test set where feasible. For language-model systems, a familiar published interaction may already be in pretraining data: a random split of extracted facts does not establish independent prediction. Compare with transparent retrieval and simple mechanistic baselines.

[5][9]
04

Report errors and limits, not only successful examples

Report all prespecified cases, including missed interactions, unsupported alerts and indeterminate outputs. Evaluate performance by mechanism, evidence strength and relevant population rather than only one pooled number. Quantitative pharmacokinetic claims require matched observed exposure data and uncertainty reporting. Retrospective success would still need independent experiments and, for any clinical intended use, suitable prospective evaluation. It would not itself establish clinical utility or authorize compound use.

[17][5][9]

How evaluation results are interpreted

Evidence identification or classification
Report sensitivity/recall and positive predictive value with uncertainty; report specificity only against adequately defined negative cases. Show the reference labels and indeterminate outputs.
Rare interaction detection
If precision–recall AUC is used, publish test-set prevalence, sampling and thresholds. This metric depends on the positive-case proportion and is not comparable across arbitrary test sets.
Quantitative exposure prediction
For matched observed AUC ratios, assess error on a declared scale, such as log-ratio error, with bias and interval coverage where probabilistic intervals are generated. Interpretation depends on the evaluated task, exposure conditions and model version.

Published interaction literature provides scientific reference material, while CCRM’s own evaluation is described in its validation documentation. Evidence retrieval, mechanistic prediction and clinical utility answer different questions and must be interpreted separately. Compound applications require assessment and approval by the treating oncologist and medical team.

04 / EU regulatory context

EU regulatory qualification follows the intended use

Medical-device status depends on the software’s intended purpose and functionality. A statement that the organisation is not a medical provider does not settle that question.

“Not a medical provider” is not a medical-device exemption.

Define the medical purpose before assigning a class

MDR Article 2 includes software with specified medical purposes. Intended purpose is assessed through the manufacturer’s supplied information, including promotional statements and clinical evaluation. Software providing patient-specific information for oncology treatment decisions may therefore qualify as medical device software. A provider disclaimer or an oncologist’s final review does not automatically exclude it. Qualification must precede classification; independent software is classified in its own right.

[18][19]

Rule 11: IIa is a starting point, not a ceiling

For qualifying MDR software that informs diagnostic or therapeutic decisions, Rule 11 generally starts at class IIa. Potential consequences of those decisions can raise the class: serious health deterioration or surgery to IIb; death or irreversible deterioration to III. Not every software product qualifies, and Rule 11 also contains monitoring and other-software provisions. The applicable CCRM classification must be read in its documented intended-use and regulatory context, rather than inferred from a general website explanation.

[18][19][22]

Assess the actual function and intended use

The actual function and intended use determine the relevant assessment. General educational content and operational patient-specific decision support are assessed separately. If software interprets in vitro-derived genetic or other specimen data to fulfil an IVD purpose, the IVDR may be relevant instead of the MDR. Merely including blood or genomic data does not decide the applicable regulation; function and intended purpose matter.

[19][21]

Document the evidence and applicable conformity pathway

A medical implementation requires a documented intended purpose, qualification and classification rationale, technical documentation, risk management and the applicable conformity-assessment route. Clinical evaluation connects the scientific association, technical performance and clinical performance to that purpose. Software lifecycle, usability, cybersecurity and post-market obligations are also considered in the applicable regulatory documentation. A general scientific explanation does not replace these records.

[18][20][22]

Documentation concerning CCRM’s regulatory qualification and conformity assessment exists. Its conclusions must be interpreted for the specific functionality and intended use assessed. This overview does not itself state a particular risk class or claim CE certification. MDCG guidance supports interpretation but is not legally binding.

Scientific workflow

01

Begin with individual analysis

CCRM begins with analysis of the individual context using the relevant information available for that person: molecular and laboratory findings, medication and treatment exposure, medical history, lifestyle and timing. The profile and its data quality define the starting point for a specific scientific question. The biological system, compound or perturbation, comparator, outcome and observation period are then made explicit. Prevention, treatment tolerability and recovery require different endpoints and evidence. The outputs described here support bounded research hypotheses and responsible expert review; they do not replace the treating medical team. Transparent reporting, including task-specific evaluation, makes the scope of each analysis clear.

[5]
02

Represent observations as a time-indexed data layer

Let x(t) denote recorded observations and metadata at time t: laboratory or molecular measurements, treatment exposure, tissue or sample origin and relevant clinical context. Preserve units, assay methods, collection times, provenance and uncertainty. Missing values remain unknown unless an explicit, justified imputation procedure is documented. Blood measurements do not directly reveal every tumor cell’s molecular state. A model must state which variables are observed, which are inferred and which are unavailable. FAIR principles guide discoverable and reusable data records, while access and consent constraints still govern sensitive information.

[1][7]
03

Build a typed evidence graph with exposure constraints

Following individual analysis, evidence mapping relates the resulting profile to identified compounds, proteins, reactions and pathways. Typed relationships distinguish binding, activation, inhibition and association. Each relationship is interpreted through its source, experimental system, supported direction and relevant conditions. Curated pathway resources provide biological structure, while pathway membership alone does not demonstrate an individual response. Exposure context includes formulation, metabolites and absorption, distribution, metabolism and elimination assumptions. Total, unbound and intracellular concentrations are distinguished. Drug–supplement interaction concerns and contradictory evidence remain visible rather than being reduced to one compatibility score.

[2][4]
04

Specify a mechanistic state model and observation model

Mechanistic state models can be expressed, for example, as dz/dt = S·v(z,θ,u,t), with y(t) = H(z(t)) + ε. Here z represents modeled molecular states, S reaction stoichiometry, v reaction rates, θ kinetic parameters, u specified inputs, H a mapping to measured observables and ε measurement error. These generic equations explain the mathematical reasoning; they do not reproduce the complete CCRM implementation. A reproducible numerical model is specified through selected reactions, units, initial conditions, parameter sources, solver settings and an experiment record. Sparse observations may leave parameters unidentifiable, so a fitted trajectory does not uniquely establish a biological mechanism.

[3][6][7]
05

Give each AI component a defined and testable task

AI-supported work is described through defined tasks such as literature retrieval, entity mapping and relationship extraction, with source verification and expert review. A learned predictor is evaluated against its stated inputs and target, training data, architecture, optimization objective and independent test protocol. The documentation of these components is separate from this public explanation; the website does not specify a neural-network type, parameter count or training history. Generated text is not itself a biochemical simulation or experimental evidence. Data partitions, leakage controls, error analysis and generalizability checks are matched to the particular task.

[1][5]
06

Return a hypothesis record and preserve the uncertainty

The output record connects a defined question with supporting and conflicting sources, the applicable experimental context, model version, parameter choices, exposure assumptions and unresolved interaction questions. Sensitivity to uncertain inputs and alternative model structures is considered; a scenario range is distinguished from a statistical confidence or credible interval. A safety flag prompts review rather than authorizing use, and an explanation is not causal proof. The scope of experimental and clinical evidence must match the claim being made. Any compound application requires the treating oncology team’s assessment.

[4][5][6][7]

A representation of mechanistic modeling

This reaction-based formulation explains a class of mechanistic methods. It is an educational representation, rather than a full specification of the operational CCRM implementation.

dz/dt = S · v(z, θ, u, t)y(t) = H(z(t)) + ε
x(t)
Observed measurements and context, with time, units, sample origin and missingness.
z(t)
Selected modeled molecular states; not an observable whole-person cellular state.
S and v
Reaction stoichiometry and rate laws; each reaction and its units must be defined.
θ and u(t)
Kinetic parameters and specified perturbations; sources and uncertainty must be documented.
H and ε
The observation map and measurement error; predictions must be compared with measured observables.

A numerical solution requires a defined network, rate laws, parameter values, initial conditions and a simulation protocol. These details belong to the relevant technical model documentation; the equations here explain the scientific principle.

Evaluation principles

How CCRM evidence is assessed

CCRM is supported by existing technical and evaluation documentation. The criteria below explain how reproducibility, evidence quality, model performance and the scope of clinical claims are assessed. They do not imply that every evaluation has the same intended use or establishes the same conclusion.

1 · Computational verification and reproducibility

Assessment criterion
Archive equations, code, model and data versions, parameters with units, initial conditions, solver settings and the execution environment. Check numerical convergence and replay results independently. A reproducible calculation still requires separate biological validation.
Evaluation endpoint
Prespecified numerical tolerances; reproducible outputs; documented verification failures.
[6][1]

2 · Evidence extraction and graph curation

Assessment criterion
Create an independently annotated reference set with explicit entity, relationship and provenance definitions. Hold out source documents from development; assess extraction and human curation separately. Review contradictory, missing and unsupported edges rather than counting retrieved papers.
Evaluation endpoint
Precision, recall, agreement and traceability coverage on a held-out reference set.
[1]

3 · Retrospective molecular hypothesis evaluation

Assessment criterion
Freeze the intended molecular task and analysis before testing. Use appropriate compound-family, biological-system and temporal splits, with an external dataset where possible. Prevent duplicate and preprocessing leakage; compare with simpler baselines and a version omitting individual context.
Evaluation endpoint
Task-specific prediction error or retrieval metrics, uncertainty, and performance by held-out group.
[1][5][9]

4 · Independent experimental testing

Assessment criterion
Test prespecified hypotheses in relevant experimental systems at justified exposures, using authenticated materials, negative and positive controls, replication and blinded assessment where feasible. Measure the proposed mechanism and adverse effects; assess effects of concurrent cancer-treatment exposure when relevant.
Evaluation endpoint
Reproducible perturbation effects at relevant exposures, with failed hypotheses and harms reported.
[8]

5 · External evaluation of clinical prediction tasks

Assessment criterion
For a defined clinical prediction task, specify the population, outcome, time horizon and intended decision. Evaluate independent cohorts with representative measurements; report calibration, discrimination, missing-data handling, subgroup performance and risk of bias, without reusing evaluation data for development.
Evaluation endpoint
Externally assessed calibration, discrimination and applicability for the stated clinical task.
[5][9]

6 · Prospective clinical utility for a defined intended use

Assessment criterion
Before claiming clinical benefit, evaluate the complete intended workflow, including clinician interaction, safety and error cases. Choose a prospective design and comparator suited to the claim; use randomization where appropriate to estimate patient benefit, harms and effects on care.
Evaluation endpoint
Prespecified patient-relevant outcomes, adverse events and workflow effects against a suitable comparator.
[10][11]

Evaluation and reporting are specific to the model version, task and intended use. Scientific reporting frameworks support assessment; they do not by themselves certify a model. Evidence for one task does not automatically validate every other application. Model outputs must not replace the treating oncologist and medical team.

Scientific references

  1. Wilkinson et al. (2016) · The FAIR Guiding Principles for scientific data management and stewardship · Scientific Data 3:160018 ↗
  2. Reactome · Pathway Browser: curated reactions, pathways and experimental overlays ↗
  3. Hucka et al. (2019) · SBML Level 3 Version 2 Core specification ↗
  4. National Cancer Institute · Cancer Therapy Interactions With Foods and Dietary Supplements · Health Professional PDQ ↗
  5. Collins et al. (2024) · TRIPOD+AI statement: updated guidance for reporting clinical prediction models · BMJ 385:e078378 ↗
  6. Waltemath et al. (2011) · Minimum Information About a Simulation Experiment (MIASE) · PLOS Computational Biology 7:e1001122 ↗
  7. Raue et al. (2009) · Structural and practical identifiability analysis of partially observed dynamical models by exploiting the profile likelihood · Bioinformatics 25:1923–1929 ↗
  8. OECD (2025) · Guidance Document on Good In Vitro Method Practices (GIVIMP), Second Edition ↗
  9. Moons et al. (2025) · PROBAST+AI: quality, risk of bias and applicability assessment for prediction models · BMJ 388:e082505 ↗
  10. Vasey et al. (2022) · DECIDE-AI: reporting early-stage clinical evaluation of AI decision support · Nature Medicine 28:924–933 ↗
  11. Liu et al. (2020) · CONSORT-AI: reporting clinical trials of interventions involving artificial intelligence · Nature Medicine 26:1364–1374 ↗
  12. National Cancer Institute · Biomarker Testing for Cancer Treatment ↗
  13. UniProt · Stable accession numbers ↗
  14. ChEMBL · Assay and activity data documentation ↗
  15. PubChem · Data sources and provenance ↗
  16. IntAct · User Guide: experimentally derived molecular interactions ↗
  17. FDA · ICH M12 Drug Interaction Studies ↗
  18. Regulation (EU) 2017/745 (MDR) · Article 2, Article 61 and Annex VIII, including Rule 11 ↗
  19. MDCG 2019-11 rev.1 (June 2025) · Qualification and classification of software under MDR and IVDR ↗
  20. MDCG 2020-1 (March 2020) · Clinical evaluation / performance evaluation of medical device software ↗
  21. Regulation (EU) 2017/746 (IVDR) · Article 2 and Annex VIII ↗
  22. MDCG 2021-24 rev.1 (April 2026) · Guidance on classification of medical devices ↗
  23. Mathijssen et al. (2002) · Effects of St. John’s wort on irinotecan metabolism · Journal of the National Cancer Institute 94:1247–1249 ↗
  24. Frye et al. (2004) · Effect of St John’s wort on imatinib mesylate pharmacokinetics · Clinical Pharmacology & Therapeutics 76:323–329 ↗

These publications support the methods, reporting standards and clinical constraints. They are not studies validating CCRM.

← Back to homepage · Technology