Using digital twins and AI to reimagine cancer treatment.
Every patient is unique. Their treatment should be too.
Cancer treatment still starts with an average, not a patient.
Every regimen in use today was validated on a population, a trial cohort with a shared diagnosis but different tumour genetics, different comorbidities, different histories. The clinician's task is to take that population-level evidence and apply it to one specific person. OncoTwin exists to make that translation computational instead of purely experiential.
Population trials, individual patients
Standard-of-care protocols are built from cohort averages. The patient in front of a clinician is not an average. They're one biology, with one history.
Trial-and-error dose-finding
When a first-line regimen doesn't fit, adjustment often happens by observing the patient on it, after the fact, not before.
Feedback arrives late
Whether a treatment path is working is frequently visible only weeks in, after imaging or lab results, not at the moment the decision is made.
A computational twin of the disease, not just the diagnosis.
A digital twin, here, is a patient-specific model, calibrated to one person's data, capable of being simulated forward under different treatment choices before any of them are tried on the patient. It's built from three layers working together.
Clinical & tumour data
Imaging, lab markers, treatment history, and where available, genetic and histological data: the raw signal a twin is built from.
Mechanistic models
Systems of equations that describe how a tumour grows and responds to treatment over time, grounded in established tumour biology.
AI-assisted calibration
Patient-specific parameters are estimated from that patient's own data, with the uncertainty of that estimate quantified, not hidden.
A living model that updates as the patient does.
A twin isn't built once and left static. Each new scan, lab result, or clinical observation feeds back into recalibration: the model stays current with the patient, and the loop closes with every visit.
Data in
Imaging, labs, treatment history, and biomarkers, structured into a patient-specific case file.
Calibrate
The mechanistic model is fit to that patient's own trajectory, not a population average.
Simulate
Candidate treatment paths are run forward against the calibrated model and compared.
Review & decide
A clinician reviews the projections alongside their own judgement and chooses a path.
Equations grounded in tumour biology, not a black box.
OncoTwin's core is mechanistic: systems of differential equations that describe tumour growth and treatment response over time, built from established pharmacokinetic and pharmacodynamic theory, not a pattern-matcher trained to guess an outcome from historical records.
That distinction matters clinically. A mechanistic model's behaviour can be inspected, constrained by known biology, and reasoned about: a projection can be traced back to the assumptions that produced it.
dV/dt = r·V·ln(K/V) − c·C(t)·V
- V
- Tumour volume
- r
- Growth rate
- K
- Carrying capacity
- c
- Treatment efficacy
- C(t)
- Drug concentration
A Gompertzian growth term with a treatment kill term, shown to illustrate the kind of relationship these models encode: the calibrated parameters (r, K, c) are what make it patient-specific.
AI calibrates the model. It doesn't replace it.
It would be easy to build a system that predicts outcomes directly from historical data. OncoTwin deliberately doesn't. AI's role here is narrower, and more accountable, than that.
AI calibrates
It fits the mechanistic model's parameters to one patient's data. It doesn't generate the prediction on its own.
AI estimates patient-specific parameters
Growth rate, treatment sensitivity, and related terms are inferred from that patient's own trajectory.
AI quantifies uncertainty
Every projection carries a confidence range. The model reports what it doesn't know, not just a single number.
AI assists, clinicians decide
The output is a projection for a clinician to weigh alongside their own judgement, never an instruction.
Five layers, one pipeline. Explore how they connect.
Select a layer to see what it does and how it hands off to the next.
Data layer
- Structured intake: history, biometrics, treatment record
- Imaging and lab markers, linked to the same case file
- Genetic and histological data where available
From data to decision, in one continuous loop.
Six steps, repeating for as long as a patient is on treatment: not a one-time report, but a model that stays in the loop.
Data intake
Clinical history, imaging, and lab markers are structured into a case file.
Twin calibration
The mechanistic model is fit to this patient's own data.
Scenario simulation
Candidate treatment paths are simulated forward and compared.
Clinician review
Projections, and their uncertainty, are presented for clinical judgement.
Treatment decision
The clinician selects a path. OncoTwin informs, it doesn't decide.
Outcome feedback
New results flow back in, and the twin recalibrates for what comes next.
What this looks like for one patient.
An illustrative timeline. Real pacing varies by diagnosis, regimen, and how quickly response data becomes available.
Diagnosis
A new case is opened. Imaging, labs, and history are recorded.
Twin created
A patient-specific model is calibrated from that intake data.
Scenarios compared
Two or three candidate regimens are simulated and reviewed with the clinician.
Treatment begins
The clinician selects a path, informed by the twin's projections.
Twin recalibrates
New scans and labs feed back in; the model updates to reflect the real response.
Plan adjusts if needed
If the twin's updated projection diverges from the original plan, the clinician revisits it.
Where OncoTwin is, and where it's going.
We're early. Being direct about that, and about what still needs validating before any clinical claim, is part of how we intend to earn trust in this space.
Foundation
- Mechanistic model architecture for priority cancer types
- Structured clinical intake and case-file design
- AI calibration framework with uncertainty quantification
Validation
- Retrospective validation against historical treatment outcomes
- Pilot deployments with partner clinicians
- Expanded coverage across additional cancer types
Scale
- Multi-site deployment across East African hospitals
- Continuous recalibration from ongoing care data
- Open research collaboration and peer-reviewed publication
A model that admits what it doesn't know.
Precision oncology tools carry real weight. These are the commitments we hold ourselves to while building one.
Clinician-in-the-loop, always
OncoTwin never makes a treatment decision. It projects outcomes for a clinician to weigh: the decision, and the accountability, stay human.
Uncertainty is shown, not hidden
Every projection carries a confidence range. A model that doesn't know is more useful admitting it than pretending certainty.
Limitations are disclosed openly
Where a model hasn't been validated for a cancer type, stage, or population, that gap is stated, not papered over with a confident number.
Patient data is minimized and protected
Only the data a model needs is collected. It stays local by default, and any wider use requires explicit, informed consent.
We're building this with people, not in isolation.
OncoTwin doesn't have a roster of partners to name yet. We're early, and we'd rather earn those relationships than list them prematurely. If your work overlaps with any of this, we want to hear from you.
Start a conversationHospitals & clinical partners
Oncology wards and clinicians willing to pilot, stress-test, and shape the workflow against real cases.
Research institutions
Groups working in mechanistic modelling, computational biology, or oncology who want to validate or extend the models.
Funders & mission-aligned partners
Organisations investing in health technology for under-resourced settings, particularly across Africa.
Engineers & researchers
People who want to work on the modelling, calibration, or product itself, as contributors or as a team.
What's happening at OncoTwin.
Answers, before you ask.
The questions clinicians, researchers, and reviewers ask first.
A patient-specific computational model built from that patient's clinical, imaging, and lab data. It can be simulated forward under different treatment choices before any of them are tried on the patient.
Nairobi-based. Reachable anywhere.
Clinical pilot, research question, or funding conversation: pick whichever channel is easiest.
The beginning of a new chapter in precision oncology.
We're building OncoTwin in the open, with clinicians, researchers, and institutions who want to shape it alongside us.