Why OncoTwin exists.

OncoTwin began with a question rather than a plan: what would it take to give a clinician in a resource-constrained ward the same kind of forward-looking, patient-specific insight available in the best-resourced cancer centres in the world, without assuming the connectivity, staffing, or infrastructure those centres take for granted.

That question came from working alongside the everyday reality of healthcare delivery in Kenya, where the distance between what medicine knows and what a given hospital can act on is often a matter of access, not ambition. Mechanistic models of tumour growth and treatment response have existed in the research literature for decades. Methods for calibrating those models to an individual patient using AI have matured rapidly in recent years. What was missing was a platform that brought the two together, and that was built first for the hospitals that needed it most, not retrofitted for them afterward.

Every cancer treatment protocol in use today was validated on a trial population, then applied by a clinician to one specific patient, with a different tumour biology, a different history, and different constraints than the cohort it came from. Translating population-level evidence into an individual decision has always depended on experience and judgment.

OncoTwin exists to give part of that translation a computational foundation: a model of the patient's own tumour, calibrated to their own data, that can be simulated forward under different treatment choices before any of them are tried for real.

OncoTwin's core models are mechanistic: systems of equations describing how a tumour grows and responds to treatment, grounded in established pharmacokinetic and pharmacodynamic theory. That choice is deliberate. A mechanistic model's behaviour can be inspected, constrained by known biology, and traced back to the assumptions that produced it, a property a black-box predictor doesn't share.

AI has a specific, narrower role inside that structure: it estimates the parameters of the mechanistic model for one specific patient, and it reports the uncertainty of that estimate alongside it. It calibrates the model to the patient. It does not decide what the model outputs, and it does not decide what a clinician does with that output.