image: ©BlackJack3D | iStock
Researchers at University College London (UCL) and the University of Cambridge have developed an artificial intelligence platform that creates personalised computational tumours, or “digital twins”, to identify effective treatments for hard-to-treat cancers
Published in Cell Reports Medicine, the study demonstrates how logic-based symbolic AI can analyse a patient’s genetic fingerprint to simulate therapeutic responses, successfully predicting treatment outcomes in 74% of evaluated clinical cases for Triple-Negative Breast Cancer (TNBC).
Explainable symbolic AI vs. traditional “Black Box” models
Clinicians have often approached traditional machine learning and Large Language Models (LLMs) with caution due to their “black box” architecture, where decision-making pathways remain opaque.
To establish clinical trust, the research team utilised symbolic AI, an interpretable framework grounded in explicit logical reasoning:
Transparent Logic
Unlike probabilistic deep learning, symbolic AI executes deterministic, rule-based operations. Oncologists can directly inspect and interrogate the step-by-step logic used by the software to arrive at a therapeutic recommendation.
Mechanistic Modelling
The program maps specific gene mutations, enzymatic pathways, and protein interactions, allowing doctors to evaluate whether suggested targeted therapies are clinically viable or potentially toxic.
Decision Support Integration
Designed as an advisory tool rather than an autonomous decision-maker, the system helps clinicians uncover secondary therapeutic targets to overcome drug resistance when primary treatments fail.
Validation via digital twins and clinical trial data
To validate the model, the team constructed 31 individualised digital twins using genetic profiles from patients enrolled in the PARTNER clinical trial (led by Cambridge University and Addenbrooke’s Hospital):
Simulation Phase:
Each digital twin underwent simulated treatment regimens consisting of chemotherapy alone, chemotherapy combined with the PARP inhibitor Olaparib, or an untreated control baseline.
Predictive Accuracy:
The computer simulation matched actual clinical trial responses with 74% accuracy.
Rescuing Non-Responders:
In 12 of 13 patient models that did not fully respond to standard therapies, the digital twin identified at least one additional targeted treatment that could restore drug sensitivity.
Complete Remission
Potential: For 10 of those non-responding cases, the model predicted that adding the suggested secondary therapy could eliminate all invasive cancer cells in the breast and regional lymph nodes.
Targeted mechanisms in triple-negative breast cancer
TNBC accounts for roughly 15% of all breast cancer diagnoses and presents significant therapeutic challenges because its cells lack oestrogen receptors, progesterone receptors, and the HER2 protein, rendering standard hormonal and HER2-targeted therapies ineffective.
Additionally, over 15% to 20% of TNBC cases exhibit Homologous Recombination Deficiency (HRD). While HRD enables tumours to mutate rapidly, it also creates a metabolic vulnerability. PARP inhibitors like Olaparib exploit this flaw by blocking single-strand DNA repair enzymes, causing double-strand breaks that lead to selective cancer cell death.