← Back to Home

Testing drugs on virtual tumors first: symbolic AI retrospectively predicts breast cancer response, with 74% agreement across 31 cases

Researchers used patients’ genetic data to build tumor “digital twins,” reproducing existing treatment responses and identifying potential drug options for those with incomplete responses. However, the improvements predicted by computer simulations still require validation in prospective clinical studies.

By SURL BioNews

When breast cancer does not respond to treatment as expected, choosing the next drug is often more difficult. A UK research team tried exposing patients’ tumors to different treatments in a computer to see whether genetic data could reveal new avenues. This small retrospective study offers preliminary clues, but it has not demonstrated that choosing drugs based on the model’s recommendations can improve patients’ actual outcomes.

The study, a collaboration between University College London and the University of Cambridge, was published on October 26 in *Cell Reports Medicine* and focused on triple-negative breast cancer. These tumors lack three common treatment markers—estrogen receptor, progesterone receptor, and HER2—making hormone therapies and HER2-targeted drugs unlikely to work and increasing the importance of identifying treatment vulnerabilities in individual tumors.

The team used “symbolic artificial intelligence”: reasoning through inspectable logical rules that allow clinicians to trace how the model arrives at its recommendations. Researchers used genetic data from 31 participants in the PARTNER clinical trial to build digital twins of their tumors, then separately simulated chemotherapy alone, chemotherapy combined with the PARP inhibitor olaparib, and no treatment.

When the simulations were compared with actual treatment responses in the trial, the model’s predictions matched in 74% of cases. This figure describes predictive performance within the dataset of 31 cases; it cannot be interpreted as a treatment success rate, nor is it sufficient to establish the model’s accuracy in other hospitals or different patient populations.

The researchers then focused on 13 patients with incomplete treatment responses. For 12 of them, the model identified at least one additional targeted therapy that might improve outcomes; further simulations even predicted that 10 might achieve a response with no residual invasive cancer in either the breast or the lymph nodes. However, these were all computer-generated treatment hypotheses, rather than observed outcomes after patients actually received additional treatment.

The model’s reasoning can be inspected, which may help clinicians assess whether its recommendations are reasonable. The research team also noted that even when the computer proposes a drug, clinicians may still rule it out because of factors such as toxicity. Interpretability creates room for discussion and revision, but cannot by itself ensure efficacy or safety.

The value of this work lies in bringing existing treatment responses and potential alternative strategies into a single reasoning process that can be scrutinized. Prospective studies are still needed to test whether the model can reliably predict outcomes before treatment begins, and whether choosing drugs according to its recommendations delivers real clinical benefits. Until then, digital twins are better suited to providing leads for research and treatment discussions.

References

  1. University College London
  2. Medical Xpress (provided by University College London)
  3. Bioengineer.org