This post explains our published research: Stage-Aware Multi-Mechanism Optimization of Breast Cancer Treatment (Rai, 2026, DOI: 10.5281/zenodo.19845275). The code is open-source (CureNet on GitHub).
Designing a multi-drug cancer regimen is a brutal balancing act. Each added drug brings efficacy — and also toxicity, interactions with the other drugs, strain on the liver and heart, and cost. A human specialist juggles perhaps a handful of candidate combinations in their head. Our study asked what happens if you make the computer juggle all of them.
Concretely: a library of 53 compounds — standard chemotherapies, targeted drugs, repurposed medicines, phytochemicals — filtered to 48 clinically available agents, then combined four at a time and scored across three breast cancer subtypes and five disease stages. That is roughly 190,000 combinations, each evaluated by a nine-dimensional optimizer that weighs pharmacokinetic feasibility (can these drugs actually coexist in a body?), drug–drug interactions, cumulative toxicity, immune-related adverse events, diversity of mechanisms (four drugs hitting the same pathway is one drug in a trench coat), and cost.
Two findings stand out.
The first is an old, cheap drug that would not stop winning. Niclosamide — an anti-parasitic on the market for decades — appeared in 11 of the 15 optimal protocols. Suspicious of a bias (was the optimizer just picking it because it is cheap?), the study ran cost-bias diagnostics: the drug’s selection rate stayed at an identical seventy-three percent no matter how the cost weighting was turned up or down. It was being chosen for its mechanisms, not its price. The single highest-scoring protocol in the entire search was for one of the hardest situations in breast oncology — HER2-positive disease that has spread to the brain — pairing lapatinib and capecitabine (both brain-penetrant, both standard) with niclosamide and sulfasalazine, at a score of 0.805.
The second finding is sobering rather than exciting. For heavily pretreated disease — patients who have already been through several lines of therapy — the optimizer visibly ran out of good moves. When the mechanism classes a patient can still tolerate are restricted, no clever arithmetic conjures new options; the study concludes that this is precisely where investigational agents and clinical trials are not a luxury but the main road. Robustness testing (Monte-Carlo, wobbling every drug’s potency by ±50%) confirmed the recommended protocols are stable — but stability inside a shrinking option space is exactly why we keep telling patients to ask about trials early, while eligibility is still open.
As always: these are computational rankings, not prescriptions. Several drugs in the winning combinations are used off-label or are investigational in this setting. What a study like this is for is to give oncologists and tumour boards a systematically searched, honestly stress-tested shortlist — and to give patients sharper questions. Read the full paper free at zenodo.org/records/19845275, and if you want your own case mapped against the evidence, open a ticket below.