CTX stops asking how likely recurrence is and predicts who gains from chemotherapy, catching benefit a genomic score missed in TAILORx

CTX changes the question in adjuvant breast cancer decisions from how high the recurrence risk is to how much adding chemotherapy would gain, outputting CTX-τ, the gain in five-year recurrence-free probability for an individual patient. The 21-gene, 50-gene and 70-gene assays used to decide on chemotherapy are prognostic tools built to estimate recurrence risk and only later applied to chemotherapy decisions; recurrence risk and chemosensitivity are distinct properties of a tumor. In 1,994 held-out real-world patients from five cohorts in three countries, the treatment-by-biomarker interaction for CTX-τ was p=2.1×10⁻⁶; in the same data age, stage, grade, histological subtype and Ki67 showed no significant interaction, and the Oncotype DX recurrence score gave p=0.025.

The model chains a pathology foundation model with a causal inference module, predicting outcomes separately under endocrine therapy alone and under added chemotherapy, then taking the five-year difference as personalized chemosensitivity. Causal methods had not been successfully applied in oncology, attributed to insufficiently expressive inputs and inadequate data scale; pathology slides are gigapixel images that experts cannot annotate at scale. The pathology foundation model Falcon was pretrained self-supervised on more than two billion H&E image patches, and CTX was developed on 9,141 patients from twelve cohorts in nine countries; training penalizes drift between the model's internal pictures of the two treatment groups so it does not learn biases in real-world data.

In randomized trials, CTX-τ separated patients who benefit from chemotherapy from those who do not, and flagged high-risk patients for whom escalation should be considered. In real-world cohorts chemotherapy allocation is non-random and carries the biases of clinical judgement, so the model was locked and tested as a predictive biomarker in randomized trials. In TAILORx, the 11% identified by CTX-τ as high-benefit gained 6% absolute five-year disease-free interval benefit (p=0.011) while the remainder gained 0% (p=0.83), with interaction p=0.016 versus p=0.20 for the genomic score; in UNIRAD, the CTX-τ low-benefit group gained 10% in five-year disease-free survival from everolimus (p=0.005) while the high-benefit group did not (6% loss, p=0.48), with interaction p=0.01.

The demonstrated scope is adjuvant chemotherapy decisions in HR+/HER2− breast cancer: training on 9,141 patients from twelve cohorts in nine countries, evaluation on 1,994 held-out patients from five cohorts in three countries, and locked-model testing in the phase 3 TAILORx and UNIRAD trials; cross-cancer results come from zero-shot application to 6,692 patients across 16 other cancer types.

Cross-cancer transfer is not universal: risk predictions were associated with survival in 11 of 16 cancers, and CTX-τ correctly ordered benefit in only 5 of the 7 cancers treated with platinum- or taxane-based regimens, mostly failing in cancers treated with alkylators or targeted and immune therapies. The evaluation-set estimate of the same 6% recurrence rate with only 10.5% treated is a counterfactual calculation ranking patients by predicted benefit, not an observed allocation.