Breast cancer is the most common cancer in women, and about 70% of cases belong to the hormone receptor-positive, HER2-negative subtype. All such patients receive endocrine therapy, but adding chemotherapy helps only some of them. Existing genetic tests, such as Oncotype DX, assess the risk of recurrence rather than the benefit of chemotherapy, which leads to overtreatment. This work sits among studies that predict cancer outcomes from medical images and genetic data. Neighboring studies use deep learning to forecast recurrence or response to therapy, but almost all of them remain prognostic rather than predictive: they do not estimate how much a specific treatment will help a specific patient. This paper fills that gap by applying a causal approach to multimodal data.
The CTX model combines two data sources: histology slides of the tumor, processed by the pretrained neural network Falcon (trained on two billion tissue patches), and standard clinical features such as age, tumor stage, and receptor status. To estimate the individual treatment effect, the authors used a counterfactual regression network (CFRNet), which aligns feature distributions between groups of patients who did and did not receive chemotherapy. This allows the model to predict the probability of recurrence for each patient under two scenarios: endocrine therapy alone and endocrine therapy plus chemotherapy. The difference between these predictions is the personal benefit of chemotherapy. The model was trained on data from 9,141 patients across 12 cohorts in 9 countries and tested on 1,994 patients from 5 independent cohorts.
The model showed nearly perfect calibration: predicted recurrence probability matched the actual event rate. Discrimination reached an AUC of 0.711 for the 5-year horizon and 0.765 for the 10-year horizon. The key metric, the interaction between predicted chemotherapy benefit and actual treatment, was statistically significant (p = 8.11×10⁻⁵). Among patients with high predicted benefit, chemotherapy reduced the risk of recurrence by 58% (HR = 0.42), whereas among patients with low predicted benefit there was no effect (HR = 1.40). In a direct comparison with Oncotype DX on the same patients, CTX showed a stronger association with treatment benefit (p = 9.68×10⁻⁴ versus p = 0.0247). The model also revealed morphological and molecular correlates: tumors with high predicted sensitivity to chemotherapy were enriched for genes involved in the cell cycle, replicative stress, and immune response.
The work differs from neighboring studies in that it directly estimates the personal benefit of treatment, not just prognosis. Most existing models, including those that use deep learning on histology, predict the risk of recurrence regardless of therapy, which does not help decide who needs chemotherapy. Here, a causal model separates prognosis from treatment sensitivity, which fundamentally changes clinical use. In addition, the model works on routine data (histology and clinical records), without expensive genetic tests, which broadens access to personalized oncology.
The approach can be extended to other cancer types and other treatments. The authors showed that a model trained only on breast cancer transfers without fine-tuning to 16 other cancer types, especially those where similar cytotoxic drugs are used. In the future, spatial molecular data or circulating tumor DNA could be added to improve predictions. It may also be possible to train the model to estimate the effect of targeted therapy and immunotherapy.
The results will affect clinical oncology by helping to reduce overtreatment with chemotherapy and direct it to those who will truly benefit. This will also accelerate the development of new biomarkers and change the design of clinical trials, where patient selection will become more precise.
Prospective validation of the model in randomized clinical trials is needed to confirm its predictive accuracy under controlled conditions. It is also worth exploring whether the model can predict the effect of specific chemotherapy regimens and combinations with targeted drugs.
Oncology still faces the unresolved problem of identifying patients who will truly benefit from adjuvant chemotherapy. Existing prognostic biomarkers mix recurrence risk and treatment sensitivity, which leads to overtreatment. This work directly addresses that challenge by separating prognosis from treatment effect at the individual level.
The Falcon model used to analyze histology slides was trained on two billion tissue patches, roughly 40,000 whole slides, comparable to the archive of a major cancer center accumulated over decades.
The idea of predicting an individual's response to treatment echoes the film Gattaca, where genetic analysis determines a person's fate. Here, however, it is not about fatalism but about the ability to choose the optimal therapy based on data.