Explainable AI (XAI) is emerging as a critical tool to help researchers understand which genetic factors drive AI predictions for breast cancer drug selection, yet significant barriers still prevent these AI-powered recommendations from reaching patients safely. A recent review published in npj Genomic Medicine examined how interpretable machine learning could accelerate drug repurposing for precision breast cancer treatment, revealing both the promise and the pitfalls of making AI decisions transparent in oncology.

Breast cancer remains one of the most common and deadly cancers globally, with an estimated 2.3 million new cases and 670,000 deaths worldwide in 2022. Developing entirely new medications typically requires more than 10 years and billions of dollars, creating a bottleneck for patients who need faster treatment options. AI and genomic technologies offer a potential shortcut by identifying existing, already-approved drugs that could be repurposed to treat aggressive breast cancer subtypes, potentially delivering therapies to patients more quickly and affordably.

What Is Explainable AI, and Why Does It Matter for Cancer Treatment?

Explainable AI refers to machine learning methods designed to reveal which data inputs most heavily influence a model's predictions. In cancer research, this transparency is essential because doctors and patients need to understand not just what a model recommends, but why. Two widely used XAI techniques are SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), which help researchers identify which genomic variables most affect AI model predictions.

However, interpretability tools come with an important caveat: they reveal correlation, not causation. These post-hoc explanations still require biological validation to confirm that the identified genomic factors actually drive the predicted drug response, rather than merely appearing correlated in the data.

How Can AI Help Researchers Repurpose Existing Cancer Drugs?

Drug repurposing involves finding new uses for medications already approved for other conditions. Because these drugs have established safety profiles, they can move into clinical testing far more rapidly than entirely new compounds. AI accelerates this process by integrating genomic, pharmacologic, and clinical data to identify which existing drugs might reverse cancer-associated pathways in specific tumor types.

Consider two real examples: metformin, a diabetes medication, activates a protein called AMPK and can inhibit a cancer-growth pathway known as PI3K/AKT/mTOR in breast cancer models. Statins, drugs prescribed to manage cholesterol, inhibit a pathway linked to cancer cell proliferation and metastasis. Advanced AI systems, including graph neural networks and transformer-based models, can now predict whether an existing drug might effectively target a specific genomic abnormality in a patient's tumor.

Steps to Validate AI-Predicted Cancer Drugs Before Clinical Use

Researchers propose an integrated framework that combines AI predictions with rigorous biological validation before any drug reaches patients. This approach emphasizes mechanistic validation rather than relying solely on statistical associations.

• Pathway Analysis: Researchers analyze the biological pathways affected by both the tumor's genetic mutations and the candidate drug to ensure the predicted mechanism makes biological sense.
• Molecular Docking Studies: Computational simulations test whether the drug molecule can physically bind to its intended protein target in the cancer cell.
• Experimental Validation: Cell cultures, patient-derived organoids, and animal models confirm that the drug actually inhibits cancer growth before human testing begins.
• Feedback Loop Integration: Data from experimental and clinical research is reintegrated into AI models to enhance future predictions and support adaptive learning, creating an iterative refinement process.

What Barriers Still Prevent AI Cancer Predictions From Reaching Patients?

Despite promising advances, several critical challenges limit clinical implementation. Many genomic datasets lack diversity and are heavily biased toward populations of European ancestry, reducing prediction reliability for underrepresented groups. This bias could reduce model performance in African, Asian, Latin American, and other underrepresented populations, potentially worsening inequities in breast cancer care.

Validation remains another major hurdle. Most computationally predicted drug candidates never progress from AI prediction to laboratory testing. Even drugs that have shown promise computationally, such as metformin and statins, have not consistently translated into clear clinical benefit in randomized controlled trials. This gap between computational prediction and real-world clinical outcomes underscores why experimental testing with cell cultures, patient-derived organoids, and animal models remains essential for establishing biological activity and pharmacological safety.

The complexity of breast cancer itself adds another layer of difficulty. Breast cancer is characterized by great variability between different tumors in different patients and even between tumors within the same patient. The primary molecular subtypes include luminal A and B, HER2 (human epidermal growth factor receptor 2)-enriched, and triple-negative breast cancer (TNBC), each requiring different treatment approaches.

How Can Researchers Address These Gaps?

The review concluded that integrating AI with genomic analysis offers a promising strategy for accelerating drug repurposing in breast cancer. Combining multi-omics data, which includes DNA mutations, gene expression, protein expression, and epigenetic regulation, with interpretable machine learning methods helps researchers prioritize candidate repurposed drugs and therapeutic hypotheses for validation.

However, before these technologies can be safely deployed in clinical settings, researchers must address challenges related to data diversity, reproducibility, experimental validation, and ethical governance. The goal is not to replace human expertise or clinical judgment, but to augment it, giving oncologists and patients a clearer understanding of why an AI system recommends a particular drug and what biological evidence supports that recommendation.

As AI continues to reshape precision oncology, the emphasis on explainability and mechanistic validation represents a maturation of the field. Rather than treating AI as a black box that produces predictions, the field is moving toward transparent, interpretable systems that clinicians and patients can understand and trust.