Predicting Treatment Response in Breast Cancer with Artificial Intelligence

Personalized Treatment and Artificial Intelligence in Breast Cancer Care

Breast cancer has several different subtypes. Some tumors are sensitive to hormones, such as estrogen or progesterone receptor-positive tumors, while others may grow because of excess amounts of a protein called "HER2." In particular, "HR-positive, HER2-negative" is the most common subtype. When this subtype is metastatic, treatment options generally combine hormone-suppressing drugs with specialized treatments called CDK4/6 inhibitors. This combination generally improves treatment outcomes, but not every patient benefits equally.

Why do some patients respond well to treatment while others benefit less than expected?

This is where "artificial intelligence" comes into play. Artificial intelligence refers to computer systems that imitate human intelligence to solve complex problems. Machine learning is a branch of artificial intelligence that enables computer systems to gain experience through data analysis and make predictions by learning from these data. In other words, machine learning allows systems to develop their own decision-making mechanisms from patterns in data without needing preprogrammed rules.

At the San Antonio Breast Cancer Symposium (SABCS 2024) researchers presented a machine-learning model developed by Memorial Sloan Kettering Cancer Center that analyzes patients' clinical data, such as age and metastatic status, and genomic data, such as genetic mutations, to predict which patients will respond better to treatment. The aim is to offer patients the most effective treatment plan while avoiding unnecessary adverse effects and treatment costs.

With this new model, patients will be able to find a clearer answer to the question, "What is the most appropriate treatment plan for me, considering my genetic makeup and clinical characteristics?" For doctors and patients alike, these personalized approaches may signal a promising era in breast cancer treatment.

Study Details

Research Tool: OncoCast-MPM

This machine-learning model, developed by Memorial Sloan Kettering Cancer Center, combined clinical and genomic data to predict disease course. The study developed three different models:

1. A model based on clinicopathological data
2. A model based on genomic data
3. A model integrating clinical and genomic data

Patient Groups and Data Used

• Training group: 761 patients
• Test group: 326 patients
• Genomic analysis used: MSK-IMPACT testing to identify gene mutations in tumor samples obtained before treatment or within 2 months of starting treatment

Findings

1. Model Based on Clinicopathological Data: This model divided patients into three risk groups according to their clinicopathological characteristics:

• High-risk group: Median progression-free survival (PFS): 6.3 months
• Intermediate-risk group: Median PFS: 15.2 months
• Low-risk group: Median PFS: 24.5 months

2. Model Based on Genomic Data: Analysis using genomic mutations and biomarkers again identified three risk groups:

• High-risk group: Median PFS: 9.9 months
• Intermediate-risk group: Median PFS: 18.1 months
• Low-risk group: Median PFS: 23.1 months

3. Integrated (Clinical + Genomic) Model: This model combined clinical and genomic data to identify four different risk groups:

• High-risk group: Median PFS: 5.3 months
• Low-risk group: Median PFS: 29 months
• Two intermediate-risk groups: Median PFS: 10.7 months and 19.8 months

The Advantage of the Integrated Model

The integrated model stood out by distinguishing more clearly between the high-risk and low-risk groups. In this model, the difference between these groups was 6.5-fold , whereas in models based only on clinicopathological or only on genomic data, the difference was limited to 3-4-fold .

Key Markers

The machine-learning model identified the following factors as predictors of treatment response:

Genomic Markers (Indicators of Poor Prognosis):

• TP53 loss: Loss of a gene important for DNA repair
• MYC amplification: Increased copies of a gene that promotes cell proliferation
• PTEN loss: Loss of a tumor suppressor gene
• RTK-MAPK pathway alterations: Mutations in cell-growth signaling pathways
• RB1 loss: Plays a role in the development of resistance to CDK4/6 inhibitors
• Whole-genome duplication: Doubling of genetic material
• High loss of heterozygosity: Reduced genetic diversity

Clinical Markers:

• Liver metastasis
• A treatment-free interval of less than 1 year
• Progesterone receptor negativity
• Low estrogen receptor expression
• Presence of visceral (internal organ) metastases

The ease of identifying these factors at metastatic diagnosis makes the model broadly applicable in clinical practice.

Future Applications and Plans

Models of this type may reshape treatment strategies. For patients in the high-risk group:

• Close clinical follow-up
• Monitoring disease course through liquid biopsy and tumor-derived biomarkers
• Early referral for second-line treatment and clinical trial options

Because this study was a retrospective analysis, the model needs validation in a larger patient population. The MSK team is therefore testing it with external datasets and developing an online prediction tool that could be made available to clinicians.

Study Limitations

• Single-center study: Limited to data from Memorial Sloan Kettering Cancer Center.
• Retrospective analysis: Lack of prospective studies.
• Referral bias: Patients attending a specialist cancer center may differ from the general population.

The research team aims to improve the model's reliability using international databases to address these limitations.

Conclusion

This study showed that combining clinical and genomic data may allow more accurate identification of individual risk groups in breast cancer treatment. Developing personalized treatment approaches to improve the effectiveness of CDK4/6 inhibitors could usher in a new era in breast cancer care.