Breast cancer is a multifactorial disease influenced by genetic variants, hormonal factors, reproductive history, and lifestyle. Most existing risk prediction models treat these factors as either independent or linearly additive, assumptions that fail to capture the complex biological interactions that actually drive individual risk. A polygenic risk score (PRS) aggregates the effects of many genetic variants identified in genome-wide association studies (GWASs), but it assumes SNPs act independently and additively, missing potentially important interaction effects.
More than 200 SNPs have been identified as significantly associated with breast cancer risk through large-scale GWASs. However, individual SNPs have small effect sizes, and when analyzed in isolation, their combined predictive value remains limited. The key insight motivating this study is that breast cancer risk may emerge not from individual SNPs or individual demographic factors alone, but from networks of interacting features across both genetic and demographic domains.
Demographic risk factors related to familial history (such as having a first-degree relative with breast cancer) and to oestrogen metabolism (such as menopausal status, oral contraceptive use, and reproductive history) are independently associated with breast cancer risk. These biological mechanisms are distinct from genetic susceptibility encoded by SNPs, and combining them could provide complementary predictive information.
This Finnish study from the Kuopio Breast Cancer Project (KBCP) proposed a machine learning approach using XGBoost with adaptive iterative feature selection to identify the best-performing networks of interacting SNPs and demographic features for breast cancer risk prediction, going beyond linear models to capture non-additive interactions across high-dimensional genetic and clinical data.
The Kuopio Breast Cancer Project (KBCP) dataset contained 445 breast cancer cases and 250 age-matched and area-matched controls, totaling 695 subjects, all Finnish women. Controls were carefully selected to match cases by age and long-term residence to minimize confounding from population stratification. Genotyping was performed using the Illumina iCOGS custom array with 211,155 SNPs, of which 125,041 passed quality control for analysis.
Demographic risk factors were organized into two predefined groups based on biological mechanism. Group 1 consisted of 9 features related to familial history of breast cancer, including whether any family member had cancer, the type of cancer in each affected family member, whether affected relatives were first-degree relatives, the number of family members with breast cancer, and bilateral versus unilateral tumor status. Group 2 consisted of 11 features related to oestrogen metabolism, including oral contraceptive use and duration, menopausal status, breastfeeding history, pregnancy history, hormone replacement therapy duration, menstrual cycle length, pregnancy lengths, and BMI.
The machine learning approach operated in two phases. First, an XGBoost model was applied to the full SNP dataset to rapidly identify candidate SNPs with the highest initial importance for breast cancer risk. Second, an adaptive iterative search algorithm combined the top-ranked SNPs with each demographic group separately to identify the optimal networks of interacting genetic and demographic features that maximized prediction accuracy on held-out validation data at each iteration.
For evaluation, the algorithm was run with 10 iterations of 5-fold stratified cross-validation, maintaining class balance across folds to handle the dataset imbalance (445 cases versus 250 controls). Since the dataset was imbalanced, mean average precision (mAP) derived from precision-recall curves was used as the primary evaluation metric rather than AUC, because precision-recall curves provide more informative assessments of model performance on imbalanced classification problems.
Combining interacting SNPs with Group 1 (familial history) features achieved a mean average precision of 77.78, substantially outperforming models using only Group 1 features (mAP = 74.19) or only the data-driven interacting SNPs (mAP = 73.65). This demonstrates that the genetic and familial history components provide genuinely complementary predictive information rather than redundant signals.
Combining interacting SNPs with Group 2 (oestrogen metabolism) features achieved a slightly higher mAP of 78.00, again outperforming models using only Group 2 features (mAP = 72.57) or only SNPs (mAP = 73.65). Combining all Group 1, Group 2, and genetic features together achieved the highest mAP of 78.36, confirming that each risk factor category contributes additive predictive value beyond what the others can provide.
A key finding was that models using data-driven interacting SNPs substantially outperformed the PRS model based on 82 published breast cancer-risk-associated SNPs (mAP = 70.02). The combination of published SNPs plus demographic features (mAP = 68.20 for Group 1, 71.36 for Group 2) also performed worse than combining data-driven interacting SNPs with demographic features (mAP = 77.78 and 78.00 respectively). This highlights that allowing the model to discover new interacting SNPs from data outperforms restricting analysis to pre-selected variants from the literature.
Feature importance analysis using leave-one-out perturbation showed that removing any demographic feature reduced mAP, confirming that each feature contributed unique predictive value. Among Group 1 features, having cancer in the family and the cancer type were individually most important (4.5 percent and 3 percent mAP reduction respectively). Among Group 2 features, oral contraceptive use and menopausal status were most important (4 percent and 3 percent reduction). Individual SNPs caused smaller reductions than individual demographic features, confirming that SNPs contribute most effectively as interacting networks rather than individually.
The SNPs identified as interacting with Group 1 (familial history) features were associated with 86 genes. A gene interaction map built from these genes revealed three cancer-related networks, most notably networks centered on EGFR, BLK, ATR, and WWOX. EGFR signaling is a well-established driver of tumor growth, angiogenesis, metastasis, and apoptosis inhibition, and high EGFR expression in breast cancer has been linked to BRCA1/2 mutations. ATR encodes a DNA damage sensor that phosphorylates tumor suppressors including BRCA1, CHEK1, and TP53. WWOX is a tumor suppressor gene involved in apoptosis regulation that is deleted in up to 80 percent of breast cancer cases.
The SNPs interacting with Group 2 (oestrogen metabolism) features revealed a prominent network centered on ESR1, the primary oestrogen receptor gene, which is one of the most important genes in oestrogen-driven breast cancer biology. The top-ranked SNP in this analysis, rs11757540, is directly linked to the ESR1 gene. The FGFR2 gene also appeared prominently: a locus within FGFR2 is consistently identified as the genetic variant most strongly associated with oestrogen receptor-positive breast cancer risk across independent GWAS analyses, and FGFR2 transactivates HER2 via c-SRC, potentially contributing to HER2-targeted therapy resistance.
The SHAP value analysis of individual feature contributions revealed biologically consistent patterns. Among Group 1 features, having cancer in the family, breast cancer type in a family member, more family members with breast cancer, and a higher familial risk score all increased the model's prediction of cancer status, as expected from established clinical knowledge. Among Group 2 features, menopause and shorter pregnancy length increased predicted risk, while breastfeeding showed a protective association in this Finnish dataset.
An interesting finding was that oral contraceptive use was associated with decreased breast cancer risk in this model, opposite to its commonly reported slight risk-increasing effect. This likely reflects the specific characteristics of the KBCP population, where most cases (67 percent) had not used oral contraceptives, causing the model to learn this data-specific pattern. The authors note that changes in oral contraceptive formulations over time may contribute to discrepancies between this dataset collected over several decades and contemporary populations.
The central methodological contribution of this study is demonstrating that non-linear feature interactions captured by gradient boosting outperform linear additive models for breast cancer risk prediction. Classical approaches including logistic regression and polygenic risk scores assume that genetic and demographic risk factors act independently and combine in a simple additive way. In reality, genetic variants and lifestyle factors interact in complex biological ways that additive models systematically underestimate.
The finding that machine learning-discovered interacting SNPs outperform published established risk SNPs combined with demographic features is particularly important. It implies that the set of 82 GWAS-identified SNPs used in standard PRS models does not fully capture the genetic architecture of breast cancer risk, and that data-driven discovery of new interacting variants can identify biologically relevant signals not yet indexed in the literature.
The study's approach offers a framework applicable to any multifactorial disease, not just breast cancer. The combination of XGBoost for initial feature ranking with adaptive iterative search for finding interacting networks addresses the high-dimensional nature of genomic data (tens to hundreds of thousands of SNPs) while avoiding the computational intractability of exhaustive interaction searches. This combination of speed and comprehensiveness is a practical advantage for multi-omics research.
Practically, risk-adapted screening is a compelling application. If individuals can be stratified by their combined genetic-demographic risk profile, those with high predicted risk could be directed toward more intensive screening (mammography at younger ages, MRI, more frequent intervals), while those with low predicted risk could be screened less frequently, improving both cancer detection rates and healthcare resource efficiency.
The primary limitation is the small dataset of 695 subjects, which is modest for training models with over 125,000 SNPs. Despite extensive regularization strategies including stratified cross-validation, class-weighted optimization, and performance evaluation on held-out test sets, the risk of overfitting in such a high-dimensional low-sample-size problem cannot be eliminated. The dataset imbalance (1.78:1 cases to controls) was addressed at multiple levels but the control group remains underpowered relative to the case group.
The dataset reflects a single Finnish population, which limits generalizability to other ethnic groups. Genetic architecture, allele frequencies, and the interactions between SNPs and demographic factors may differ substantially across populations. In particular, findings related to oral contraceptive use reflect patterns in an older Finnish cohort where formulations and usage patterns differ from contemporary populations worldwide.
The authors plan to extend the dataset using the Biobank of Eastern Finland, which will provide a larger cohort with additional controls and potentially higher statistical power to validate findings and identify new interacting feature networks. A larger dataset will also enable subgroup analyses by molecular subtype or menopausal status that the current sample size cannot support.
The study's most important future implication is its potential contribution to risk-adapted screening programs. A validated model combining genetic variants with accessible demographic risk factors could be integrated into clinical workflows to identify women who would benefit most from intensified surveillance, channeling limited healthcare resources toward those at highest risk while reducing unnecessary procedures in low-risk individuals.