Layer-Specific Source-Anchored Transfer Learning for Multi-Omics Survival Prediction in Younger Breast Cancer Patients

Abstract

Background: Younger breast cancer patients (=60, >=55, and >=50 years were used as the source to improve prediction in patients aged <50 years. LSAT-EN were evaluated in group-correlated simulations and in TCGA-BRCA integrating 6 omics layers. Binary survival at 2-4 years was assessed by AUC; censored survival by Harrell's C-index and 2-5-year inverse probability of censoring weighted (IPCW) AUC. Results: In simulation, LSAT-EN achieved the highest AUC under moderate source-target heterogeneity (sigma^2=0.25; 0.747- 0.802), outperforming GAS-EN (0.735-0.767), overall EN (0.683-0.710), individual EN (0.646-0.767), and glmtrans EN (0.640-0.733) across target training sample sizes (nT). Its advantage was greatest under higher source-target heterogeneity (sigma^2=4), particularly with small nT, where LSAT-EN achieved AUCs of 0.689- 0.789 compared with 0.590-0.771 across competing methods. In TCGA-BRCA data, LSAT-EN improved discrimination relative to overall and individual EN in all binary and Cox models across all time horizons and source age groups. The method also identified clinically/ biologically important mRNA and miRNA, as well as useful candidate miRNA-mRNA pairs, indicating powerful biomarker discovery capability. Conclusion: LSAT-EN is a powerful TL method for improving prediction for younger breast cancer patients; The framework may support prediction in other underrepresented populations with limited target-cohort data to advance cancer disparity ML research.