Younger breast cancer survival models borrow from older cohorts: LSAT-EN tunes borrowing per omics layer and reaches 0.819 four-year AUC
Synopsis
Younger breast cancer patients have too few samples to train reliable prognostic models, so LSAT-EN treats older patients as a source cohort and sets the strength of borrowing toward the source coefficients separately for each of six omics layers, while elastic-net regularization selects variables. In TCGA-BRCA, with patients aged 60 and older as source and those under 50 as target, four-year binary survival prediction reached an AUC of 0.819, above 0.774 for pooled fitting, 0.757 for target-only fitting, 0.728 for glmtrans and 0.780 for GAS-EN. In simulation the advantage grew as source-target differences and target-sample scarcity increased.
Interpretation
Survival prediction for younger breast cancer patients now draws information from an older patient cohort, and how much it draws is decided separately for each omics layer. Earlier practice offered two choices: pool patients of all ages, which brings more data but may obscure age-specific associations, or fit only younger patients, which leaves few observations and outcomes. In the TCGA-BRCA four-year binary model with source age 60 and older and target under 50, LSAT-EN reached an AUC of 0.819, against 0.774 for pooled elastic net, 0.757 for target-only fitting, 0.728 for glmtrans and 0.780 for GAS-EN.
One tunable parameter per layer controls borrowing strength, with source-model coefficients held as a fixed anchor while the target model is fitted near it. Conventional source-anchored transfer uses a single borrowing strength for all features, whereas LSAT-EN gives each omics layer its own parameter, tuned together with the elastic-net penalty. Parameters are chosen by five-fold cross-validation on the target training set, the model is refitted on all target training data, and AUC is evaluated on an untouched 30% target test set and averaged over 10 repeats.
The gain from layer-specific borrowing is largest when source and target associations differ and target samples are few. When source and target coefficients were identical, methods performed similarly, with glmtrans achieving the highest AUC and LSAT-EN comparable. Under the highest simulated source-target heterogeneity, LSAT-EN ranged from 0.689 to 0.789 AUC; for the comparators in that same setting the abstract and the Results detail disagree, the abstract giving 0.590 to 0.771 while the Results list pooled elastic net at 0.561 to 0.592, target-only at 0.629 to 0.747, glmtrans at 0.590 to 0.721 and GAS-EN at 0.625 to 0.747. With no coefficient shift, LSAT-EN ranged from 0.943 to 0.957 and pooled elastic net from 0.954 to 0.961.
Perspective
The analysis stays within TCGA-BRCA and uses overall survival as the endpoint; the binary target cohort contains 64 to 105 patients and 3 to 7 deaths per horizon, and the Cox analysis contains 150 patients and 16 deaths. The authors call for external validation in prospective cohorts or cancer registries and for evaluation on additional endpoints such as progression-free interval and time to recurrence.
The authors note that the model uses molecular features only and omits serum markers such as CA15-3 and CEA as well as clinicopathological variables including nodal status, tumor size and HER2 expression. The predicted miRNA-mRNA pairs are hypotheses; the cited studies do not validate these exact interactions or their age specificity, and integration alone does not establish regulatory relationships.