Predicting outcomes in breast cancer using genomic data is challenged by demographic imbalances in available cohorts. Underrepresented populations often have smaller sample sizes in public multi-omics datasets, which can degrade predictive model performance and limit equitable clinical translation. The authors frame their work around the need for methods that both leverage multi-omics information and account for data imbalance to improve outcome prediction for minority groups.
The authors introduce DAG-HEART, a directed acyclic graph–guided multi-omics transfer learning framework intended to provide health equity–aware representation learning for breast cancer. DAG-HEART extends a previously reported transfer-learning strategy by incorporating data augmentation and by explicitly modeling directional relationships among molecular features using a directed acyclic graph (DAG) constraint during representation learning.
DAG-HEART was developed and evaluated using TCGA-BRCA multi-omics datasets comprising mRNA, miRNA, and DNA-methylation profiles. The source summary reports these modalities as the inputs used for model training and representation extraction. The summary does not include detailed cohort sizes, preprocessing choices, or feature-selection thresholds; those specifics are contained in the full preprint PDF and are not reported in the source abstract.
The primary prediction task reported was progression-free interval (PFI) for a data-minority group. The authors explicitly designed the evaluation to probe performance in a demographically underrepresented subgroup, aligning the work with an equity-focused objective. The abstract does not provide the numerical definition of the minority group, subgroup sample counts, or the precise PFI endpoint construction; readers should consult the full manuscript for those analytic details.
According to the authors' reported results, DAG-guided nonlinear integration consistently improved predictive performance compared with direction-agnostic and correlation-based representations. This comparison indicates that incorporating directed structure into multi-omics integration yielded better outcomes in the reported experiments. The abstract does not supply specific performance metrics, confidence intervals, or statistical testing details in the summary; those results are available in the full preprint.
The framework tested models with biologically motivated directional constraints, reversed directional structures, and unconstrained directed structures. The authors report that biologically motivated directional constraints generally outperformed reversed or unconstrained structures. This finding supports the premise that imposing plausible causal or regulatory directionality during representation learning can be beneficial for downstream prediction in the authors' experiments.
Feature selection within the DAG-HEART workflow produced recurrently selected features that the authors report converged on extracellular-matrix and regulatory pathways. These pathway-level signals were reported to support clinically meaningful risk stratification in their analyses. The abstract does not list individual genes, miRNAs, CpG sites, or the pathway enrichment statistics; those specifics are reported in the full preprint.
A stated advantage of DAG-HEART is interpretability: by combining a directed multi-omics representation with transfer-learning and data augmentation, the approach aims to yield representations that are both predictive and biologically interpretable, especially under conditions of data imbalance across racial groups. The authors emphasize that the directed graph structure facilitates biologically meaningful representation and selection of features tied to known pathways.
The source article is an abstract and summary on bioRxiv; it does not include full methodological detail, code availability, exact quantitative results, model hyperparameters, or subgroup sample sizes in the provided summary. The work is presented as a preprint and has not been peer reviewed. For reproducibility, external validation, and assessment of clinical utility, consult the full preprint PDF and any accompanying supplementary materials.
The preprint lists Minjeong Baek, Jieqiong Wang, and Shibiao Wan (University of Nebraska Medical Center) as authors, with Shibiao Wan provided as the corresponding contact. The authors declare no competing interests. Reported funders include the U.S. National Science Foundation, National Cancer Institute (NCI), National Institute on Alcohol Abuse and Alcoholism (NIAAA), the NIH Office of the Director, the Pediatric Cancer Research Group, and the Nebraska Research Initiative. The manuscript is posted on bioRxiv as a preprint and is not peer reviewed in the source record.
To assess implementation details, reproduce experiments, or apply DAG-HEART to other cohorts, readers should download and review the full preprint, which contains methodological specifics and result tables not included in the abstract. The summary reported here captures the high-level aims and conclusions but omits the quantitative performance metrics and exact algorithmic specifications.