Learning Framework Designed for Early Prediction of Breast Cancer Metastasis Consuming Genomic, Transcriptomic, and Epigenomic Profiles

Breast cancer is the most identified cancer and the most common cause of cancer related mortality among women all over the world, with about 2.3 million cases being diagnosed every year with 685,000 deaths occurring annually [1]. Despite the recent improvements in early cancer detection methods, hormonal therapies, and molecular-specific agents, metastasis i.e., the spread of the tumor cells at the original site to other body parts have continued to claim over 90 percent of death cases related to breast-cancers [2, 3]. Early and precise forecasting of metastatic potential is therefore critical in the optimization of treatment, prognosis and mortality reduction. Nonetheless, available clinical staging and pathological models, including TNM classification and receptor profile (ER, PR, HER2) include minimal information on the molecular factors of metastatic development [4]. With the advent of multi-omics technologies, including genomics, transcriptomics and epigenomics, oncology research has undergone a paradigm shift due to the ability to study cancer biology on a multilayered level [5]. Genomic profiles are quantitative records of somatic mutations, copy-number variation, and chromosomal rearrangements that contribute to tumor formation [6]; transcriptomic data measures aberrant gene-expression patterns that mediate proliferation and invasion [7]; and epigenomic signatures, especially DNA methylation and histone changes are quantitative records of heritable but reversible regulatory changes that regulate gene activity without changing the DNA sequence [8]. By combining such heterogeneous data modalities, we can build more whole-tumor heterogeneous landscapes that are more likely to capture tumor heterogeneity and evolution than single-omics methods [9]. These complementary sources of information can be effectively integrated into a multi-omics learning framework, which will enhance the sensitivity and specificity of the prediction models of metastasis. The schematic idea of such a system is shown in Figure 1, whereby, beforehand, multi-layer biological data, which are genomics, transcriptomics and epigenomics, are pre-processed, then encoded to obtain latent representation, which is fused via advanced learning architectures to provide an early prediction of metastatic potential. This integrative approach enables the discovery of critical biomarkers and pathways involved in metastatic spread as well as supporting the interpretation of models to the clinicians. In general terms, a multi-omics learning system to forecast early metastasis, which is the focus of this study, consists of the following components:<|human|>Generally speaking, a multi-omics learning system to predict early metastasis, as is the case with this study, is made up of the following elements:

Figure 1. Construction of a multi-omics learning context for breast cancer metastasis prediction

The last few years have seen the growing usage of machine-learning and deep-learning algorithms in cancer prognosis. Common techniques like Support Vector Machines (SVMs) [10], Random Forests [11], and logistic regression [12] have been used to classify metastatic and non-metastatic samples using gene-expression data, but due to the large dimensionality and nonlinearity of omics features, their performance is limited. Deep-learning methods, such as autoencoders [13], convolutional neural networks (CNNs) [14], and graph neural networks (GNNs) [15] have been shown to be better at learning nonlinear interactions and extract biologically meaningful representations of multi-omics data. Despite these developments, there are three key challenges that exist:

1. Partial multi-omic integration-most of the studies are based on only one type of omics, e.g., transcriptomics or methylation, and hence fail to capture inter-omic interactions [16].
2. Weak interpretability-deep architectures can be viewed as black boxes, which cannot be understood biologically or be trusted clinically [17].
3. Lack of cross-cohort generalization models that are trained on a single dataset (e.g., TCGA-BRCA) often will not work with other datasets (e.g., METABRIC or GEO) because of platform bias and batch effects [18].

These gaps demonstrate the necessity to have a single, interpretable, and generalizable learning model that maximizes the complementary relationship between multi-omics data to improve the early detection of breast-cancer metastasis.

To address these constraints, the proposed learning framework in this study, which is entitled Learning Framework of Multi-Omics Metastasis Prediction (LF-MMP), is a predictive machine built through the integration of genomic, transcriptomic and epigenomic profiles in a single end-to-end architecture. This piece of work has the goals of:

• Create a multi-modal deep learning model that is integrated to represent high-order correlation across omics layers in predicting early metastasis.
• Use feature-attribution and attention models (e.g., SHAP and layer-wise relevance propagation) to obtain biologically interpretable predictions and discover biomarkers of metastasis.
• Test the framework on several benchmark datasets (TCGA-BRCA, METABRIC, GEO) to determine reproducibility, scalability, and resistance to cohort variability.

The main findings of this paper are summarized as the following:

• Unified Multi-Omics Fusion: Presentation of a profound hybrid fusion design with genomic, transcriptomic, and epigenomic latent representations to boost the accuracy of metastasis prediction.
• Interpretability and Biomarker Discovery: SHAP-based feature interpretation that allows the discovery of biologically important genes and methylation sites associated with metastatic pathways.
• Cross-Dataset Validation: Overall validation on three large-scale cohorts in terms of better generalization (AUC > 0.94) over state-of-the-art models [19-21].
• Clinical/Translational Impact: Delivery of a decision-support model potentially useful in helping oncologists to risk-stratify patients, plan individualized therapies, and minimize unnecessary systemic therapies.

In addition to the innovativeness in computation, the suggested framework has the potential to provide clinical advantages in the form of the early detection of the high-risk patients, prior to the overt progression to metastasis, which allows acting proactively and positively influencing the survival rates. Moreover, the fact that the model can be biologically interpreted facilitates the generation of hypothesis to be tested downstream in vitro and in vivo, allows a pathway between computational oncology and translational medicine.

The rest of this paper will have the following structure. Section 2 is a full review of the latest developments and current shortcomings of breast cancer metastasis prediction, with the focus on the comparative analysis of single-omics and the multi-omics methods of analysis. Section 3 describes the datasets used in this paper, such as data sources, data preprocessing pipelines, normalization steps, and the feature-engineering pipeline on genomic, transcriptomic, and epigenomic profiles. Section 4 presents the proposed Learning Framework of Multi-Omics Metastasis Prediction (LF-MMP) and describes the overall architectural design, mathematical formulations employed, training algorithm and the interpretability mechanisms used to generate biologically meaningful information. Section 5 summarizes the planning of the experiment, assessment of outcomes, and comparative studies performed to confirm the functionality of the proposed model, and then also, elaborates a biological explanation of the biomarkers and pathway enrichments identified. Lastly, Section 6 presents the conclusion of the paper summarizing the significant results, its clinical implications, and relevance to precision oncology, existing limitations, and future research directions.

In this section, the design and implementation of the proposed Learning Framework (LF-MMP) that incorporates the use of genomic, transcriptomic, and epigenomic data in the early prediction of breast cancer metastasis is described. It is based on five primary steps, which are (1) data collection and preprocessing, (2) feature extraction and dimensionality reduction, (3) multimodal fusion using deep neural representation learning, (4) classification and optimization, and (5) interpretability analysis. Figure 2 shows the general flow of proposed LF-MMP, where multi-omics inputs are combined through the feature encoding modules, the attention-based fusion layer, and the metastasis classification output, resulting in the final one. The LF-MMP model was tested on three benchmark datasets: TCGA-BRCA, METABRIC and GEO (GSE96058). These datasets consist of comprehensive and multi-omics and clinical data of thousands of breast cancer patients, which is perfect as it can be used to predict metastasis. Table 2 gives a summary of the datasets.

Each dataset has the metastasis status as a binary variable (1 = metastatic, 0 = non-metastatic). To accomplish cross-dataset generalization experiments, training was done using TCGA, validation using METABRIC and external testing using GEO as shown in Figure 3.

The treatment of missing values was mode sensitive. In the case of genomic mutation data, the missing cases were treated as lack of a mutation and coded as 0. In transcriptomic and epigenomic data, features that have over 20 percent missing data points in all samples would be eliminated. On the other characteristics that had intermittent missing data (below 20%), we used the k-Nearest Neighbors (KNN) imputation (k = 5) applied to training data individually to avoid data leakage. The imputer fitted was applied to the validation set and the test set.

After imputation a step feature selection was carried out to deal with high dimensionality and to control noise. The first step was to remove features whose variance was almost zero, i.e., the ratio of the frequency of the most frequent value to the second most frequent value is at least 19:1, the fraction of distinct values is less than 10%. Second, we used a univariate statistical filter applied on the two-sample t-test (Eq. (4)) to obtain characteristics that significantly differ in their expression/abundance between the metastatic and the non-metastatic populations. To mitigate against false discoveries, we selected the 5,000 most significant features of each omics modality according to a composite measure of absolute log2 fold-change and false discovery rate (FDR) adjusted p-value less than 0.05. This strict procedure allowed passing only the most biologically significant and statistically strong features to autoencoders in order to reduce dimensions.

Figure 2. Flowchart of the proposed study

Table 2. Characteristics of data to be used in this study

Figure 3. The diagram of the proposed study

All omics datasets were preprocessed by modality-specific methods to make the data consistent and comparable:

Genomic features: Data of somatic mutation and copy-number variations (CNVs) were coded as binary and continuous matrices.

Transcriptomic data: The RNA-seq expression data was normalized with the use of log-transformed values of the FPKM as:

$x^{\prime}=\log _2(x+1)$      (1)

where, x is the raw FPKM expression count.

Epigenomic data: DNA methylation intensity values were transformed into $\beta$-values using:

$\beta_i=\frac{M_i}{M_i+U_i}$      (2)

where, $M_i$ and $U_i$ represent methylated and unmethylated probe intensities, respectively [1].

Batch effects across platforms were corrected using the ComBat algorithm [2], while z-score normalization ensured zero-mean and unit variance:

$z_{i j}=\frac{x_{i j}-\mu_j}{\sigma_j}$      (3)

where, $x_{i j}$ is the feature $j$ of sample $i, \mu_j$ and $\sigma_j$ are the mean and standard deviation of feature $j$.

Stratified sampling was used to divide each dataset into 70% training (validation) and 15% testing subsets (to maintain balance between classes). To determine the generalization performance of the model, the five-fold cross-validation has been used. Omics data with high dimensions (> 60,000 features) are extremely challenging to compute and overfitting. To reduce this, statistical filtering was used together with dimensionality reduction by using autoencoders.

Attributes whose variance is close to zero or those whose inter-correlations are too high (|human|>Attributes with near-zero variance or high inter-correlations (|human|) were eliminated. Selection was done using:

$t_j=\frac{\bar{x}_j^{(1)}-\bar{x}_j^{(0)}}{s_p \sqrt{\frac{1}{n_1}+\frac{1}{n_0}}}$      (4)

where, $\bar{x}_j^{(1)}$ and $\bar{x}_j^{(0)}$ denote the mean feature values for metastatic and non-metastatic samples, $s_p$ is the pooled standard deviation, and $n_1, n_0$ are sample counts per group.

For each omic type $o \in\{g, t, e\}$ (genomic, transcriptomic, epigenomic), a deep autoencoder compresses high-dimensional data into latent representations:

$h_o=f_o\left(X_o\right)=\sigma\left(W_o X_o+b_o\right)$     (5)

$\hat{X}_o=\sigma\left(W_o^{\prime} h_o+b_o^{\prime}\right)$     (6)

where, $W_o$ and $W_o^{\prime}$ are encoder and decoder weights, $b_o, b_o^{\prime}$ are biases, and $\sigma(\cdot)$ denotes the ReLU activation. The autoencoder minimizes reconstruction loss:

$\mathcal{L}_{\mathrm{rec}}=\left\|X_o-\hat{X}_o\right\|_2^2$      (7)

The learned latent vectors $h_o$ serve as compact, noise-robust omic representations for the fusion network.

After encoding, the latent representations $h_g, h_t$, and $h_e$ are concatenated and processed through an attention-weighted fusion layer that adaptively assigns importance to each omic modality. The fusion representation $H_f$ is computed as:

After encoding, the latent representations $h_g, h_t$, and $h_e$ are concatenated and processed through an attention-weighted fusion layer that adaptively assigns importance to each omic modality. The fusion representation $H_f$ is computed as:

$H_f=\sum_{o \in\{g, t, e\}} \alpha_o h_o$      (8)

where, attention weights $\alpha_o$ are obtained by the softmax function:

$\alpha_o=\frac{\exp \left(\beta_o\right)}{\sum_k \exp \left(\beta_k\right)}$      (9)

and $\beta_o$ are learnable parameters capturing the relative significance of each modality. This enables the model to dynamically emphasize omic features, which is most informative for metastasis prediction.

The final classification layer receives the fused representation $H_f$ and outputs a metastasis probability:

$\hat{y}=\sigma\left(W_c H_f+b_c\right)$      (10)

where, $W_c$ and $b_c$ are the weights and bias of the classifier, and $\sigma(\cdot)$ denotes the sigmoid activation function. The model is trained using the binary cross-entropy (BCE) loss function:

$\mathcal{L}_{B C E}=-\frac{1}{N} \sum_{i=1}^N\left[y_i \log \left(\hat{y}_i\right)+\left(1-y_i\right) \log \left(1-\hat{y}_i\right)\right]$      (11)

where, $y_i$ and $\hat{y}_i$ are the true and predicted labels for the $i^{\text {th }}$ sample. L2 regularization was applied to mitigate overfitting:

$\mathcal{L}_{\text {reg }}=\lambda\left\|W_c\right\|_2^2$      (12)

The total objective function is therefore:

$\mathcal{L}_{\text {total}}=\mathcal{L}_{B C E}+\mathcal{L}_{\text {rec}}+\mathcal{L}_{\text {reg}}$      (13)

where, $\lambda$ is the regularization coefficient empirically set to 0.001.

Model parameters were optimized using the Adam optimizer [3] with a learning rate of 0.0005. Early stopping was triggered when validation loss did not improve for 20 epochs. The main hyperparameters and computational settings are provided in Table 3.

Table 3. Training and simulation settings

Model performance was assessed using Accuracy, Precision, Recall, F1-score, and Area Under the ROC Curve (AUC), defined as follows:

Accuracy $=\frac{T P+T N}{T P+T N+F P+F N}$     (14)

Precision $=\frac{T P}{T P+F P}$      (15)

Recall $=\frac{T P}{T P+F N}$     (16)

F1-Score $=2 \times \frac{\text {Precision × Recall}}{\text {Precision}+ \text {Recall}}$      (17)

$\mathrm{AUC}=\int_0^1 T P R(F P R) d(F P R)$      (18)

where, $T P, T N, F P$, and $F N$ denote true positives, true negatives, false positives, and false negatives respectively. These metrics collectively quantify the model's predictive capability, sensitivity to metastasis cases, and overall discriminative power.

The LF-MMP workflow is summarized in Algorithm 1 to provide the pseudocode of the computational steps that play the major part in training, feature fusion, and optimization.

To increase the biological interpretability, SHapley Additive exPlanations (SHAP) values were calculated on each omic feature to determine its effect on the predicted metastasis score [4]. KEGG and Gene Ontology databases were used to map the most influential genes and methylation sites onto biological pathways, and metastasis-related biological pathways, including PI3K-AKT, Wnt, and TGF-B signaling were found to be associated with them.

All the simulations were performed on high-performance computing environment with Ubuntu 22.04, TensorFlow 2.13 and CUDA 12.2. The average training time per fold was 180 seconds, and convergence normally took 120 epochs. All experiments were repeated 5 times to achieve reproducibility, and all the metrics were reported in terms of mean and standard deviation as seen in Tables 4, 5, and 6.

Table 4. Simulation environment setting

Table 5. Algorithmic hyperparameters

Table 6. Evaluation protocol

• Time: dominated by Algorithms 2-3; approximately $O\left(E \cdot n \cdot\left(k_g+k_t+k_e\right)\right)$.
• Memory: stores latent embeddings per omic-size $O\left(n \cdot\left(k_g+k_t+k_e\right)\right)$.
• Seeds & Determinism: fix PRNG seeds; log package versions; persist scaler/ComBat/threshold/calibration artifacts.
• No-leakage guarantee: all normalizers, imputers, ComBat, and calibration are fit on train and applied to val/test.

Inference-Time Procedure (Deployment)

Input: New patient $x_g, x_t, x_e$ (possibly missing a modality).

Steps:

1. Apply training scalers/ComBat/imputers to each available omic.
2. Compute $h_o=f_o\left(x_o\right)$ for available modalities; if one is missing, set $\alpha_o=0$ and renormalize remaining $\alpha$.
3. Compute $H_f$ and $\hat{y}$; apply calibration and threshold $\tau^*$.
4. Provide SHAP-based explanation at feature and modality levels.

Output: Predicted metastasis risk, calibrated; interpretable attributions.

Overall, this paper proposed LF-MMP, a single learning framework that combines genomic, transcriptomic, and epigenomic cues to predict breast-cancer metastasis early with uniform improvements on discrimination, reliability and interpretability relative to powerful single- and multi-omic controls. LF-MMP has AUCs of 0.956 (TCGA-BRCA), 0.946 (METABRIC), and 0.938 (GEO) in three cohorts and strong cross-cohort performance in TCGA-BRCA-trained and METABRIC-tested (AUC = 0.942), and GEO-tested (AUC = 0.935). Probabilistic results were well calibrated (Brier = 0.085 -0.098; ECE = 0.021-0.028), and the computational overhead was practical to deploy (about 18.4M parameters, about 3 minutes per-fold on one A100, about 38 ms per-case inference). SHAP-based analyses identified biologically relevant markers -e.g., TP53 PIK3CA BRCA2 CDH1 (genomic), ESR1/GATA3/TWIST1/MKI67 (expression), and methylation at PTEN/TWIST1), and the value of modality-level attention confirmed the complementary value of methylation when combined with expression. Despite these strengths, the work has limitations: it uses only retrospective public cohorts and may be susceptible to batch effects and label noise; performance may change under unknown clinical protocols or ancestries; metastasis labels approximate early risk over time-to-event; and in spite of our efforts to reduce oscillations on ambiguity with attention and SHAP, causal interpretability and mechanistic validation is not complete. Future work then will focus on prospective, multi-center assessment with standardized wet-lab protocols; integration of histopathology, radiomics as further modalities; domain adaptation and federated learning to accommodate site-specific changes and data-sharing limitations; generative imputation to missing modalities and semi-supervised learning to utilize unlabeled samples; pathway- and cell-state-aware prior to promote biological faithfulness; longitudinal modeling to dynamically risky; decision-curve, cost-sensitive analysis of clinical thresholds; fairness audits across subgroups; and real All of these instructions put LF-MMP in the line of a clinically actionable, transparent, and generalizable early-metastasis decision-support instrument.