The breast cancer care would require the tools that will help to identify the patients who may develop metastasis at an early stage, when the treatment decision could be altered. Models that use single types of data (as in the case of using a single transcription factor only) can tend to overlook significant information and may not work well when applied to different hospitals. Our framework, LF-MMP, is a learning framework that integrates three types of molecular data, namely genomics (DNA changes), transcriptomics (gene activity), and epigenomics (DNA methylation) to give an early patient-level risk score of metastases. The framework normalizes and cleans every dataset, trains a compact representation of each omics layer, and lastly combines them together with an attention mechanism that allows the model to pay attention to the most informative signals. An optimized classifier transforms the fused representation into well-behaved probabilities that may be used to support clinical thresholds. We tested LF-MMP on three external populations, namely, TCGA-BRCA, METABRIC and GEO (GSE96058). The model performed better than powerful single-omic and deep multi-omic controls, and AUCs were 0.956 (TCGA-BRCA), 0.946 (METABRIC), and 0.938 (GEO). Performance was also high when trained on TCGA-BRCA and externally tested (AUC 0.942 on METABRIC; 0.935 on GEO). There was good calibration of the expected risks (Brier 0.085-0.098; ECE 0.021-0.028). The descriptions of the features showed familiar biology (such as TP53 and PIK3CA mutations, ESR1 and GATA3 expression, and PTEN/TWIST1 methylation). Inference and training were sufficiently quick to be used on regular GPU. The limitations of this study are as follows: the research is based on retrospective publicly available data, labels are not directly related to time-to-event but to early risk, and new environments may differ in terms of performance. Future directions will incorporate prospective, multi-centric validation; imaging and radiomics; enhancement to site differences and missing data; tracking of model calibration in real-life use.