Precision oncology increasingly depends on integrating heterogeneous evidence across molecular profiling, medical imaging, and clinical records, yet robust deployment is limited by data fragmentation across hospitals, missing modalities, batch effects, privacy constraints, and weak mechanistic interpretability. We propose a causal and explainable federated multimodal learning framework for cancer prediction and target discovery that fuses multi-omics, radiology or digital pathology imaging, longitudinal EHR features, and CRISPR dependency signals. The system trains across sites without centralizing raw data using federated optimization with secure aggregation and optional differential privacy, and is designed to remain reliable under non-IID site heterogeneity and structured missingness. To move beyond correlational risk scoring, we introduce a causal layer that encodes structural assumptions for treatment response and survival, supports counterfactual prediction, and applies invariant learning style regularization to improve transportability. For clinical safety, the framework outputs calibrated uncertainty and multi-level explanations, including modality contribution reporting, feature attributions over genes, imaging regions, and EHR variables, and causal what-if narratives for treatment changes and gene perturbations. We define a fully public experimental protocol using TCGA and CPTAC for multi-omics and outcomes, TCIA for imaging domain shift evaluation, and DepMap for CRISPR based dependency mapping and pathway level target rationale. This work provides an end-to-end, reproducible blueprint for privacy-preserving, mechanism-aware cancer AI, enabling benchmark driven validation prior to prospective multi-hospital deployment.