<p>Precise diagnosis of complex diseases increasingly depends on the integration of multimodal data. However, the high dimensionality of such data makes it difficult for traditional modeling methods to efficiently identify biomarkers, while deep learning approaches, despite their strong predictive power, often lack interpretability. Here, we introduce BPX-Net, a generalizable deep learning framework that learns compact multimodal representations through a biomarker-preserving dropout mechanism, where dropout probabilities are modulated by feature importance scores self-learned from a built-in interpretability module and, when available, further informed by clinical priors. This design enables BPX-Net to make robust predictions during inference by selectively attending to a sparse set of disease-informative predictors, making it insensitive to missing values in less relevant or redundant variables. Across multi-center cohorts covering diverse clinical tasks (disease diagnosis, prognosis, treatment response prediction, and risk stratification), BPX-Net yields substantial performance gains, e.g. achieving an average AUC of 85.43% across four tasks, outperforming baselines by 4% to 20% in the presence of missing data. More importantly, it identifies predictors aligned with established clinical knowledge. Cross-hospital validation further confirms the robustness and clinical relevance of these predictors. Collectively, BPX-Net offers a clinically grounded deep learning framework for multimodal analysis, intrinsically robust to data incompleteness and equipped with built-in interpretability, thereby eliminating reliance on computationally intensive post-hoc tools such as SHAP.</p>