Ricky Nguyen, Fatemeh Vafaee
Abstract Multi-omics integration has become central to precision medicine, yet in real-world clinical cohorts, complete multi-layer profiling is rarely achieved. Cost, assay failure, and evolving study design frequently produce block-wise modality missingness, where entire omics layers are absent for subsets of patients. This setting poses a distinct challenge for predictive modelling and biomarker discovery, as conventional integrative methods often assume fully paired data or rely on complete-case filtering or point-wise imputation, leading to information loss or biased inference. In response, a growing class of methods has been developed to learn from partially observed multi-omics data through missingness-aware fusion, shared latent representations with subset-conditioned inference, and modality-completion frameworks. This review presents a methodological taxonomy of approaches designed to support outcome modelling under heterogeneous modality availability, with a primary focus on patient-level bulk data and clinical prediction tasks, while also discussing representation-learning frameworks developed for single-cell or unsupervised settings where the missing-modality handling mechanism is architecturally transferable to supervised clinical prediction. We contrast design philosophies, inference mechanisms, and robustness properties, and examine when strategies developed for single-cell mosaic integration transfer to cohort-level modelling. By clarifying the assumptions, robustness properties, and empirical behaviour of missing-modality strategies, we aim to provide a principled framework for selecting and developing models suited to partially observed multi-omics datasets in real-world clinical contexts.