Profile-guided Hybrid Approach for block-wise missing data handling in multi-omics: a breast cancer case study

Abstract

Block-wise missingness is a common challenge in multi-omics data, hindering the development of robust and generalizable machine learning models, as real-world cohorts rarely contain complete omic profiles. Many current methods either discard incomplete samples, use available-case models that need retraining when faced with new missingness patterns, or depend on full-dataset imputation, which can risk biological integrity and model stability. Using a complete four-omics breast cancer dataset (705 patients, 1,937 features), up to 60% block-wise missingness was simulated across five clinically relevant scenarios and used to compare four strategies for handling missing data: an Imputation-Based model, Dynamic and Exhaustive Available-Case approaches, and the proposed Hybrid Approach that combines profile-guided modeling with selective, test-time imputation. Performance was evaluated using accuracy, F1 score, balanced accuracy, inference time, and variability across 15 random seeds, with significance assessed using the Wilcoxon signed-rank test. The Hybrid Approach consistently achieved the strongest and most stable performance. Relative to the complete-data baseline, it reached an average accuracy of 103.7%, F1 score of 123.3%, and balanced accuracy of 104.8%, outperforming the Imputation-Based method and matching or exceeding both Dynamic and Exhaustive Available-Case strategies. Statistical testing confirmed that these improvements were significant. The method also demonstrated fast and predictable inference (~ 2 s) and an average total runtime of ~ 49 s per configuration—nearly three times faster than the Exhaustive approach (~ 124 s)—while maintaining high reproducibility and low variance across seeds, a key indicator of computational stability. By selectively combining lightweight imputation with profile-specific modeling, the Hybrid Approach provides a computationally efficient and statistically robust solution for block-wise missing data. This framework offers a generalizable strategy for multi-omics data mining, and lays the foundation for future systems incorporating cross-profile learning and advanced imputation.