DIABLO

Unlike unsupervised methods like MOFA, DIABLO explicitly incorporates phenotypic outcome classes into the integration framework. It maximizes the covariance between latent components and the outcome of interest, ensuring that the extracted multi-omics signatures are directly relevant to the biological question—such as distinguishing responders from non-responders or disease subtypes.

DIABLO applies L1 (lasso) regularization to select a small, interpretable subset of molecular features from each omics block that are maximally correlated across modalities and discriminative of the outcome. This addresses the curse of dimensionality inherent in multi-omics data, where the number of features (genes, proteins, metabolites) vastly exceeds the number of samples. The selected features form a multi-omics biomarker panel.

DIABLO is built on the block sparse Partial Least Squares Discriminant Analysis (block sPLS-DA) framework. It extends sPLS-DA to multiple data blocks by computing latent components that are linear combinations of selected features from each block. These components are constrained to be highly correlated with each other and with the discriminant vectors separating outcome classes, enabling simultaneous dimension reduction and classification.

DIABLO provides a distinctive Circos plot to visualize multi-omics correlations. This circular diagram displays the selected features from each omics block and draws connecting lines between features with strong cross-block correlations. Positive correlations are shown in one color, negative in another, allowing researchers to quickly identify coordinated molecular modules—such as a gene whose expression correlates with a specific metabolite and protein level across samples.

DIABLO uses cross-validation to tune two critical hyperparameters: the number of latent components and the number of features to retain per block. Performance is evaluated using balanced error rates and AUC metrics across multiple distance measures (centroids.dist, mahalanobis.dist, max.dist). The framework also outputs classification error rates per class, enabling assessment of model performance for each outcome category independently.

DIABLO is implemented in the mixOmics R package, a comprehensive toolkit for multivariate analysis of biological data. It integrates seamlessly with upstream preprocessing functions and downstream visualization tools within the ecosystem. The framework supports arbitrary numbers of omics blocks (genomics, transcriptomics, proteomics, metabolomics) and handles both continuous and categorical phenotypic outcomes, making it a versatile choice for biomarker discovery pipelines.

The foundational statistical framework upon which DIABLO is built. SGCCA extends traditional Canonical Correlation Analysis to handle more than two data blocks simultaneously. It incorporates L1 regularization to produce sparse weight vectors, selecting only the most relevant molecular features from each omics modality. This sparsity constraint is critical for biological interpretability, preventing the model from assigning non-zero weights to thousands of noisy variables and instead focusing on a compact, highly correlated signature.

A non-linear integration method that constructs patient similarity networks for each omics data type and iteratively fuses them into a single comprehensive network. SNF excels at capturing complementary information where relationships are strong in one data type but weak in another. The fused network can be used for spectral clustering to identify cancer subtypes. Unlike DIABLO's linear latent variable approach, SNF operates on the principle of message passing across networks, making it robust to heterogeneous data distributions and noise.

A deep learning alternative to DIABLO's linear framework. Cross-modal attention allows a neural network to dynamically weigh the relevance of features from one omics modality when processing another. For example, a transformer can learn that a specific somatic mutation in genomics should direct attention to a particular gene expression change in transcriptomics. This captures complex, non-linear inter-modal molecular relationships that linear correlation-based methods like DIABLO may miss, though at the cost of reduced interpretability.

The specific penalty function used within DIABLO's feature selection engine. Elastic net combines both L1 (Lasso) and L2 (Ridge) regularization. The L1 penalty drives many feature weights to exactly zero, performing automatic variable selection. The L2 penalty encourages a grouping effect, ensuring that highly correlated features—such as co-expressed genes in a pathway—are selected together rather than arbitrarily picking one. This dual penalty is essential for handling the strong correlation structures inherent in omics data.

A critical preprocessing step for any supervised integration framework like DIABLO. Unmeasured confounders—such as batch effects, age, or sex—can induce spurious correlations between molecular features and the outcome of interest. DIABLO's discriminative objective can inadvertently amplify these technical artifacts rather than true biological signals. Rigorous confounder adjustment using methods like ComBat for batch effects or residualization against known covariates is essential before training to ensure the selected latent components are driven by biology, not experimental design.