MultiOmics VAE: MOVE for Biomarker Discovery

• MOVE is a multi-omics VAE/CVAE model that compresses heterogeneous omics data into a lower-dimensional latent space for biomarker discovery.
• It operates as part of an ensemble pipeline, using graph-derived features from GAT and engaging Elastic-net regression to manage high-dimensional low-sample-size data.
• The model employs a VAE-style objective with modality-specific reconstruction, KL divergence, and cross-modal alignment despite some incomplete architectural details.

Searching arXiv for MOVE and closely related multi-omics VAE papers. MultiOmics Variational AutoEncoder (MOVE) is a multi-omics latent embedding module described as part of an ensemble biomarker-discovery pipeline for dementia under high-dimensional low-sample-size conditions. In the paper that names it explicitly, MOVE is positioned between a Graph Attention Network (GAT) stage and downstream Elastic-net regression plus Storey’s False Discovery Rate (FDR), and is intended to compress heterogeneous omics signals into a lower-dimensional latent representation while preserving biologically meaningful structure (Lee et al., 4 Sep 2025). The available specification is partial rather than complete: the manuscript provides a VAE-style objective, refers to MOVE both as “Multi-Omics Variational AutoEncoder” and as a “Conditional Variational Autoencoder (CVAE) for multi-omics data,” and attributes to it cross-modal feature compression, reconstruction fidelity, and cross-modal alignment, but it does not provide full architectural detail, a complete mathematical derivation, or a full training configuration (Lee et al., 4 Sep 2025).

1. Nomenclature and scope

The term MOVE appears explicitly in “An Interpretable Ensemble Framework for Multi-Omics Dementia Biomarker Discovery Under HDLSS Conditions” (Lee et al., 4 Sep 2025). In that work, MOVE is one of four components in the section “Multi-Omics Biomarker Discovery via GAT-MOVE-ElasticNet-FDR Framework,” alongside GAT, Elastic-net regression, and Storey’s FDR (Lee et al., 4 Sep 2025). Its stated role is to encode graph-derived multi-omics representations into a lower-dimensional latent space suitable for downstream biomarker prioritization.

The paper uses partially inconsistent terminology. MOVE is expanded as “Multi-Omics Variational AutoEncoder,” while the introduction refers to a “Conditional Variational Autoencoder (CVAE) for multi-omics data (MOVE),” and the framework section also describes “MOVE, a manifold optimization technique” (Lee et al., 4 Sep 2025). The manuscript does not formally resolve these labels. The safest evidence-bound characterization is therefore that MOVE is treated by the authors as a multi-omics VAE/CVAE-style latent embedding model (Lee et al., 4 Sep 2025).

This limited formalization distinguishes MOVE from better-specified neighboring architectures. For example, CustOmics is presented as a hierarchical mixed-integration system with “One autoencoder per source” feeding a “central variational autoencoder,” with Maximum Mean Discrepancy replacing the standard KL term (Benkirane et al., 2022). MODIS is described as “multiple coupled variational auto-encoders (VAEs), as many as the number of modalities,” aligned by an adversarial discriminator in a shared latent space (Lepe-Soltero et al., 24 Mar 2025). By contrast, MOVE’s published description is more concise and leaves key implementation details unstated (Lee et al., 4 Sep 2025).

2. Position in the ensemble biomarker-discovery framework

Within the GAT-MOVE-ElasticNet-FDR framework, the pipeline begins with GAT “to uncover latent gene–gene associations,” after which “graph-derived representations are then encoded into a lower-dimensional latent space using MOVE” (Lee et al., 4 Sep 2025). Elastic-net regression is then used for sparse and interpretable feature selection, and Storey’s FDR is used for statistical validation of selected biomarkers (Lee et al., 4 Sep 2025).

The conceptual division of labor in the framework is explicit. GAT “captures heterogeneous and hidden dependencies among genes and across omics.” MOVE performs “cross-modal feature compression” and preserves “intrinsic geometry/manifold structure while mitigating the curse of dimensionality.” Elastic-net translates the integrated representation into sparse variable selection, and Storey’s FDR imposes statistical reliability on the final feature ranking (Lee et al., 4 Sep 2025). The paper presents this combination as especially suited to “high-dimensional, low-sample-size, sparse” multi-omics data (Lee et al., 4 Sep 2025).

The omics setting is four-layer and is consistent across simulated and ADNI analyses: genomics, transcriptomics, proteomics, and metabolomics (Lee et al., 4 Sep 2025). In the simulation section, the synthetic dataset contains 500 genes, 300 mRNAs, 200 proteins, and 100 metabolites (Lee et al., 4 Sep 2025). For ADNI, the modalities are SNP arrays and APOE genotyping, RNA-seq from peripheral blood, CSF protein levels, and plasma metabolite concentrations (Lee et al., 4 Sep 2025).

This placement makes MOVE an intermediate integration module rather than a terminal predictor. The paper does not assign feature selection to MOVE itself; feature selection is explicitly assigned to Elastic-net, and statistical validation to FDR (Lee et al., 4 Sep 2025). A plausible implication is that MOVE is intended to prepare a compact, structured representation rather than to act as the final decision layer.

3. Probabilistic formulation and latent representation

The manuscript gives MOVE a VAE-style objective with three stated ingredients: a reconstruction term for each modality, a KL divergence regularizer toward a prior $p(\mathbf{z})$, and an auxiliary cross-modal alignment loss weighted by $\lambda$ (Lee et al., 4 Sep 2025). As printed, the objective is syntactically incomplete:

$\mathcal{L}_{\text{MOVE} = \sum_{m=1}^{M} \mathbb{E}_{q_\phi^{(m)}(\mathbf{z}|\mathbf{x}^{(m)})}[\log p_\theta^{(m)}(\mathbf{x}^{(m)}|\mathbf{z})] - \beta \cdot D_{KL}(q_\phi^{(m)}(\mathbf{z}|\mathbf{x}^{(m)}) || p(\mathbf{z}))$

and the total loss is likewise truncated:

$\mathcal{L}_{\text{total} = \mathcal{L}_{\text{MOVE} + \lambda \cdot \mathcal{L}_{\text{cross}$

(Lee et al., 4 Sep 2025).

Despite this malformed presentation, the intended semantics are identifiable from the paper itself. The notation uses modality-indexed approximate posteriors and decoders,

$$q_\phi^{(m)}(\mathbf{z}\mid \mathbf{x}^{(m)}) \quad\text{and}\quad p_\theta^{(m)}(\mathbf{x}^{(m)}\mid \mathbf{z}),$$

which indicates modality-specific inference and generation functions tied through a common latent variable $\mathbf{z}$ (Lee et al., 4 Sep 2025). This suggests a shared latent representation across modalities. Because the manuscript does not state in prose that the latent space is shared, this should be treated as an inference from the displayed formula rather than a directly verbalized architectural claim.

The regularization structure is explicitly $\beta$-weighted KL plus $\lambda$-weighted cross-modal alignment (Lee et al., 4 Sep 2025). The paper says only that the auxiliary term “encourages cross-modal alignment”; it does not define whether $\mathcal{L}_{\text{cross}}$ is contrastive, cosine-based, adversarial, mean-squared alignment, or another form (Lee et al., 4 Sep 2025). Likewise, the paper does not provide an ELBO derivation, Gaussian parameterization $(\mu,\sigma)$, a reparameterization formula, modality-specific likelihood families, or modality weights (Lee et al., 4 Sep 2025).

Relative to other multi-omics VAEs, the implied design is conventional at a high level but under-specified in detail. OmiVAE explicitly defines a Gaussian posterior $\lambda$0, uses the reparameterization trick $\lambda$1, and trains with a joint loss $\lambda$2 (Zhang et al., 2019). CustOmics reviews the standard VAE formulation and then departs from it by replacing KL with MMD in its central latent module (Benkirane et al., 2022). MOVE, in contrast, remains only partially specified at the level of formal variational inference (Lee et al., 4 Sep 2025).

4. Data fusion, modality handling, and relation to neighboring architectures

The paper frames MOVE as the multi-omics fusion module of the framework. The integration sequence supported by the text is: GAT derives graph-informed representations from gene interaction structure; MOVE maps these heterogeneous omics-derived signals into a common lower-dimensional latent space; reconstruction terms are computed per modality; and a cross-modal alignment loss promotes agreement across omics (Lee et al., 4 Sep 2025). On that basis, MOVE is best classified as an intermediate integration model (Lee et al., 4 Sep 2025).

The explicit feature embedding strategy in prose is to preserve “gene-gene interactions extracted via GAT,” compress omics features into “biologically meaningful latent representations,” preserve “intrinsic geometry/manifold structure,” and estimate entity distributions while reconstructing the original data structure (Lee et al., 4 Sep 2025). The paper therefore does not describe MOVE as a raw concatenation model. Nor does it describe decision-level late integration.

The manuscript does not discuss missing-modality handling at all (Lee et al., 4 Sep 2025). There is no mention of partial modality availability, imputation, modality dropout, masked reconstruction, or paired/unpaired alignment objectives (Lee et al., 4 Sep 2025). This omission matters when MOVE is compared with methods explicitly built for incomplete data. MODIS, for instance, is designed for “small and unpaired datasets,” learns a shared latent space with one VAE per modality plus an adversarial discriminator, and supports “generation of missing data / missing modality reconstruction” (Lepe-Soltero et al., 24 Mar 2025). The “Multi-View Variational Autoencoder for Missing Value Imputation in Untargeted Metabolomics” uses a product-of-experts multi-view VAE to jointly model burden score, polygenetic risk score, and LD-pruned SNPs for missing metabolomics imputation (Zhao et al., 2023). MOVE’s paper offers no analogous missingness mechanism (Lee et al., 4 Sep 2025).

MOVE also differs from non-variational multi-omics autoencoders. The “Multi-view Factorization AutoEncoder with network constraints” is explicitly not variational and is described as a multi-view deterministic autoencoder / matrix-factorization hybrid with graph regularization (Ma et al., 2018). AIME is likewise deterministic and cross-modal rather than variational, learning an embedding of one omics modality that reconstructs another while optionally adjusting for confounders (Yu, 2019). MOVE, by contrast, is explicitly framed as a VAE/CVAE-style model with KL regularization and a latent prior (Lee et al., 4 Sep 2025).

5. Role in biomarker discovery and reported empirical outcomes

MOVE’s clearest contribution in the ensemble is dimensionality reduction and representation learning. The paper states that it is used to encode graph-derived multi-omics features into a “lower-dimensional latent space,” thereby mitigating the curse of dimensionality under HDLSS conditions (Lee et al., 4 Sep 2025). It also explicitly attributes to MOVE “cross-modal feature compression” and preservation of “biologically meaningful latent representations” (Lee et al., 4 Sep 2025).

The paper further states that MOVE “preserves gene-gene interactions extracted via GAT,” preserves “intrinsic geometry of the data,” and provides latent variables that serve as “the basis for estimating entity distributions and reconstructing the original data structure” (Lee et al., 4 Sep 2025). The downstream implication is that Elastic-net operates on a representation already shaped by graph-informed, cross-modal compression (Lee et al., 4 Sep 2025).

The reported quantitative results are framework-level rather than MOVE-specific. On the simulated dementia dataset, the proposed framework achieves AUC 0.93, F1 0.91, and Feature Precision 0.88, compared with DIABLO at AUC 0.84, F1 0.81, Feature Precision 0.72; MOCAT at AUC 0.86, F1 0.83, Feature Precision 0.75; AMOGEL at AUC 0.88, F1 0.85, Feature Precision 0.78; and MOMLIN at AUC 0.89, F1 0.86, Feature Precision 0.80 (Lee et al., 4 Sep 2025). On ADNI, the proposed framework achieves AUC 0.91, F1 0.89, and Feature Precision 0.87, compared with DIABLO at AUC 0.85, F1 0.82, Feature Precision 0.74; MOCAT at AUC 0.86, F1 0.83, Feature Precision 0.76; AMOGEL at AUC 0.88, F1 0.85, Feature Precision 0.79; and MOMLIN at AUC 0.89, F1 0.86, Feature Precision 0.81 (Lee et al., 4 Sep 2025).

These results cannot be attributed uniquely to MOVE. The paper does not report an ablation study removing MOVE, does not compare GAT+ElasticNet+FDR with and without MOVE, and does not provide a reconstruction error table, modality alignment metric, or latent-space quality metric specific to MOVE (Lee et al., 4 Sep 2025). The manuscript’s direct attribution is limited to the statement that “the Conditional Variational Autoencoder (CVAE) for multi-omics data (MOVE) compressed heterogeneous omics features into biologically meaningful latent representations” (Lee et al., 4 Sep 2025). A plausible implication is that MOVE is important to the framework’s performance, but its marginal contribution cannot be quantified from the manuscript.

6. Limitations, ambiguities, and interpretation

The most immediate limitation is incomplete specification. The MOVE equations are truncated or malformed, the total loss is not fully written, and $\lambda$3 is undefined (Lee et al., 4 Sep 2025). The manuscript also does not disclose the encoder or decoder architecture in concrete form: it gives no number of encoder branches, hidden layer sizes, activation functions, decoder symmetry, conditioning-variable design, or explicit modality-specific likelihoods (Lee et al., 4 Sep 2025).

Training details are likewise absent. The paper does not report latent dimensionality, learning rate, batch size, number of epochs, optimizer, scheduler, encoder/decoder widths, dropout rates, $\lambda$4 value, $\lambda$5 value, initialization, cross-validation design for MOVE, early stopping, or hyperparameter tuning method (Lee et al., 4 Sep 2025). As a result, MOVE is described at the level of intended function and partial objective, but not at a level that permits direct reconstruction of the model from the paper alone.

The paper also does not provide a dedicated interpretability mechanism for MOVE’s latent factors (Lee et al., 4 Sep 2025). It states that MOVE yields “biologically meaningful latent representations,” and the framework as a whole is presented as interpretable, but there is no explanation of how latent dimensions correspond to pathways, whether decoder weights are inspected, or whether saliency, attribution, or latent traversals are performed on MOVE specifically (Lee et al., 4 Sep 2025).

Several contextual concerns follow from these omissions. This suggests that MOVE, as published, is better understood as a sketched shared latent-space multi-omics fusion engine than as a fully documented standalone method. It also suggests caution in equating MOVE with more fully specified multimodal VAEs such as OmiVAE (Zhang et al., 2019), CustOmics (Benkirane et al., 2022), MODIS (Lepe-Soltero et al., 24 Mar 2025), or the product-of-experts multi-view VAE for metabolomics imputation (Zhao et al., 2023). In the broader technical review literature, methods in this family are typically situated within the taxonomy of multimodal VAEs with modality-specific encoders and decoders, using fusion strategies such as direct concatenation, mixture-of-experts, product-of-experts, or mixture-of-product-of-experts (Baião et al., 29 Jan 2025). MOVE’s notation is consistent with that broader family, but the paper does not reveal which specific fusion mechanism it instantiates (Lee et al., 4 Sep 2025).

The result is a precise but limited characterization. MOVE is a multi-omics VAE/CVAE-style module used after GAT and before Elastic-net and FDR; it is intended to compress genomics, transcriptomics, proteomics, and metabolomics into a biologically structured low-dimensional representation; it includes modality-wise reconstruction, KL regularization weighted by $\lambda$6, and a cross-modal alignment term weighted by $\lambda$7; and its notation suggests a common latent variable $\lambda$8 shared across modalities (Lee et al., 4 Sep 2025). Beyond that, the published account remains incomplete.