Variational Autoencoders for Biomedical Data: New Study Maps Their Inner Workings

Variational autoencoders, a class of neural networks widely used to fuse messy, high-dimensional biomedical datasets, have gained a rare look under the hood. In a study published on 2 September 2026 in BMC Artificial Intelligence, researchers led by Marc Pielies Avellí, Ricardo Hernández Medina, Henry Webel and Simon Rasmussen dissect, step by step, how their multimodal VAE framework called MOVE actually behaves inside, and where its celebrated in silico perturbation analyses can quietly mislead. The work combines synthetic datasets with well-defined ground truths, a multi-omics cohort of inflammatory bowel disease patients, and single-cell CRISPR perturbation data from two human cell lines.

A variational autoencoder consists of an encoder that compresses high-dimensional inputs into a lower-dimensional latent representation and a decoder that reconstructs the original data from that compressed code. The model is trained to minimize reconstruction error while organizing samples in latent space according to a prior distribution. VAEs have been applied to patient stratification, single-cell data embedding and metagenome binning, and the team’s earlier MOVE framework was demonstrated on a cohort of type 2 diabetes patients, integrating continuous and categorical biomedical data in a single model.

The central problem the authors tackle is that perturbation analyses in such models were originally built for categorical variables, such as flipping a patient’s medication status, while many biomedical questions demand perturbations of continuous quantities, like raising the concentration of a metabolite or protein. To address this, the team extended MOVE to handle continuous perturbations, introduced the Kolmogorov-Smirnov distance as a new metric for the magnitude and consistency of reconstruction shifts, and built synthetic datasets in which the true associations between variables were known by construction, allowing a rigorous benchmark.

The synthetic experiments revealed that a VAE’s behavior depends strongly on the properties of the dataset it is fed. With few samples relative to features, spurious correlations emerged in the learned representations, but these vanished as sample numbers grew. Under a low-correlation regime, the model could only learn a limited number of independent input features, constrained by the network’s bottleneck, echoing earlier theoretical work showing that separate latent dimensions are needed for each independent input component. In contrast, when variables were entangled, as gene co-expression patterns typically are, the data lay on a lower-dimensional manifold that the VAE could compress and reconstruct far more effectively.

The study also clarified how the training objective sculpts the latent space itself. To ease reconstruction, the encoder separated samples into clusters based on shared categorical labels and ordered them along value gradients of learned continuous variables, while features the model failed to learn showed no such structure and their reconstructions suffered. The relative weighting of categorical versus continuous data modalities proved to be a key hyperparameter: heavy emphasis on categorical reconstruction produced tightly clustered but sparsely populated latent spaces, whereas upweighting continuous data yielded a more homogeneous arrangement. A SHAP-inspired analysis, in which each feature was set to zero and the resulting movement of samples in latent space was measured, faithfully reflected this organization and offered a first hint of which variables the model had actually learned.

Perturbation design turned out to matter enormously. Simply setting a feature to the cohort’s minimum or maximum value affected samples unevenly and risked pushing them into unexplored, out-of-distribution regions of latent space, where the decoder, trained only on plausible unperturbed data, could produce reconstructions lacking biological meaning. The authors instead proposed a context-preserving scheme that adds or subtracts one standard deviation from each feature’s value, moving samples smoothly within the learned manifold and modeling physiological variation. Substitution remains suitable for simulating knock-out scenarios, but the choice between methods, they argue, should be dictated by the biological context and the risk of out-of-distribution effects.

To quantify perturbation effects, the team compared three statistical approaches. The original t-test on reconstructed values produced a high false-positive rate. A Bayesian-inspired method achieved better precision but scored only the consistency of change across samples, saturating at its maximum for many associations and making prioritization difficult. The new Kolmogorov-Smirnov distance, the maximum vertical difference between the cumulative distributions of reconstructed values before and after perturbation, captured both consistency and magnitude, and its stricter sample-size-dependent significance threshold yielded fewer but better-ranked associations, with 55 true positives versus 96 for the Bayesian method in one synthetic benchmark.

Applied to a real multi-omics dataset of inflammatory bowel disease from Lloyd-Price and colleagues, containing metagenomics, metatranscriptomics and metabolomics from 293 samples, MOVE clustered samples by disease and dysbiosis status and arranged them along gradients of learned metabolites such as cholate. Perturbing dysbiosis status in non-dysbiotic samples reproduced known biology, predicting increases in the primary bile acid cholate and decreases in the secondary bile acids litocholate and deoxycholate, and confirming previously reported associations with carnitines and with the bacterium Roseburia hominis. Yet the perturbed reconstructions deviated only slightly from baseline, with KS distances around 0.08 against a significance threshold of roughly 0.11 for this sample size, and perturbed samples often reverted to their original categorical label rather than transitioning to the target state.

The team then extended the perturbation paradigm to the fashionable quest for a virtual cell, benchmarking MOVE against methods including scGPT, GEARS, CPA and a bilinear baseline on Perturb-seq knockdown data from K562 and RPE1 cells. MOVE’s L2 errors were comparable to those of the other models, but its Pearson delta correlations, ranging from −0.178 to 0.119 for RPE1 and −0.107 to 0.022 for K562, underperformed all non-VAE-based approaches, including even a simple mean baseline. Low regularization improved accuracy but invited overfitting, while high regularization triggered posterior collapse, in which the model simply predicted the mean perturbed profile. Notably, the tendency to predict changes smaller than those observed was not unique to MOVE but appeared in competing baselines as well, and standard evaluation metrics, being rank-based or dominated by barely changing genes, did not penalize this shrinkage.

The authors’ overarching conclusion is a cautionary one: while VAEs can correctly infer the direction of change after a perturbation, the magnitude is often small, perturbed samples may revert to their original state, and current benchmarks reward approximating average perturbation effects without capturing how meaningful the predicted perturbed state actually is. Because latent space dynamics are local rather than global, simple vector arithmetic borrowed from language models does not transfer to multimodal biomedical latent spaces. The researchers suggest using Bayesian-inspired scores when hunting for small but consistent shifts and KS-based scores when prioritizing larger effect sizes that should be confirmed in larger cohorts, and they point toward integrating explicit network topologies with VAEs as a promising future direction for multi-omics integration.

Subject of Research: How variational autoencoders integrate multimodal biomedical data and perform in silico perturbation analyses

Article Title: On the use of variational autoencoders for biomedical data integration

Article References: Pielies Avellí, M., Hernández Medina, R., Webel, H., & Rasmussen, S. (2026). On the use of variational autoencoders for biomedical data integration. BMC Artificial Intelligence, 2(1), Article 18. https://doi.org/10.1186/s44398-026-00034-9

Image Credits: AI Generated

DOI: 10.1186/s44398-026-00034-9

Keywords: variational autoencoders, biomedical data integration, multi-omics, in silico perturbations, latent space, inflammatory bowel disease, Perturb-seq, Kolmogorov-Smirnov distance, single-cell transcriptomics, MOVE framework, machine learning, synthetic benchmarks