Generative artificial intelligence models applied to high-dimensional, low-sample-size biological data frequently rely on aggressive dimensionality reduction and remain insufficiently validated when replicate counts are extremely low, according to a critical review published in Artificial Intelligence Review. The authors surveyed the research literature from 2015 through 2025 to evaluate major generative families, focusing on their behavior, stability, and generative fidelity under data-scarce biological conditions.
Generative artificial intelligence has become central to modern computational analysis within high-dimensional, low-sample-size biological domains. To organize this field, the authors established a taxonomy of current generative paradigms, critically examining Generative Adversarial Networks, Variational Autoencoders, diffusion-based probabilistic models, and energy-based models. Their analysis specifically examined how researchers adapt these architectures to molecular and omics data settings.
The surveyed literature demonstrates that generative models are applied across multiple specialized omics modalities, notably single-cell RNA sequencing, bulk RNA sequencing, spatial transcriptomics, and metabolomics.
A central recurring finding across the surveyed publications is that most generative models depend heavily on aggressive preprocessing procedures prior to model training. Workflows routinely implement highly variable gene selection, calculate principal component analysis, or depend on encoder-decoder latent spaces. By relying on these initial steps, existing pipelines effectively replace the original high-dimensional, low-sample-size problem with a lower-dimensional proxy.
Furthermore, the review revealed that existing generative architectures are predominantly evaluated on moderate-sample or large-sample biological datasets. Consequently, these models remain insufficiently validated in biological applications where sample sizes are extremely constrained, such as experimental designs with fewer than ten replicates.
The authors also identified a key methodological gap in current generative modeling: the limited development of embedded sparsity-aware feature selection within these frameworks.
To complement their literature survey, the researchers conducted experimental comparisons of representative Variational Autoencoder, Generative Adversarial Network, and diffusion-based models using benchmark biological datasets in the high-dimensional, low-sample-size regime. This empirical comparison directly tested model behavior, output stability, and generative fidelity when operating under genuine data-scarce constraints across distinct algorithmic paradigms.
Across the representative implementations tested, the Variational Autoencoder exhibited comparatively stable behavior under severe high-dimensional, low-sample-size conditions. In contrast, the evaluated LSH-GAN and scDiffusion implementations demonstrated greater sensitivity to sample availability, with the degree of generative improvement observed at larger sample sizes differing notably between the evaluated machine learning architectures.
Subject of Research: Generative artificial intelligence for high-dimensional, low-sample-size biological data
Article Title: Generative AI for high-dimensional biological data under HDLSS constraints: a critical review and comparative study
Article References: Ghosh, T., Surineela, S. V. K., Hendren, C., Kanakamalla, P., Ghaffarnia, A., Harikumar, H., & Markey, M. (2026). Generative AI for high-dimensional biological data under HDLSS constraints: a critical review and comparative study. Artificial Intelligence Review. https://doi.org/10.1007/s10462-026-11758-5
Image Credits: AI Generated
DOI: 10.1007/s10462-026-11758-5
Keywords: generative artificial intelligence, high-dimensional biological data, HDLSS regime, variational autoencoders, generative adversarial networks, diffusion-based probabilistic models, energy-based models, single-cell RNA sequencing, bulk RNA sequencing, spatial transcriptomics, metabolomics, data scarcity, dimensionality reduction, principal component analysis, highly variable gene selection, sparsity-aware feature selection, LSH-GAN, scDiffusion
News Source: Juliet Wilcox. (November 20, 2026). Generative AI in High-Dimensional Biology Remains Insufficiently Validated in Small-Sample Studies. Scienmag.



