Researchers at the Indian Institute of Technology Kanpur, working with a collaborator at the University of California, Irvine, have unveiled a new artificial intelligence framework that promises to make sense of one of the most data-rich but technically unruly corners of modern biology: single-cell multiomics. The tool, called multiHIVE, is a hierarchical multimodal deep generative model designed to weave together measurements of different molecular layers taken from the very same cell, producing a unified picture of cellular state that is more accurate, more complete, and more interpretable than what existing methods can achieve. The work, published in BMC Bioinformatics, arrives at a moment when experimental technologies are racing ahead of the computational methods needed to analyze their output, and it addresses head-on the statistical headaches that arise when biologists try to combine, say, gene expression counts with measurements of chromatin accessibility or protein abundance from a single nucleus.
Single-cell multiomics technologies represent a major leap over earlier single-cell approaches, which could typically measure only one molecular modality at a time. Modern assays can now jointly profile multiple modalities — such as RNA transcripts, DNA accessibility at open chromatin regions, and surface protein abundances — within individual cells. This joint profiling is extraordinarily valuable because cellular identity and behavior are governed by layered molecular processes: chromatin state shapes which genes can be transcribed, transcription defines messenger RNA output, and translation and protein regulation determine the functional machinery of the cell. However, each modality arrives with its own technical quirks. RNA counts follow sparse, zero-inflated count distributions driven by dropout and sequencing depth; chromatin accessibility data are extremely sparse binary-like signals; protein measurements obtained through antibody-based tagging carry different noise structures altogether. Methods that treat these heterogeneous data as if they were interchangeable tend to blur the very biological distinctions researchers are trying to detect.
MultiHIVE confronts this problem with a probabilistic architecture built on variational inference, the same family of techniques that underpins variational autoencoders. At its core, the model learns a compressed latent representation of each cell — a mathematical embedding in which similar cells cluster together — but it does so in a deliberately structured, hierarchical way. Rather than forcing all information into a single latent space, multiHIVE factorizes the cellular signal into two components: a shared latent variable that captures biological programs common across all measured modalities, and modality-specific latent variables that retain information unique to each data type. These latent variables are stacked hierarchically, meaning the model can represent biological variation at multiple levels of abstraction, from broad cell-type identity down to finer, modality-dependent subtleties. This factorization is the conceptual heart of the method: it acknowledges that multiomics data contain both a common biological core and modality-specific structure, and it refuses to sacrifice either one.
The generative side of the model is equally carefully constructed. For each modality, multiHIVE uses an observation likelihood matched to that data type’s statistical behavior — for instance, negative binomial or related count distributions for RNA and accessibility data, which can absorb overdispersion and zero inflation. By conditioning the generative process on both the shared and private latent variables, the decoder can reconstruct each modality faithfully while the encoder, guided by the hierarchical prior, learns to disentangle shared biology from technical and modality-specific noise. The upshot is that the model performs three tasks simultaneously: integration, in which cells from different batches or experimental conditions are aligned into a common space; denoising, in which the model’s reconstructions suppress technical artifacts; and imputation, in which missing measurements in one modality are predicted from the others. In datasets where some cells have RNA profiles but lack protein measurements, or vice versa, multiHIVE can fill in the gaps with statistically grounded estimates rather than naive averages.
To validate the approach, the team carried out extensive benchmarking against established methods on gold-standard real datasets, spanning both bi-modal configurations, which typically pair gene expression with chromatin accessibility, and the more demanding tri-modal setting, which adds protein abundance into the mix. Across these benchmarks, multiHIVE consistently outperformed competing approaches in integrating the multiomics measurements and in imputing missing features. The evaluation metrics probed whether cells of the same known biological type were correctly aligned across modalities and whether imputed values recovered ground-truth measurements held out during testing. The superiority in these tests suggests that the hierarchical factorization is not merely an architectural nicety but a genuine statistical advantage: by modeling what is shared and what is private separately, the model avoids both the over-smoothing that destroys modality-specific signal and the fragmentation that prevents coherent integration.
The researchers then pushed the method beyond static snapshots into the dynamic realm of cellular development. Using a dataset of thymocytes — the developing immune cells that mature within the thymus before becoming T cells — they showed that multiHIVE’s learned cellular embeddings support improved trajectory inference, the computational reconstruction of the paths cells travel as they differentiate, and more accurate identification of gene trends along those paths. Trajectory inference is notoriously sensitive to noise and to the quality of the underlying embedding; small distortions early in the analysis can scramble the apparent ordering of developmental stages. The fact that multiHIVE’s embeddings sharpened both the trajectories and the gene-expression dynamics along them indicates that the integrated representation preserves genuine developmental signals that noisier or poorly integrated embeddings tend to wash out. For immunologists, this means a clearer, more faithful map of how progenitor cells acquire the molecular hallmarks of functional T cell lineages.
Perhaps the most biologically consequential demonstration came from datasets spanning development and disease, where the authors applied multiHIVE’s denoising machinery to extract gene expression programs from the integrated data. Gene program analysis seeks to decompose the expression matrix into coherent modules of genes that act together — reflecting processes such as cell cycling, immune signaling, metabolic rewiring, or stress responses. Because multiHIVE explicitly separates shared from modality-specific variation and denoises the reconstructed expression, the resulting programs were cleaner and more interpretable, and they revealed biological processes operating at multiple levels of the cellular hierarchy, from broad lineage-defining programs down to finer, state-specific modules. This capacity to move from raw, noisy multiomic measurements to named biological processes is precisely what experimentalists need, and it positions multiHIVE as more than a data-processing utility — it is a hypothesis-generating engine.
Scalability, often the Achilles heel of sophisticated generative models, also receives attention in the design. Single-cell datasets now routinely contain hundreds of thousands to millions of cells, and any method that cannot scale will remain a curiosity regardless of its statistical elegance. The authors describe multiHIVE as a robust and scalable framework, and the model’s implementation choices reflect the engineering realities of modern genomics: the software is written in Python, integrates with the standard ecosystem of single-cell analysis tools, and has been released publicly on GitHub at https://github.com/Zafar-Lab/multiHIVE, allowing any laboratory to adopt, scrutinize, and extend it. Open availability matters in this field, where reproducibility and community benchmarking drive rapid methodological progress, and where the best ideas are often refined through widespread real-world testing on diverse tissue types and disease contexts.
The broader significance of the work lies in what it says about the trajectory of computational biology. As multiomic profiling becomes routine — tri-modal assays combining transcriptome, epigenome, and proteome are increasingly accessible — the bottleneck is shifting from data generation to data interpretation. Models like multiHIVE illustrate a maturing design philosophy: rather than bolting together separate single-modality analyses and hoping the results align, researchers are building generative models that respect the joint structure of the data from the ground up. The hierarchical latent-variable approach also echoes developments in machine learning at large, where structured representations have proven superior to monolithic embeddings when data contain genuinely heterogeneous sources of variation. Biology, with its layered regulatory logic, may be an ideal proving ground for these ideas.
For working biologists, the practical implications are immediate. Labs studying development can trace cell fates with greater confidence; immunologists can map maturation pathways across modalities; cancer researchers can integrate epigenetic and transcriptional dissection of tumor heterogeneity to identify malignant subpopulations that might otherwise hide within a single modality’s noise. Because the model’s denoised outputs feed directly into interpretable downstream analyses such as gene program discovery, the path from raw data to biological insight is shortened and made more reliable. The authors emphasize that multiHIVE preserves the complementary information carried by each modality — the epigenome’s glimpse into regulatory potential, the transcriptome’s record of current activity, the proteome’s report on functional capacity — rather than collapsing them into a lossy average.
The study was led by Musale Krushna Pavan and Anirudh Nanduri, who contributed equally, together with Kushagra Pandey of the University of California, Irvine, and corresponding author Hamim Zafar of IIT Kanpur, where Zafar holds appointments in Computer Science and Engineering, Biological Sciences and Bioengineering, and the Mehta Family Centre for Engineering in Medicine. The research was supported by the DBT/Wellcome Trust India Alliance, the DBT Har Gobind Khorana Innovative Young Biotechnologist Award, an IIT Kanpur initiation grant, and the Prime Minister’s Research Fellowship, with additional support from the Research-I Foundation of the Department of Computer Science and Engineering at IIT Kanpur. Published open access, the paper invites the global single-cell community to test the method on their own multimodal datasets — and if the benchmark results hold up in the wild, multiHIVE may soon become a standard tool in the multiomics analyst’s arsenal, turning the noisy, fragmentary measurements of individual cells into coherent, biologically meaningful stories.
Subject of Research: Single-cell multiomics integration, denoising, and imputation using hierarchical multimodal deep generative modeling
Subject of Research: Biology
Article Title: multiHIVE: hierarchical multimodal deep generative modeling for single-cell multiomics
Article References: Pavan, M. K., Nanduri, A., Pandey, K., & Zafar, H. (2026). multiHIVE: hierarchical multimodal deep generative modeling for single-cell multiomics. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06613-x
Image Credits: AI Generated
DOI: 10.1186/s12859-026-06613-x
Keywords: single-cell multiomics, hierarchical deep generative model, variational autoencoders, multimodal integration, imputation, denoising, trajectory inference, gene expression programs, tri-modality, cellular embeddings
Cite Scienmag News
APA
MLA
Chicago
Juliet Wilcox. (September 10, 2026). Hierarchical AI model unifies single-cell multiomics data through deep generative modeling. Scienmag. https://scienmag.com/hierarchical-ai-model-unifies-single-cell-multiomics-data-through-deep-generative-modeling/
Juliet Wilcox. “Hierarchical AI model unifies single-cell multiomics data through deep generative modeling.” Scienmag, 10 September 2026, https://scienmag.com/hierarchical-ai-model-unifies-single-cell-multiomics-data-through-deep-generative-modeling/. Accessed 10 September 2026.
Juliet Wilcox. “Hierarchical AI model unifies single-cell multiomics data through deep generative modeling.” Scienmag. September 10, 2026. https://scienmag.com/hierarchical-ai-model-unifies-single-cell-multiomics-data-through-deep-generative-modeling/
Copy citation
Download RIS
Tags: advanced bioinformatics tools for multi-modal dataadvancements in single-cell multiomics technologyAI-driven biological data integrationcombining gene expression and chromatin accessibilitycomputational methods for multi-modal biologycomputational methods for single-cell multiomicsdeep learning for multiomics datadeep learning in single-cell multiomicshierarchical deep generative models in biologymulti-modal single-cell data analysismultiHIVE AI framework for cellular datamultiHIVE artificial intelligence frameworkmultiomics data interpretation toolsmultiomics data visualization and interpretationmultiomics measurement integrationsingle-cell multiomics data challengessingle-cell multiomics data integrationsingle-cell multiomics data visualizationstatistical challenges in multiomics dataunified cellular state modeling
