TarDis: Achieving Robust and Structured Disentanglement of Multiple Covariates

Abstract

Addressing challenges in domain invariance within single-cell genomics necessitates innovative strategies to manage the heterogeneity of multi-source datasets while maintaining the integrity of biological signals. We introduce TarDis, a novel deep generative model designed to disentangle intricate covariate structures across diverse biological datasets, distinguishing technical artifacts from true biological variations. By employing tailored covariate-specific loss components and a self-supervised approach, TarDis effectively generates multiple latent space representations that capture each continuous and categorical target covariate separately, along with unexplained variation. Our extensive evaluations demonstrate that TarDis outperforms existing methods in data integration, covariate disentanglement, and robust out-of-distribution predictions. The models capacity to produce interpretable and structured latent spaces, including its pioneering work in ordered latent representations for continuous covariates, markedly enhances its utility in hypothesis-driven research. Consequently, TarDis offers a promising analytical platform for advancing scientific discovery, providing insights into cellular dynamics, and enabling targeted therapeutic interventions. Progress and potentialModern single-cell genomics provides an unprecedented view into cellular heterogeneity, yet the very richness that propels new discoveries also complicates downstream analysis. Gene-expression patterns emerge from overlapping biological processes (e.g., differentiation programs, disease progression) and extrinsic factors (e.g., laboratory protocols, technical artifacts). Disentanglement, in this context, aims to parse these intertwined influences into interpretable latent representations, a crucial step for elucidating how complex covariates shape cellular states. While methods that correct for batch effects have become standard, these strategies often fall short in achieving the deeper objective of capturing subtle, high-dimensional biological dynamics. In single-cell experiments, cells navigate intricate developmental trajectories, respond nonlinearly to environmental or pharmaceutical perturbations, and exhibit myriad context-specific behaviors. Without disentanglement, these diverse signals frequently remain intermingled, limiting biological interpretability and hindering hypothesis-driven research. Disentangling biological covariates is particularly vital for addressing nuanced questions in single-cell research. For example, in a disease model involving multiple genetic variants and variable drug dosing, researchers may wish to examine the effect of each variant independently or investigate how dosage influences a specific mutant background. Similarly, in developmental biology, uncovering how cells evolve across a continuum of pseudotime (e.g., from pluripotent to fully differentiated states) is critical for identifying the genes that orchestrate fate decisions while isolating the influence of developmental time from tissue-specific contexts, along with other confounding factors such as culture conditions, sample preparation, or donor genetic characteristics. Alternatively, disentangling lineage commitment signals from spatial patterning cues enables the identification of master regulators driving fate decisions. Moreover, by explicitly isolating and representing each covariate as an independent latent dimension, one can systematically navigate and interrogate a rich multidimensional covariate space. This approach extends beyond merely observing biological states, it enables exploration of novel or unmeasured cellular conditions through latent-space manipulations. For instance, disentangled latent spaces could allow researchers to computationally predict cellular responses at drug dosages or developmental stages that were never experimentally observed, significantly broadening the scope and predictive power of experimental datasets. Such analyses yield testable hypotheses for unexplored biological phenomena and enable informed planning of subsequent experimental validations. The challenge of covariate disentanglement stems fundamentally from the complexity of modeling joint distributions of gene expression conditioned simultaneously on multiple covariates, both categorical (e.g., tissue type, disease condition) and continuous (e.g., pseudotime, dosage). This is inherently an underdetermined problem because single-cell measurements represent only sparse snapshots within a vast combinatorial space of covariate conditions. Conventional modeling approaches often conflate correlated covariates, collapsing biological variability into ambiguous latent factors, and typically fail to explicitly create separate latent representations for disentangled covariates. Moreover, continuous covariates introduce an additional layer of complexity; yet discretizing them artificially imposes arbitrary boundaries, obscuring subtle transitions and hindering accurate capture of biological gradients. Therefore, preserving the continuous nature of such covariates in disentangled representations is critical, as it maintains their intrinsic ordering and enables researchers to discern nuanced biological shifts--such as identifying thresholds in dose-response relationships or characterizing gradual developmental transitions--in a naturally interpretable manner. The key idea in this paper is to devise a tailored deep generative model for systematically separating both categorical and continuous covariates into independent latent dimensions, while still ensuring coherent integration of the underlying gene-expression data. By explicitly targeting these covariates and preserving continuous variables as smooth, ordered latent axes, our approach clarifies complex interactions and uncovers nuanced patterns that remain concealed under standard analyses. The resulting disentangled representations can then support robust out-of-distribution generalizations, refined differential analyses, and more principled hypotheses about how diverse factors interact to drive cellular variation.