Executive Summary: Synchronized Optimal Transport Reconstructs Coupled Transcriptomic and
Core discovery and governing mechanism. The study introduces Cross‑Omics Aligned Trajectory Inference (COATI), a computational framework that extends unbalanced synchronized dynamical optimal transport to jointly model time‑dependent single‑cell RNA‑seq (scRNA‑seq) and ATAC‑seq (scATAC‑seq) landscapes. By treating RNA and chromatin accessibility as distinct probability measures evolving over developmental time, COATI solves a coupled optimal‑transport problem that simultaneously aligns modalities and infers population‑mass changes, thereby capturing both transcriptional and regulatory dynamics in a unified causal trajectory.
Primary empirical benchmark and theoretical advancement. COATI was evaluated on (i) synthetic benchmarks designed to mimic known lineage bifurcations and (ii) four real developmental datasets spanning early embryogenesis to organogenesis. Using a held‑out‑stage reconstruction protocol, COATI consistently outperformed existing single‑modality and shared‑representation methods in recapitulating the composition of unseen stages, notably recovering intra‑type chromatin‑state heterogeneity that other approaches missed. The unbalanced optimal‑transport formulation further revealed stage‑specific mass fluxes that correlated with DNA‑replication programs in early development and with lineage‑maturation programs later, providing a quantitative link between population dynamics and underlying biological processes.
Strategic significance and practical implications. By delivering coupled, high‑resolution trajectories across transcriptomic and epigenomic spaces, COATI enables precise dissection of the temporal ordering of regulatory events, facilitating the identification of causal transcription‑regulation relationships. This capability accelerates hypothesis generation for developmental biology, improves the design of CRISPR‑based perturbation screens, and informs regenerative medicine strategies that require coordinated manipulation of gene expression and chromatin state. Moreover, the unbalanced transport paradigm offers a generalizable template for integrating any pair of destructive, time‑course single‑cell modalities, opening avenues for multi‑omics atlases that faithfully reconstruct dynamic biological systems.
Theoretical Framework & Mechanisms: Synchronized Optimal Transport Reconstructs Coupled Transcriptomic and
COATI formalizes the joint evolution of transcriptional ($\mathcal{X}_R$) and chromatin ($\mathcal{X}_A$) state spaces as a pair of time‑dependent probability measures $$\mu_t^{R}\in\mathcal{P}(\mathcal{X}_R),\qquad \mu_t^{A}\in\mathcal{P}(\mathcal{X}_A),\qquad t\in[0,1],$$ each obeying an unbalanced continuity equation
$$\partial_t\mu_t^{\kappa}+\nabla\!\cdot\!\bigl(v_t^{\kappa}\,\mu_t^{\kappa}\bigr)=\sigma_t^{\kappa},\qquad \kappa\in\{R,A\},$$where $v_t^{\kappa}$ denotes the modality‑specific velocity field and $\sigma_t^{\kappa}$ a source term that captures net cell proliferation or loss. The unbalanced formulation permits $\int d\mu_t^{\kappa}\neq\int d\mu_{t'}^{\kappa}$, a critical departure from classical optimal transport (OT) that assumes mass conservation.
Coupling between modalities is introduced through a time‑varying alignment map $$\psi_t:\mathcal{X}_R\rightarrow\mathcal{X}_A,$$ assumed differentiable and invertible on the support of $\mu_t^{R}$. The push‑forward $\psi_{t\#}\mu_t^{R}$ is required to remain close to $\mu_t^{A}$, which is enforced by a synchronized penalty term based on the Kullback–Leibler (KL) divergence:
$$\mathcal{L}_{\text{sync}}(t)=\lambda_{\text{sync}}\;\mathrm{KL}\!\bigl(\psi_{t\#}\mu_t^{R}\,\|\,\mu_t^{A}\bigr),$$with $\lambda_{\text{sync}}>0$ controlling the strength of cross‑modal synchrony. The total dynamical OT problem becomes
$$\begin{aligned} \inf_{\substack{\mu^{R},\mu^{A},\\ v^{R},v^{A},\psi}} \; &\int_{0}^{1}\!\Bigg[ \underbrace{\int_{\mathcal{X}_R}\tfrac12\|v_t^{R}(x)\|^{2}\,d\mu_t^{R}(x)}_{\text{RNA kinetic cost}} +\underbrace{\int_{\mathcal{X}_A}\tfrac12\|v_t^{A}(y)\|^{2}\,d\mu_t^{A}(y)}_{\text{ATAC kinetic cost}}\\ &\qquad\qquad +\underbrace{\lambda_{\text{bal}}\bigl[\mathrm{KL}(\mu_t^{R}\|\nu^{R})+\mathrm{KL}(\mu_t^{A}\|\nu^{A})\bigr]}_{\text{unbalanced mass regularization}} +\mathcal{L}_{\text{sync}}(t) \Bigg]dt, \end{aligned}$$subject to the continuity constraints above and fixed reference measures $\nu^{R},\nu^{A}$ that encode prior knowledge of baseline cell‑type distributions. The kinetic terms arise from the Benamou–Brenier formulation of OT, guaranteeing that the optimal $v_t^{\kappa}$ are gradients of scalar potentials $\phi_t^{\kappa}$, i.e. $v_t^{\kappa}=\nabla\phi_t^{\kappa}$, which permits efficient computation via convex duality.
Algorithmically, COATI solves the saddle‑point problem by alternating between:
- Entropic Sinkhorn updates for the transport plans $\gamma_t^{\kappa}$ that approximate the optimal couplings under the regularized cost $c^{\kappa}(x,y)=\|x-y\|^{2}$;
- Gradient descent on $\psi_t$ using the synchronized KL gradient $\nabla_{\psi_t}\mathcal{L}_{\text{sync}}$, which aligns the RNA manifold onto the ATAC manifold while preserving smoothness via a Sobolev regularizer $\|\nabla\psi_t\|_{2}^{2}$;
- Closed‑form updates for the source terms $\sigma_t^{\kappa}$ derived from the first‑order optimality condition $$\sigma_t^{\kappa}= \lambda_{\text{bal}}\bigl(\nu^{\kappa}-\mu_t^{\kappa}\bigr).$$
This architecture resolves two entrenched bottlenecks. First, by allowing $\sigma_t^{\kappa}\neq0$, COATI captures developmental expansions and contractions of cell populations that balanced OT cannot represent, linking observed mass changes to underlying replication or differentiation programs. Second, the synchronized alignment term enforces temporal concordance across distinct high‑dimensional manifolds without collapsing them into a single latent space, thereby preserving modality‑specific regulatory information while still enabling joint trajectory inference. The resulting coupled dynamical system yields a mathematically rigorous, biologically interpretable reconstruction of transcriptional and chromatin evolution across developmental time points.
Empirical Evidence & Methodology: Synchronized Optimal Transport Reconstructs Coupled Transcriptomic and
The empirical validation of the Cross-Omics Aligned Trajectory Inference (COATI) framework was conducted through a rigorous dual-track assessment involving synthetic benchmarks and four distinct developmental datasets. This experimental design specifically addressed the inherent challenge of destructive sequencing, where continuous trajectories must be inferred from independently sampled cell populations across time-course snapshots. The primary metric for evaluation was held-out-stage reconstruction, a methodology that tests the model’s capacity to accurately predict cell populations at developmental stages excluded from the training data. This approach ensures that the inferred dynamics are not merely interpolative but possess genuine predictive power regarding the temporal evolution of single-cell states.
Quantitative analysis revealed that COATI significantly improved the reconstruction of cell populations at these held-out stages compared to baseline methods. A critical finding was the framework’s ability to recover chromatin-state heterogeneity within defined cell types, a nuance often lost in single-modality or shared-representation integration approaches. Furthermore, the model better preserved developmental lineage relationships, ensuring that the inferred trajectories remained biologically coherent with known differentiation hierarchies. The unbalanced formulation of the optimal transport problem proved particularly informative, allowing for the inference of population mass changes. These mass dynamics were empirically shown to associate with DNA-replication programs during early developmental stages and with lineage-maturation programs in later phases, providing a mechanistic link between transport geometry and biological regulation.
Crucially, the validity of the coupled trajectories was corroborated through independent experimental verification. Lineage-specific responses inferred by COATI demonstrated strong agreement with independent CRISPR perturbation results, confirming the causal relevance of the identified regulatory interactions. Additionally, the inferred regulatory patterns were consistent with independent epigenetic data, further validating the accuracy of the reconstructed chromatin dynamics. This convergence of computational inference with orthogonal experimental evidence underscores the robustness of the synchronized optimal transport approach in modeling complex, multi-omic developmental landscapes.
Research Attribution:
Lead Authors: Jingfeng Wang, Yanxiang Zhao, Qing Nie, Zixuan Cang
Primary Affiliations: Chongqing University, Tianjin University
Publication Venue: arXiv Preprint (Quantitative Biology — Genomics (q-bio.GN))
Identifier: arXiv:2611.08853 | 📄 DOI: 10.48550/arXiv.2611.08853
Canonical Link: https://arxiv.org/abs/2611.08853
Scientific Impact & Future Horizons: Synchronized Optimal Transport Reconstructs Coupled Transcriptomic and
While COATI demonstrates that synchronized optimal transport can faithfully reconstruct coupled transcriptional‑chromatin trajectories, several methodological bottlenecks temper its immediate deployment. First, the unbalanced dynamical formulation incurs a computational complexity that scales super‑linearly with the number of cells and the dimensionality of each modality, limiting routine analysis of million‑cell atlases. Second, COATI assumes that the temporal ordering of snapshots is known a priori; in practice, experimental designs often contain irregular or missing time points, which can bias the inferred transport maps. Third, the choice of ground‑cost metric remains heuristic—Euclidean distances in latent space may not capture the true energetic landscape of chromatin remodeling, especially when batch effects or sparsity dominate ATAC‑seq signals. Finally, the current implementation treats RNA and ATAC as independent state spaces linked only through the transport coupling, precluding the exploitation of known cis‑regulatory relationships that could regularize the solution.
Future research will therefore pursue three synergistic directions. (i) Algorithmic acceleration through stochastic Sinkhorn iterations and GPU‑native kernels will render COATI tractable for >10⁶ cells while preserving the fidelity of unbalanced mass transport. (ii) Adaptive cost learning—embedding biologically informed priors such as TF‑motif similarity or chromatin‑contact maps—will refine the geometry of the transport problem and mitigate over‑smoothness. (iii) Extension to multimodal hierarchies (e.g., spatial transcriptomics, proteomics, and lineage‑tracing barcodes) will enable a unified dynamical model that simultaneously respects spatial constraints and clonal ancestry.
Core Takeaways
- Fundamental Mechanism: COATI casts the joint evolution of RNA‑seq and ATAC‑seq cell populations as a pair of time‑dependent probability measures linked by an unbalanced synchronized optimal‑transport map, thereby recovering both population‑size dynamics and modality‑specific state transitions.
- Real‑World Value: By exposing the temporal ordering of chromatin accessibility changes relative to transcriptional programs, COATI provides a quantitative scaffold for pinpointing regulatory checkpoints that can be targeted in stem‑cell differentiation protocols, cancer lineage tracing, and epigenetic drug discovery.
Applications & Future Outlook
The ability to reconstruct coupled trajectories opens concrete translational pathways: (a) in regenerative medicine, COATI can guide the design of differentiation cocktails that synchronize epigenomic priming with gene‑expression milestones; (b) in oncology, the framework can map tumor evolution under therapy, revealing windows where chromatin‑modifying agents are most effective; (c) in policy‑driven biobank curation, the unbalanced mass component offers a principled metric for assessing sampling bias across developmental stages. Open challenges include robust handling of dropout‑induced sparsity, integration of lineage‑tracing barcodes to validate inferred ancestry, and standardization of cost‑function selection across diverse tissue contexts.
- Wang, J., Zhao, Y., Nie, Q., & Cang, Z. (2026). Synchronized Optimal Transport Reconstructs Coupled Transcriptomic and Chromatin Dynamics. arXiv Preprint, Quantitative Biology — Genomics (q‑bio.GN). 📄 DOI: 10.48550/arXiv.2611.08853
- Schiebinger, G., et al. (2019). Optimal‑Transport Analysis of Single‑Cell Gene Expression Identifies Developmental Trajectories. Cell, 176(4), 925–938.e22. 📄 DOI: 10.1016/j.cell.2019.01.006
- Peyré, G., & Cuturi, M. (2019). Computational Optimal Transport: With Applications to Data Science. Foundations and Trends® in Machine Learning, 11(5‑6), 355–607. 📄 DOI: 10.1561/2200000073
💬 Comments