CRAK-Velo: chromatin accessibility kinetics integration improves RNA velocity estimation Authors: Nour El Kazwini, Mingze Gao, Idris Kouadri Boudjelthia, Fangxin Cai, Yuanhua Huang, Guido Sanguinetti Published: 2026 ABSTRACT AbstractRNA velocity has recently emerged as a key tool in the analysis of single-cell transcriptomic data, yet connecting RNA velocity analyses to underlying regulatory processes has proved challenging. Here, we propose CRAK-Velo, a semi-mechanistic model that integrates chromatin accessibility data in the estimation of RNA velocities. CRAK-Velo provides biologically consistent estimates of developmental flows and enables accurate cell-type deconvolution, while additionally shining light on regulatory processes at the level of interactions between genes and chromatin regions.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13059-026-04086-y.Keywords: RNA velocity, Chromatin accessibility, Region kinetics Background Recent advances in single-cell sequencing technologies are leading to unprecedented insights in the molecular biology of cells [1–10]. However, the intrinsically destructive nature of sequencing creates a significant limitation in using such technologies to elucidate the dynamics of cellular processes. To circumvent this problem, La Manno et al. [11] proposed the concept of RNA velocity, a ground-breaking method which exploits splicing information to determine whether individual genes are being up-regulated or repressed at the time the sequencing happened. When aggregated at the cell level, this information provides valuable insight into the developmental or adaptation processes happening through the population of cells. The RNA velocity concept has been extended and refined in a number of recent papers [12–17], yet all of these methods need to specify or learn from data the transcription rate of individual genes in each cell. This either imposes significant constraints on the models (when the transcription rate is postulated to be e.g. a binary switch), or greatly expands the number of parameters required by the model. Alternatively, one might use additional epigenomic information, such as chromatin accessibility, to explain transcription rates; such information is increasingly becoming available through multi-omic single-cell techniques, including SNARE-seq [18], SHARE-seq [15], and the 10x Genomics Multiome commercial kit. This idea has been recently exploited by MultiVelo [19], an elegant model which aggregates chromatin accessibility and expression data through a differential equations latent variable model. Here, we present CRAK-Velo (ChRomatin Accessibility Kinetics integration in RNA Velocity), a simpler model that directly integrates chromatin accessibility data in the estimation of individual gene transcription rates. Results CRAK-Velo draws inspiration from UniTVelo [12], a recently proposed RNA velocity model which estimates transcription rates using a parametric approximation, but expands it by explicitly connecting transcription rates to observed open region probabilities (quantified using a probabilistic topic model [20, 21]; see Methods for the mathematical formulation of the model). This achieves two main goals: it regularizes the estimation of transcription rates, ensuring consistency with chromatin information, and it quantifies the impact of each regulatory region during transcriptional dynamics. In the following, we demonstrate CRAK-Velo on two challenging developmental data sets, as well as on a data set included in the additional files (Additional file 1: Table S2). We report run-time comparisons (Additional file 1: Table S1) between the three methods, and focus on the ability of CRAK-Velo to extract biological insights from complex data; technical specifications, including robustness, are discussed as Supplementary Material (Additional file 2: Figs. S1, S2). Figure 1 shows the application of CRAK-Velo to a 10x Multiome data set of human stem cells undergoing hematopoietic differentiation [19]. Panel (a) shows the visualisations of the data set obtained by three different methods: CRAK-Velo (left), UniTVelo (centre) and MultiVelo (right). This challenging benchmark was used in [19] (see Fig. 5 in the original publication) to demonstrate how integrating chromatin accessibility could correct spurious velocity estimates. We also observe that the purely RNA-based method UniTVelo cannot correctly identify the platelet state as a terminally differentiated state in this data set. CRAK-Velo is, however, also able to identify three separate terminal states (granulocytes, erythrocytes and platelets), while MultiVelo detects a biologically dubious flow from erythrocytes to granulocytes. This and other observations can be quantified using a cross-boundary direction (CBDir) analysis shown in Fig. 1b, which compares inferred transitions with expected transitions from biological knowledge. Figure 1b shows histograms of this metric for the three methods, clearly showing the superiority of the flows inferred by CRAK-Velo compared to the other methods. a Visualizations of velocity field showing inferred HSPC differentiation on the Umap embedding using three methods: CRAK-Velo (left), UniTVelo (center), and MultiVelo (right). b Histograms of cross-boundary direction (CBDir) analysis for the three methods, comparing inferred transitions with expected biological transitions. The black dashed line is the average CBDir per method. c Visualising cell type deconvolution performance for three genes in un/spliced phase space versus chromatin accessibility-unspliced space obtained by CRAK-Veelo or MultiVelo. The number below each subplot is the accuracy of a KNN cell-type classifier. d Histogram of the performance of the KNN cell type classifier across all genes for the phase space and accessibility-unspliced space inferred by the CRAK-Velo and MultiVelo. e Dotplot showing the biological processes enriched within the list of low entropy genes. f Log enrichment of binding at high-weighted regions (from CRAK-Velo) relative to what is expected by chance. g Regulatory dynamics of the textitKLF1 gene. Left is the region kinetic plot with double y axes that show the average accessibility crg for each region around KLF1 on the left and the average normalized unspliced reads u on the right one. The left bottom subplot shows the number of intersections of the binding sites of TF with the different regions (relative distance d of the regions from the gene is shown). The bottom right subplot is an arrow plot showing where the thickness is the regions contribution/weight wrg inferred by CRAK-Velo located at distance d from the gene We then considered the usefulness of the CRAK-Velo representation to distinguish cell types. Figure 1c shows the standard RNA velocity phase portraits (spliced vs unspliced read counts) for selected genes within the first row, while the second and third rows show unspliced reads vs chromatin effect for the same genes for CRAK-Velo and MultiVelo, respectively. Different cell types are coloured differently; visually, using chromatin accessibility shows a better separation of cell types compared to using only un/spliced reads. This can be quantified by evaluating the accuracy of a nearest neighbour classifier on these two-dimensional data sets, as proposed in [21] (see Methods). The accuracy achieved by the classifier is reported in the inset in the individual panels in Fig. 1c, showing a stronger performance of CRAK-Velo in all the cases. Figure 1d shows histograms of the accuracy of such a classifier on all genes for the different methods. Methods using chromatin information retrieve in general a higher accuracy, but while this effect is modest for MultiVelo, CRAK-Velo provides a much greater accuracy for a very large number of genes. While the previous results demonstrate the ability of CRAK-Velo to extract meaningful information at the cell level, we now examine its ability to recover meaningful biological information at the level of individual genes and gene-region interactions. In order to do so, we propose a novel ranking of genes based on their interactions with nearby accessible regions. We reason that, while individual genes may be associated with multiple nearby cis-regulatory regions, in any particular biological condition only few regions are likely to control the expression dynamics of a gene. We therefore use the distribution of inferred region-gene interaction weights to define an entropy metric which we use to rank genes (see Methods for mathematical details). This novel entropy metric explicitly integrates both expression and chromatin information, and can be used to identify key genes and pathways involved in the biological process under study. Figure 1e reports the results of a GO analysis summarizing functional enrichment among low-entropy genes. Several pathways involved in hematopoietic and lineage-specific biological processes Fig prominently, indicating that CRAK-Velo preferentially identifies genes whose function is consistent with the differentiation trajectories inferred earlier. Next, we investigated whether inferences from CRAK-Velo can also be used to detect signals about the regulators of these genes. To do so, we interrogated ChIP-ATLAS, a large database of ChIP-seq experiments [22], for transcription factor binding assayed in relevant cell types (more details in data availability section). Figure 1f shows the enrichment of transcription factor (TF) binding within the high-weighted chromatin regions inferred by CRAK-Velo, demonstrating that the model not only captures transcriptional dynamics but also recovers biologically meaningful regulatory elements. The significant log-enrichment values suggest that CRAK-Velo’s weighting scheme effectively prioritizes regions that correspond to genuine TF binding sites, confirming the model’s interpretability in linking accessibility features to gene regulation. The enriched transcription factors include well characterized TFs like IKZF1 (Ikaros), key in lymphoid lineage commitment and chromatin organization, and MEIS1 an essential for HSC self-renewal, once again highlighting the biological relevance of our findings. Finally, we illustrate the power of CRAK-Velo in deciphering the mechanisms of RNA expression kinetics on an example. Figure 1g further illustrates this for the KLF1 gene, a key regulator in erythroid differentiation. The plot shows in blue the average unspliced read counts as a function of pseudotime, while coloured traces represent the activity of individual chromatin regions, defined as the product of the open chromatin probability and the inferred regional weight. It shows a clear temporal coupling between chromatin accessibility and unspliced RNA abundance across pseudotime. For KLF1, it is evident that the increasing transcriptional trajectory is driven by an accessible region-proximal region with high accessibility that later declines, producing a transient expression pattern. Notice that the expression of the gene follows a delayed pattern with respect to the region’s activity, as implied by the dynamical system of CRAK-Velo. The final decrease in activity of the region is only partly compensated by the increase in activity in other more distal regions which are assigned lower weight by the model. The accompanying arrow plot visualises the weight associated with each regulatory region and its genomic distance from the transcription start site, revealing that accessible regions-proximal sites dominate regulatory influence during activation. We also consider the intersection of chromatin loci with TF binding sites; regions showing both high intersection counts and strong CRAK-Velo weights are likely to act as major control points for gene activation, as shown in the bar plot of Fig. 1g. Together, these analyses validate that CRAK-Velo integrates chromatin accessibility in a way that reflects real regulatory hierarchies and transcriptional control mechanisms. Next, we applied CRAK-Velo to another challenging data set, detailing mouse embryonic brain development at day 18 using 10X Multiome technology. While all three methods obtain broadly similar flows (Fig. 2a), a detailed observation reveals subtle but relevant differences, particularly regarding the dynamics at the boundary between the terminally differentiated states of the Upper and Deeper Layer. We use the PAGA graph approach [23] to summarize transitions in Fig. 2b. Both MultiVelo and UniTVelo recover a spurious flow from Upper Layer to Deeper Layers, while CRAK-Velo correctly recognises these cell states as independent terminally differentiated states.This is appreciated more clearly in Fig. 2c, where the inset shows in detail the bifurcating flow at the boundary of subplate and Upper and Deeper Layer. An analysis of the inferred latent time (Fig. 2a) also confirms this observation, as both MultiVelo and UniTVelo predict the deeper layers to develop after the upper layers, while CRAK-Velo assigns them both a similar pseudotime, consistent with biological knowledge. Limitations nevertheless remain, as all three methods were not able to identify ependymal cells as a terminal state. Next, we quantified the ability of the various methods to deconvolve cell types based on their inner representation. Figure 2d shows that chromatin-based approaches are more effective at detecting different cell types on three example genes, while Fig. 2e quantifies across all genes the ability to separate cell types, once again confirming a significant superiority in performance for CRAK-Velo. a UMAP visualization of the velocity fields (shown in streamlines) inferred by CRAK-Velo, MultiVelo, and UniTVelo. The cells are colored according to the inferred pseudotime by the corresponding method. b PAGA graph summarizing transitions between different cell types obtained from the velocity field embeddings. c Detailed view of the bifurcating flow at the boundary of the subplate and Upper and Deeper Layers. d Visualising cell type deconvolution performance for three genes in un/spliced phase space versus chromatin accessibility-unspliced space obtained by CRAK-Veelo or MultiVelo. The number below each subplot is the accuracy of a KNN cell-type classifier. e Histogram of the performance of the KNN cell type classifier across all genes for the phase space and accessibility-unspliced space inferred by the CRAK-Velo and MultiVelo. f Log enrichment of binding at high-weighted regions (from CRAK-Velo) relative to what is expected by chance. g Dotplot showing the biological processes enriched within the list of low entropy genes. h Regulatory dynamics of the Jag2 gene. Left is the region kinetic plot with double y axes that show the average accessibility crg for each region around Jag2 on the left and the average normalized unspliced reads u on the right one. The left bottom subplot shows the number of intersections of the binding sites of TF with the different regions (relative distance d of the regions from the gene is shown). The bottom right subplot is an arrow plot showing where the thickness is the region’s contribution/weight wrg inferred by CRAK-Velo located at distance d from the gene Figure 2f and h provide a second, independent demonstration of how CRAK-Velo recovers meaningful gene–regulatory structure, which shows the model’s reliability across distinct biological contexts. In Fig. 2f, the enrichment of TF binding within high-weighted chromatin regions implies that CRAK-Velo assigns weights to regulatory elements that are likely to be biologically active. The transcription factors enriched in the Fig collectively regulate neural stem cell maintenance, neurogenesis, and neuronal differentiation. For example, the most enriched TF, Pou3f2, is a central regulator of neuronal differentiation and identity, particularly important for cortical and hippocampal neuron development. The next most enriched factor, Sox2, helps maintain neural stem cell self-renewal and prevents premature differentiation. Additionally, Junb acts as an activity-responsive transcription factor that adjusts gene expression in response to neuronal stimulation. Using the entropy-based ranking introduced earlier, a similar theme emerges in the enrichment analysis of Fig. 2g, where low-entropy genes—those regulated by a concentrated subset of regions—cluster within pathways fundamental to cortical development, especially neuronal projection morphogenesis and synaptic organization. These specific enriched processes prove that CRAK-Velo reliably identifies genes whose regulatory profiles are tightly coupled to differentiation. Thus, CRAK-Velo can capture important genes that are relevant to the developmental trajectories captured by the velocity field. The regulatory dynamics of the Jag2 locus in Fig. 2h deepen this validation. Jag2, an important ligand in Notch-mediated neuronal specification, exhibits a tightly orchestrated coordination between chromatin accessibility and unspliced RNA accumulation. As pseudotime progresses, distinct regulatory regions modulate their contributions, with proximal elements dominating precisely when Jag2 expression rises. This structured sequence of region activity resembles what was previously observed for KLF1 in hematopoiesis: thus, the method detects differential regional engagement that not only accompanies transcriptional change but anticipates and shapes it. Conclusion Taken together, these analyses demonstrate that CRAK-Velo consistently uncovers a coherent regulatory architecture across diverse tissues, species, and developmental processes. By tightly integrating chromatin accessibility with RNA kinetic information, the method reliably recovers known biology, highlights lineage-specific regulatory programs, and resolves the temporal structure of gene activation with fine granularity—all while maintaining strong performance at the systems level. Across benchmark datasets, CRAK-Velo accurately reconstructs complex dynamic flows and outperforms existing approaches in distinguishing cell types, underscoring the practical utility of its unified representation. In particular, compared to the recent method MultiVelo, CRAK-Velo offers both a simpler (and faster, see Supplementary Table 1) approach and a clear interpretation of how individual chromatin regions shape transcriptional trajectories, providing mechanistic insights that can be translated into testable hypotheses. As such, CRAK-Velo offers an interpretable framework for extracting dynamic regulatory information from high-throughput single-cell multi-omic data. Methods Data preprocessing For the HSPC dataset, the same cells and their corresponding annotations in MultiVelo are used. For the embryonic E18 mouse brain we followed the cell annotations of MultiVelo. All cells that belong to the cell types list chosen by MultiVelo (‘RG, Astro, OPC’, ‘IPC’, ‘V-SVZ’, ‘Upper Layer’, ‘Deeper Layer’, ‘Ependymal cells’, ‘Subplate’) are kept for downstream analysis. All the scRNA-seq part of the multi-omics datasets are preprocessed using the same approach as UniTVelo: The highly variable genes (HVGs), which contribute to cell-cell variation, were selected (default 2,000). Further selected informative genes with similar settings in scVelo. For the scATAC-seq data part we chose to keep regions that are captured in more than 800 cells in the HSPC dataset and 400 cells for the embryonic E18 mouse brain dataset. Then, regions that are within 104 upstream and downstream from the gene starting site are kept. The final HSPC dataset is composed of 11605 cells, 2000 genes, and 3939 regions. For the embryonic E18 mouse brain dataset, we ended up with 3365 cells, 2000 genes, and 4002 regions. For the human cerebral cortex, we used the same scRNA-seq dataset obtained by the MUltiVelo preprocessing and kept the regions within a distance of 10kb from the genes. The final input data has the dimensions of 4693 cells, 954 genes, and 844 regions. Smoothing scATAC-seq data: CisTopic is used to smooth the binary and sparse scATAC-seq data. We implemented our version of cisTopic in python (the link to our cisTopic implementation is provided in the code availability section). CisTopic extracts a latent representation of dimension T for each cell n and region r: We computed the open probability of a region r in a cell n (ϕrn) using the cisTopic representation: CisTopic is implemented with a Gibbs sampler. The number of samples is 3000 with thinning every 10 samples to avoid sample correlations. The extracted latent space has a dimension T=30, and the sample mean was used to estimate λn and ψr. Reconciling data-driven and chromatin-based transcription rates in velocity estimation UniTVelo is based on a data-driven, parametric estimate of the amount of spliced reads as a function of time (using flexible basis function regression). Following the default workflow of UniTVelo, the spliced read of a gene is modelled explicitly using RBF basis function regression as [12]: Applying time derivative on Eq. 5, we obtain u^g′(t) written in terms of s^g(t) and its time derivative s^g′(t) : Alternatively, we can use scATAC-seq to define the production rate of unspliced reads as follows. The scATAc-seq data represent the accessibility information of the cis-regulatory regions across the genome. Fixing a window around a specific gene g we obtain a set of regions of size Rg. Using this set of regions, we define the ATAC-derived transcription rate of gene g, cg, as follows: Time tn is the inferred unified pseudotime for a cell n. The dynamic variable crg(tn) represents the accessibility of a region r associated with a gene g in cell n which has a pseudotime tn. Periodically the accessibility function is constructed crg(t) by reordering the cell’s accessibility according to the inferred pseudotime of the cells (more details in the parameter inference section). Different regions around a specific gene contribute with different weights to the regulation of this gene. A region r around gene g regulates this gene with a weight wrg that need to be inferred. In all the results shown Rg=104bp. Therefore, we can define the ATAC-derived dynamics of unspliced read counts as We end up with two forms of the derivative of unspliced reads. The unspliced derivative that is derived from the proposed paramtrised RBF (Eq. 6) while the other form encodes the scATAC-seq data information within it (Eq. 11). In the next section, we show how to reconcile it through a bespoke likelihood function. Parameter inference of CRAK-Velo In order to reveal the accessible chromatin region and splicing kinetics we need to infer gene-specific parameters θg and cell-specific time points tng. The set of parameters, θg, are used to define the shape of the mean function of both un/spliced reads, where θg is: Let xgi=(ugi,sgi) where ugi and sgi are the normalized observation of un/spliced counts for a particular gene and i∈(1,...,N) such that N is the number of cells. Similarly, let x^gi=(u^gi,sgi^) be the model’s estimation of un/spliced counts of that gene. The objective function used to find the optimal parameter settings of phase portrait and the accessible region kinetics is the negative log-likelihood that includes multi-omics information of the un/spliced reads and scATAC-seq smoothed data: where k = 0.5 to tune the effect of noise coming from the scATAC-seq data on the inference procedure, and σg is defined in UniTVelo as the standard deviation of the Euclidean distance residuals under the assumptions that the residuals are normally distributed N(0,σg2). We used the same optimization procedure in UniTVelo [12] to fit the model parameters by iteratively minimizing the negative log-likelihood using the Gradient Descent algorithm: Given x^i(θg|tn), parameterized by unified cell time, the gradients of the objective function are computed and θg is updated. Given x^i(θg|tn), parameterized by unified cell time, the gradients of the objective function are computed and θg is updated. For a fixed θg gene-cell specific time tngis reassigned by minimizing the Eculidean distance between xi and updated x^i. For a fixed θg gene-cell specific time tngis reassigned by minimizing the Eculidean distance between xi and updated x^i. The unified pseudotime of the cells tn is computed. The unified pseudotime of the cells tn is computed. The variable ϕr for a region r is ordered according to the unified pseudotime. Note that the loss of the genes that has no regions within the chosen window is the same as the UniTVelo one. The variable ϕr for a region r is ordered according to the unified pseudotime. Note that the loss of the genes that has no regions within the chosen window is the same as the UniTVelo one. We used non-informative initialization as defined in UniTVelo with initial value tn=0.5. We run for 10000 epochs for each dataset. Choosing k : For the first two datasets, we choose k to be: such that G and R are the number of genes and regions respectively, considered in the dataset. For the third dataset where the G>R and since scATAC-seq is noisier than scRNA-seq, we calculated k as follows: where ngc and nrc are the total number of reads for a cell c in scRNA-seq and scATAC-seq data, respectively. E[nRNAc] and E[nATACc] are their corresponding averages across cells. Evaluating model performance We used cross boundary direction on the HSPC dataset to evaluate the performance of CRAK-Velo, UniTVelo, and Multivelo in capturing meaningful differentiation trajectories of the hematopoietic stem cells. We used K-nearest neighbors(KNN) classifier to evaluate CRAK-Velo and MultiVelo performances in deconvoluting cell-types once the inferred accessibility information c is used. Also, we compare it with the observations (un/spliced reads) to quantify the improvement the accessibility can add to the observed data. The number of neighbors used in KNN is 5, and the metric is Euclidean. We trained the KNN on 70% of the cells in each dataset and validated on the rest. In the histograms, we show the common genes between CRAK-Velo and MultiVelo. The PAGA graph in Fig. 2b is obtained by using latent time as a time prior and velocity field embeddings inferred by the three methods. All the possible edges are obtained through PAGA by setting threshold = 0 and the minimum spanning tree option to False. Low entropy genes To identify genes with preferentially utilized chromatin regions, we quantified the entropy of accessibility weights across a window (104bp) for each gene. For every gene, the distribution of accessibility weights within its associated regions was compared against the entropy of a uniform distribution over the same number of regions. Because the number of accessible regions is gene-specific, the uniform entropy was computed separately for each gene as a baseline. We then calculated the difference between the observed entropy and the corresponding uniform entropy. Genes for which this difference ranked in the top 20% were classified as low-entropy genes. A low entropy score indicates that accessibility is not evenly distributed across a gene’s regions but is instead concentrated in a subset of preferred binding sites. This property highlights regulatory specificity: while some genes display broadly distributed accessibility across many potential sites, low-entropy genes exhibit stronger localization of chromatin opening, which may correspond to key regulatory elements driving transcriptional dynamics. Region kinetics We monitored the accessibility kinetics of regions around a chosen gene by dividing the timeline (between 0 and 1) into intervals of width 0.05 and computing the average accessibility crg per interval. We showed in parallel the change in the unspliced reads with cell pseudotime by also averaging the normalised unspliced reads over an interval of 0.05. The kinetic plots have double y axes that show the accessibility on the left and the normalised ug on the right. Computing enrichment of transcription factors at low-entropy genes For each transcription factor (TF), we quantified enrichment by comparing its global binding probability across all accessible regions with its binding probability restricted to the most accessible sites. Let R denote the set of all accessible regions in the dataset, and RTF⊆R is the subset predicted to be bound by a given TF. For each gene g, let wmaxg⊆R represent its maximally weighted accessible site as inferred by CRAK-Velo, and define M=⋃gwmaxg as the union across all such genes. The global binding probability of the TF is given by: while the binding probability is: Finally, TF enrichment is quantified as the ratio: The log (E(TF))>0 indicates that the TF is disproportionately associated with the maximally accessible regions of low-entropy genes, suggesting potential regulatory relevance in driving their transcriptional dynamics. Supplementary information Additional file 1. Evaluating the performance of CRAK-Velo on the human cerebral cortex 10x multiome data set and its running time. Additional file 2. Sensitivity analysis of CRAK-Velo to the number of topics and the window size around the gene. Acknowledgements We thank Fabian Theis and Weixu Wang for useful discussions towards the beginning of this project, and Maria Ayub for help in interpreting the biological results. Peer review information Tim Sands was the primary editor of this article and managed its editorial process and peer review in collaboration with the rest of the editorial team. The peer-review history is available in the online version of this article. Code availability CRAK-Velo code is publicly available on zenodo [24] (DOI: https://zenodo.org/records/19247214) and github [25] (https://github.com/StatBiomed/CRAK-Velo.git). Detailed workflows to reproduce Figs and results in this paper are written as Jupyter notebook in the repository. Our implementation of cisTopic is available on https://github.com/Nour899/cisTopic.git. Authors' contributions NeK GS, and YH conceived and designed the research. NeK and MG implemented the research code. NeK, MG,IK, FC, YH, and GS and analyzed the results. NeK and GS wrote the paper. All authors read and approved the final manuscript. Funding The authors acknowledge support from the Italian Association for Cancer Research (AIRC) under grant IG 27631. GS acknowledges co-funding from Next Generation EU, in the context of the National Recovery and Resilience Plan, Investment PE1 - Project FAIR “Future Artificial Intelligence Research”. This resource was co-financed by the Next Generation EU [DM 1555 del 11.10.22]. Data availability 10x embryonic mouse brain dataset can be accessed at the 10x website at https://www.10xgenomics.com/resources/datasets/fresh-embryonic-e-18-mouse-brain-5-k-1-standard-1-0-0. The processed files of the newly sequenced 10x Multiome HSPC samples are available at the GEO (GSE209878). The instructions to download the preprocessed version of the data by MultiVelo utilized in our analysis are available in the Multivelo tutorials. Human brain multi-ome dataset 43 can be found at the GEO GSE162170 and the authors’ GitHub page. The Chip-seq data for the TF binding sites coordinates is obtained from CHIPseq Atlas. To obtain different sites for the three datasets we chose “TF and others” from the “Track type class” list. For the HSPC dataset we used hg38 database. We chose from the “Cell type class” list the option “Blood”, and chose relevant cell types from the “cell type” list (the precise cell types follows). Regarding the human cerebral cortex, we also used hg38 with cell type class “Neural”, and relevant cell types. Finally for the 10 embryonic mouse brain dataset, we used the mm10 database with cell type class “Neural”, and relevant cell-specific types. Cell types for HSPC: hematopoietic cells, hematopoietic stem cells, peripheral blood mononuclear cells, peripheral blood mononuclear stem cells, iPSC-derived hematopoietic cells, erythroid progenitors, and CD34 hematopoietic stem cells. Cell types for human cerebral cortex: Neural stems cells, fetal brain, fetal neural cells, fetal neural progenitor cells, neural progenitor cells, and fetal astrocytes, astrocytes. Cell types for embryonic mouse brain: Neural stems cells, induced neural progenitors, brain cortex, cortical neuron, oligodendrocyte progenitor, cortical oligodendrocyte progenitor, astrocytes. CRAK-Velo code is publicly available on zenodo (DOI: https://zenodo.org/records/19247214) and github (https://github.com/StatBiomed/CRAK-Velo.git). Detailed workflows to reproduce Figs and results in this paper are written as Jupyter notebook in the repository. Our implementation of cisTopic is available on https://github.com/Nour899/cisTopic.git. Declarations Ethics approval and consent to participate Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Yuanhua Huang, Email: yuanhua@hku.hk. Guido Sanguinetti, Email: gsanguin@sissa.it. References Associated Data Supplementary Materials Additional file 1. Evaluating the performance of CRAK-Velo on the human cerebral cortex 10x multiome data set and its running time. Additional file 2. Sensitivity analysis of CRAK-Velo to the number of topics and the window size around the gene. Data Availability Statement 10x embryonic mouse brain dataset can be accessed at the 10x website at https://www.10xgenomics.com/resources/datasets/fresh-embryonic-e-18-mouse-brain-5-k-1-standard-1-0-0. The processed files of the newly sequenced 10x Multiome HSPC samples are available at the GEO (GSE209878). The instructions to download the preprocessed version of the data by MultiVelo utilized in our analysis are available in the Multivelo tutorials. Human brain multi-ome dataset 43 can be found at the GEO GSE162170 and the authors’ GitHub page. The Chip-seq data for the TF binding sites coordinates is obtained from CHIPseq Atlas. To obtain different sites for the three datasets we chose “TF and others” from the “Track type class” list. For the HSPC dataset we used hg38 database. We chose from the “Cell type class” list the option “Blood”, and chose relevant cell types from the “cell type” list (the precise cell types follows). Regarding the human cerebral cortex, we also used hg38 with cell type class “Neural”, and relevant cell types. Finally for the 10 embryonic mouse brain dataset, we used the mm10 database with cell type class “Neural”, and relevant cell-specific types. Cell types for HSPC: hematopoietic cells, hematopoietic stem cells, peripheral blood mononuclear cells, peripheral blood mononuclear stem cells, iPSC-derived hematopoietic cells, erythroid progenitors, and CD34 hematopoietic stem cells. Cell types for human cerebral cortex: Neural stems cells, fetal brain, fetal neural cells, fetal neural progenitor cells, neural progenitor cells, and fetal astrocytes, astrocytes. Cell types for embryonic mouse brain: Neural stems cells, induced neural progenitors, brain cortex, cortical neuron, oligodendrocyte progenitor, cortical oligodendrocyte progenitor, astrocytes. CRAK-Velo code is publicly available on zenodo (DOI: https://zenodo.org/records/19247214) and github (https://github.com/StatBiomed/CRAK-Velo.git). Detailed workflows to reproduce Figs and results in this paper are written as Jupyter notebook in the repository. Our implementation of cisTopic is available on https://github.com/Nour899/cisTopic.git.