pepr-hal-data / bio2flow_json /CellTrajectoryInferencebasedonSchrdingerProblemandaMechanisticModelofStochasticGeneExpression.grobid.tei.xml.json
Download bio2flow_json/CellTrajectoryInferencebasedonSchrdingerProblemandaMechanisticModelofStochasticGeneExpression.grobid.tei.xml.json from HackatonSanteNumerique/pepr-hal-data: direct link, hf CLI and curl.
- Browser
- Download file 61.6 kB
-
https://huggingface.co/datasets/HackatonSanteNumerique/pepr-hal-data/resolve/main/bio2flow_json/CellTrajectoryInferencebasedonSchrdingerProblemandaMechanisticModelofStochasticGeneExpression.grobid.tei.xml.json
- Command line
-
hf download hf://datasets/HackatonSanteNumerique/pepr-hal-data/bio2flow_json/CellTrajectoryInferencebasedonSchrdingerProblemandaMechanisticModelofStochasticGeneExpression.grobid.tei.xml.json
-
curl -L -o CellTrajectoryInferencebasedonSchrdingerProblemandaMechanisticModelofStochasticGeneExpression.grobid.tei.xml.json https://huggingface.co/datasets/HackatonSanteNumerique/pepr-hal-data/resolve/main/bio2flow_json/CellTrajectoryInferencebasedonSchrdingerProblemandaMechanisticModelofStochasticGeneExpression.grobid.tei.xml.json
61.6 kB
| { | |
| "num_paragraphs": 106, | |
| "paragraphs": [ | |
| { | |
| "index": 0, | |
| "text": "Cellular differentiation is the biological process that leads a cell to opt for a particular cellular identity. Recently, single-cell RNA-sequencing has enabled the simultaneous measurement of gene expression levels at specific times for a large number of individual cells and a large number of genes. Repeating such measurements at different time points gives then access to the temporal variation, or transport, of a distribution on a gene expression space. The whole temporal trajectory of distributions thus characterizes the differentiation process at population level, but trajectories of individual cells are still out of reach since most measurement techniques are destructive.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 1, | |
| "text": "The optimal transport theory that has been used so far to infer cellular differentiation trajectories from time-stamped single-cell RNA-seq data involves solving the so-called Schrödinger problem in its most common version. This implies assuming that cells move, in the gene expression space, by diffusion. Yet, real gene dynamics are much more complex.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 2, | |
| "text": "In the present work, we assume that mRNA dynamics are characterized by brief and important production of RNA, with long periods of inactivity in between, and consider the so-called Bursty model of gene dynamics. We use this model to define a reference process for the Schrödinger problem. By comparing the solutions of the Schrödinger problems with a Diffusive and a Bursty reference process, under different conditions, we show that the Bursty model provides a better approximation of the underlying gene dynamics than the standard Diffusive process when inferring cell trajectories.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 3, | |
| "text": "While two cells in a multicellular organism may possess the same genome, they can express their genes differently, leading to distinct cellular outcomes. Cellular differentiation is the biological process that leads a cell to opt for a particular cellular identity [1]. In contemporary biology, the cellular identity is defined by the expression of cell markers, mostly surface proteins.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 4, | |
| "text": "Recently, single-cell RNA sequencing (scRNA-seq) has enabled the simultaneous measurement of gene expression levels for up to millions of individual cells and tens of thousand of genes [2,3,4]. Thanks to this technique, a cell can be represented in a high-dimensional space, where each dimension corresponds to the expression level of a given gene, so the space dimension equals the number of genes. Repeating such measurements at different time points gives access to the temporal variation, or transport, of a distribution on a gene expression space. The entire temporal trajectory of the distribution characterizes the differentiation process at population level, but trajectories of individual cells are still out of reach since measurement process is destructive [5]. Hence the goal to develop tools to infer trajectories between measurements to reconstruct the differentiation process has become crucial. Since cells are destroyed in the data acquisition process, inference of cell trajectories allows to determine potential cell fate under the assumption that the cell sample used at each measurement time is representative of the whole cell population.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 5, | |
| "text": "In practice, gene expression is measured either as the number of mRNA molecules or the level of protein expression. Hence, the above-mentioned gene space is either a mRNA counts space or a protein level space. While the differentiation process is fundamentally a continuous process, most of the experimental measurements provide access only to discrete temporal distributions. In order to study a differentiation process and more specifically to infer differentiation pathways, many works focused on the reconstruction of the continuous transcriptomics evolution based on such distributions.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 6, | |
| "text": "Schiebinger et al. [5] were pioneers in using the optimal transport (OT) theory to infer such trajectories from time-stamped transcriptomics data, introduced in the Waddington Optimal Transport (WOT) algorithm. WOT enables the reconstruction of a \"timeline\" between two cell distributions measured at different times. WOT solves the OT problem in a standard way: by minimizing a cost function defined as the squared Euclidean distance between cells in a PCA-reduced space.", | |
| "entities": [ | |
| { | |
| "text": "Optimal Transport", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 176, | |
| "end": 193, | |
| "confidence": 0.9722 | |
| }, | |
| { | |
| "text": "WOT", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 195, | |
| "end": 198, | |
| "confidence": 0.9942 | |
| }, | |
| { | |
| "text": "WOT", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 318, | |
| "end": 321, | |
| "confidence": 0.8696 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 7, | |
| "text": "Since then, several works proposed methods for cell trajectory inference inspired by OT theory [6,7,8,9,10,11,12,13,14,15,16,17]. Authors of OT-based studies chose to work with quadratic OT, which consists in using the squared Euclidean distance as a cost in the OT problem (see Section Coupling between time points). More specifically, they used the so-called entropic approximation of OT-which is equivalent to solving the Schrödinger problem-in order to use the efficient Sinkhorn algorithm [18,19]. The Schrödinger problem consists in finding the stochastic process that is the closest to a given reference process while fitting the observations. Originally, the aim of the Schrödinger problem was to interpolate between cell distributions observed at discrete timepoints in order to reconstruct the full time-varying distribution. The mathematical theory supporting this problem ensures that this optimal interpolation is associated to an optimal coupling corresponding to the transition probability kernel between each pair of distributions of the underlying stochastic process. Thus, the destructive scRNA-seq measurements are perfectly aligned with the Schrödinger problem: the inferred coupling characterizes the optimal process describing the observed data under the hypothesis of the Schrödinger problem.", | |
| "entities": [ | |
| { | |
| "text": "Sinkhorn", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 475, | |
| "end": 483, | |
| "confidence": 1.0 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 8, | |
| "text": "The reference process used in the Schrödinger problem is in general a Brownian motion, hence it is implicitly assumed that cell trajectories are close to diffusion. In some cases, distances different from the square Euclidean distance have been considered, mostly for taking into account additional information like lineage information [11,20] or coexpression between genes [21].", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 9, | |
| "text": "Some methods nevertheless used a different cost function when solving the OT problem. For instance, Liu et al. [17] criticized the use of Euclidean distance when applying OT theory and proposed an implementation of a learned cost function in the OT problem based on the Sinkhorn algorithm [18]. However, they do not use OT theory to infer cell trajectories, instead they focus on alignment of scRNA-seq datasets. Additionally, the learned cost is non-interpretable, a property that is shared by computational methods based on machine learning. In order to reduce the computational cost and gain interpretability, scEGOT [13] combines OT with a cell state graph. By adding Gaussian mixture distribution on each cell's dots, scEGOT allows to consider cell clusters instead of individual cells. The Gaussian mixture makes the results interpretable, however their objectives are clearly different from ours because the authors use Euclidean distance between Gaussian distributions to infer trajectories, relying once again on a standard approach.", | |
| "entities": [ | |
| { | |
| "text": "Sinkhorn", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 270, | |
| "end": 278, | |
| "confidence": 1.0 | |
| }, | |
| { | |
| "text": "scEGOT", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 613, | |
| "end": 619, | |
| "confidence": 1.0 | |
| }, | |
| { | |
| "text": "scEGOT", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 723, | |
| "end": 729, | |
| "confidence": 1.0 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 10, | |
| "text": "The main idea of the present paper is to assert that in order to reconstruct cell differentiation trajectories, a more realistic, quantitative model of stochastic gene expression should be considered as reference process. We focus on solving the Schrödinger problem while using an alternative reference process that would be closer to realistic gene dynamics. Indeed, the Diffusive process generates unrealistic dynamics for proteins and mRNA [22]. The Schrödinger problem is not seen here as an approximation of a transport problem, but as a way to identify a process close to an a priori quantitative model that would fit the observations.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 11, | |
| "text": "When examined at the single-cell level, gene expression is undoubtedly a stochastic process [23,24,25,26,27,28]. The simplest model that can correctly reproduce experimental data distribution is the well-known two-state model of gene expression [29,30,31], a refinement of the model introduced in 1991 by Ko [32]. In this model, a gene is described by its promoter which can be either active or inactive-possibly representing a transcription complex being \"bound\" or \"unbound\" although it may be more complicated [33]-with mRNA being transcribed only during the active periods.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 12, | |
| "text": "Genes may mechanistically influence the activation rate of other genes, thus creating a Gene Regulatory Network (GRN). We consider an extension of the two-state model for several genes, that takes into account the GRN structure. For certain parameter values, mRNA dynamics are characterized by brief and important production of RNA, with long periods of inactivity in between: this regime corresponds to the widely observed phenomenon called transcriptional bursting [29]. Herbach et al. [34] introduced a mathematical framework for modeling GRN driven by transcriptional bursting, which we call the Bursty model here. This model has been shown to reproduce realistic gene dynamics at the single-cell level [35,36] and will provide our reference process.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 13, | |
| "text": "In this work we compare the performances of a Diffusive process, associated to standard OT theory, and a reference process based on the Bursty model to solve the Schrödinger problem, when initial and final distributions are based on realistic, simulated data. For the sake of clarity, in the following we refer to our approach as the Bursty model-based inference method. We first explore the case of a single gene, then of a specific two-gene GRN, the toggle switch, and finally more complex three-gene GRNs. We focus on the influence of some indicators, namely the time difference between the measurements of the two distributions and the number of cells in the distributions. We show that using the Bursty model as a reference process provides a better solution to the Schrödinger problem for all tested networks. Additionally, we show that the Bursty model-based inference method performs better than a Diffusion-based inference approach on a dataset made of mRNA counts for 41 genes.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 14, | |
| "text": "We consider two distributions of protein counts obtained from a fixed number of cells, simulated from the Bursty model (see Section Gene dynamics) and measured respectively at times t 1 and t 2 (t 2 > t 1 ). Noticeably, we focus only on protein counts since protein dynamics, contrary to mRNA dynamics, allow to preserve the markovian property in the Bursty model therefore providing a rigorous mathematical framework. We solve the associated Schrödinger problem for several reference processes, including the Bursty process and the standard Diffusive process. Solutions of the Schrödinger problem are compared using an entropy function. If the two processes are similarly able to correctly infer trajectories, the entropy equals or nears zero.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 15, | |
| "text": "When the final distribution is measured a \"long\" time after the initial distribution, the Bursty model has reached a so-called stationary state, and both initial and final distributions can be considered independent. In this case, trajectory inference should depend less on the selected reference coupling. Consequently, we focus our study on distributions in transient states, meaning the time difference t 2 -t 1 is small enough and t 1 is close to the initialisation of the simulations, as represented in Figure 12. Details are provided in Section Methods.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 16, | |
| "text": "Noticeably, since we are working with data that we generated, so-called simulated datasets, it is possible and efficient to define a ground truth. More precisely, it is convenient to use an identity-based ground truth, defined such that cell i at t 1 corresponds to cell i at t 2 . However, when using experimental single cell datasets, no ground truth is available and only marginal observations can be used to infer the cell differentiation process. In this case, no reference exists to validate the result of any method.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 17, | |
| "text": "We first considered the case of a single gene with no feedback, as a toy model to illustrate the influence of the underlying model of gene dynamics on trajectory inference.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 18, | |
| "text": "Figure 1 shows the solutions of the Schrödinger problem computed with different approximated processes: a Diffusive model, D sch (Figure 1a), the non-reduced Bursty model, B sch (Figure 1b), and the reduced Bursty model, RB sch (Figure 1c). The ground truth of the simulated dataset (Figure 1d) is a coupling of cells with the same index (cell i in the initial distribution is cell i in the final distribution). We observe that RB sch yields the closest result to the ground truth of the simulated dataset (Figure 1c), in the form of a very accurate transition matrix, while D sch yields the furthest (Figure 1a). A darker diagonal is observed for B sch (Figure 1b), meaning that this process better approximates the ground truth than D sch . Noticeably, a limitation of the representation in Figure 1 is that the process could be not expressed (light color) and yet be relevant. For example, if two cells are close in protein space at both t 1 and t 2 , then the process between those cells could be mixed up but relevant.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 19, | |
| "text": "Figure 2 presents the entropy between D sch and RB sch , denoted by H D,R , and the entropy between B sch and RB sch , denoted by H B,R (see Section Performances of the models). The curve H R also appears in Figure 2, it serves as a control by computing the entropy between the reference coupling and its approximation. The initial distribution is measured at time t 1 = 0.1 h, and the final distribution is computed at various times t 2 in the range [0.2; 0.5] h.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 20, | |
| "text": "We observe that quantities H D,R and H B,R decrease as t 2 -t 1 increases and H R remains constant. We also observe that H D,R is always higher than the entropy between B sch and RB sch , and remains much higher than H R even for large values of t 2 -t 1 (here 0.4 h). Entropy H B,R is always larger than H R , yet it is much smaller than H D,R , and for t 2 -t 1 = 0.4 h there is no difference between H B,R and H R . Both the exact temporal distribution of the process and its approximation provide comparable results. 1. Other parameters are detailed in Table 1.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 21, | |
| "text": "The study of this toy model highlights that using the exact temporal distribution of the reference process to solve the Schrödinger problem provides much better results than using the approximation of a gene dynamics model. However, it is a complex, most of the time impossible, task to find the analytical expression of this distribution for a GRN comprising at least two connected genes. Consequently, for a more relevant GRN, we will focus on the estimations provided by the reference process of the non-reduced Bursty model and the Diffusive process.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 22, | |
| "text": "We focused on a two-gene GRN, called 'toggle switch', introduced in Section Example of GRN: the Toggle Switch and illustrated in Figure 6b (blue GRN). This specific GRN consists of two self-activating genes that inhibit each other.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 23, | |
| "text": "We first investigated the influence of the time between measurements of the initial and the final distributions (i.e. t 2 -t 1 ) on the ability of two models, the Bursty and the Diffusive models, to correctly transport cell distribution.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 24, | |
| "text": "Figure 3a shows values of entropies H D,B and H B as functions of the time difference t 2 -t 1 . The H D,B curve is always above the H B curve and remains approximately constant and independent of the time difference. The entropy between the two processes is the same for all t 2 -t 1 values: whatever the time difference, the Diffusive process infers trajectories with the same uncertainty. For the time difference t 2 -t 1 = 0.2h, the entropy H D,B has a larger standard deviation, due to one extreme value.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 25, | |
| "text": "Although it is more relevant to focus on the transient state (Figure 3a), it may be informative to question what happens in the stationary state. In this case, cells may either remain in their basin of attraction or switch to the other basin of attraction. The basin of attraction is a notion related to deterministic dynamics, consequently in the current framework where we consider stochastic models of gene dynamics it refers to the basin of attraction of the coarse approximation in the form of a system of ordinary differential equations of a piecewise-deterministic Markov process [37]. Results in the stationary state are illustrated in Supplementary File (Section 3), they highlight conclusions similar to the ones obtained for the transient state. Although the Diffusive model-based inference method is less accurate than the Bursty model-based inference method for inferring trajectories, its ability increases when the final distribution becomes independent of the initial one (later time measurements).", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 26, | |
| "text": "Figure 3b illustrates the influence of the number of cells N considered when measuring both initial and final distributions in transient state. To that aim, the entropy has been normalized by the number of cells to avoid a bias due to the size of the collections in the computation of the entropy: the computation of the entropy (8) explicitly depends on the number of cells N . First, we observed that H B /N does not depend on the number of cells. Second, H D,B /N decreases as the number of cells increases. We observed (not shown) that the Diffusive process tends to couple a small sample of cells from the initial collection to all other cells in the final collection.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 27, | |
| "text": "Both numerical investigations on the dependence of H D,B and H B on the length of 1.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 28, | |
| "text": "the time interval t 2 -t 1 and the number of cells N highlighted a poor ability of the Diffusive model to reconstruct cell trajectories when the underlying model is the Bursty model of gene dynamics. These results are based on the use of the entropy to measure the differences between solutions of the Schrödinger problem. We then questioned the relevance of the entropy criterion and the parameter values of the Bursty model in reaching this conclusion.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 29, | |
| "text": "In Figure 4, we compared the performance of different indicators when evaluating the difference between the Diffusive and the Bursty process. We observe that for all indicators, the curves are all strictly positive and increasing. So, despite quantitative differences, all indicators yielded the same qualitative conclusion: the Bursty process provides a better trajectory inference than the Diffusive process.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 30, | |
| "text": "To challenge the performance of the B ref process, we generated the two distributions with a given GRN and used a different GRN for the reference process (mostly the same structure, but different interaction parameter values), that is a different Bursty model. The use of simulated data provides full knowledge of the structure of the GRN and of its parameter values, however when dealing with experimental data the GRN is not known and can be at best inferred from the data, without any way to validate per se either the structure of the GRN or the parameter values. Hence, using a GRN that does not correspond to the data used to generate the initial and final distributions challenges the method and allows to test the robustness of the approach. In practice, we computed the initial distribution µ and the final distribution ν using the default parameter values of the toggle switch GRN, but we solved the Schrödinger problem with a different B ref process (different gene interaction matrix).", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 31, | |
| "text": "Figure 5a highlights that for all the two-gene-based Bursty models (listed in Figure 5c), the curves are very close. The H D,B curves are all higher than the H B curves, indicating that any Bursty process, even one with parameter values different from the ones 1.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 32, | |
| "text": "corresponding to the data, provides a better approximation than the Diffusive process. In Figure 5b, one observes that even though H B curves are nearly indistinguishable at early times they become separated as t 2 -t 1 increases. We observed differences when parameter values of the GRN associated with self-activation were reduced (purple and yellow curves and GRN).", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 33, | |
| "text": "Overall, the Bursty model provides a framework to cell trajectory inference that performs better than standard diffusion.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 34, | |
| "text": "We studied numerically the sensitivity of our method to hyperparameters (number of simulations n, standard deviation ε in (6), precision and truncation thresholds used to prevent zero values in the Sinkhorn algorithm) and to parameter values of the GRN, specifically focusing on degradation rates and basal activity levels (Supplementary File). No significant impact was observed on the entropy H D,B , pointing towards a weak sensitivity of the results on parameterization. 1.", | |
| "entities": [ | |
| { | |
| "text": "Sinkhorn", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 198, | |
| "end": 206, | |
| "confidence": 0.9819 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 35, | |
| "text": "As illustrated in Figure 11, the toggle switch model of gene dynamics exhibits what can be considered as specific temporal dynamics. We subsequently challenged the Bursty process on other GRNs, first with various two-gene GRNs and then on three-gene GRNs. Results discussed hereafter and illustrated in Figures 6 and 7 have been obtained using various GRNs to compute both the distributions (µ and ν) and the reference coupling. This contrasts with the results shown in Figure 5, where only the reference coupling was modified by the GRNs, not the data.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 36, | |
| "text": "Figure 6 shows the entropy H D,B for different two-gene GRNs. For each GRN, the distributions and the Bursty process were computed from the corresponding Bursty model. The blue curve represents H D,B values computed for the toggle switch. We observed that, whatever the two-gene GRN, dynamics of H D,B were similar, quantitatively and qualitatively: it was mostly constant as the time difference t 2 -t 1 increased, and the variability was higher for small time differences.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 37, | |
| "text": "Figure 7 shows the entropy H D,B for two 3-gene GRNs. In the first one, gene 1 activates gene 2 which activates gene 3 (Figure 7a); in the second one, the so-called repressilator model, gene 1 inhibits gene 2, gene 2 inhibits gene 3, and gene 3 inhibits gene 1 (Figure 7b). Parameter values are the same than for the toggle switch (Table 1).", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 38, | |
| "text": "Figure 7a shows a similar range of variations for H D,B and H B as observed for the two-gene GRNs. Variations about the mean were, however, more important than for the two-gene GRNs, as highlighted by error bars and minimum and maximum values (dashed lines). The H D,B curve remains well above the H B curve, similarly to what has been obtained for two-gene GRNs. The same results held for the Repressilator model (Figure 7b), a standard three-gene GRN. Whatever the time difference, the mean of H D,B was always higher than H B . For both GRNs, when t 2 -t 1 was small, the standard deviation of H D,B was larger and the minimum and maximum values were widely separated. The two three-gene GRNs considered generated more differences between the Diffusive and the Bursty processes than the toggle switch model or any 2-gene GRN, as well as more variability.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 39, | |
| "text": "In conclusion, for any two-gene or three-gene GRN, whatever the gene-gene interactions considered, the Bursty process always performed better than the standard Diffu- 1.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 40, | |
| "text": "sive process in correctly transporting the distribution at time t 1 into the distribution at time t 2 .", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 41, | |
| "text": "Application to experimental data from Semrau et al. [38] While previous sections have been devoted to validating the Bursty-based inference method on small GRN and simulated data, here we apply the method on an experimental dataset. We use a time-stamped in vitro dataset from Semrau et al. [38] made of mRNA counts obtained from scRNA-seq on mouse embryonic stem cells. A GRN in the form of a bursty model has been inferred from these data by Ventre et al. [36], therefore we use this GRN and the parameter values obtained during the inference procedure to compute our reference process B ref .", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 42, | |
| "text": "Since all previous investigations were performed on protein data, we first check whether or not the Bursty model-based inference method would also perform well on simulated mRNA data. Similarly to what we did in Figure 3a., considering the toggleswitch GRN, we investigated the impact of time difference t 2 -t 1 on the entropy H D,B but this time using only simulated mRNA data. Figure 8 shows that the H D,B curve is higher than the H B curve. Contrary to what is observed with protein level distributions, the H D,B curve slightly decreases as t 2 -t 1 increases, yet remaining far above the control curve. mRNAs are produced and degraded more rapidly than proteins, accelerating the approach to the stationary state. The same conclusions then hold, whether mRNA or protein data are used to infer cell trajectories.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 43, | |
| "text": "We focus now on experimental mRNA data from Semrau et al. [38]. When applying the Bursty model-based inference method to experimental data, no ground truth is available therefore, in a general context, standard or recently proposed methods-such as cluster-to-cluster fate probabilities [39], ancestral/source compositions [5]-may allow a posteriori validation by focusing on checkable quantities obtained from the identified coupling, provided that the size of the sample is large enough. Here, in order to apply our Bursty model-based inference method, we chose to focus on its predictive power. We first computed an mRNA-based version of the process B sch between time points t 1 = 0 and t 3 = 72h, and estimated the distribution ρ est,36 at t 2 = 36h (see Section Application to Semrau et al. [38] dataset and Supplementary File, Section 4). Next, we computed 1.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 44, | |
| "text": "the Earth Mover's Distance (EMD) between the experimental distribution at 36h and, on one side, ρ est,36 , denoted by EM D B , and on the other side, the distribution obtained using regularized OT, ρ D,36 , denoted by EM D D .", | |
| "entities": [ | |
| { | |
| "text": "OT", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 194, | |
| "end": 196, | |
| "confidence": 0.9984 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 45, | |
| "text": "Figure 9 shows ∆ EM D = EM D D -EM D B for each of the 41 genes in the dataset. It allows a measure of the performance of the Bursty model-based inference method compared to the Diffusion-based inference method. Green bar-plots correspond to a positive ∆ EM D , i.e. EM D B is smaller than EM D D , and red bar-plots correspond to negative ∆ EM D , i.e when EM D B is larger than EM D D . Heights of bar-plots quantify the contributions of Bursty model-based versus Diffusion-based inference methods.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 46, | |
| "text": "For 11 genes, ∆ EM D = 0. This means either that both EM D B and EM D D equal 0 (genes Fgf5, Cer1, Ascl1 and Sox9), or that they are equal (Esrrb, Dppa3, Nr0b1, Fgf4, Wnt8a, Pax6 and Gata4). In the first case, both methods are able to correctly predict the gene expression level, which is close to zero for the considered genes (no expression). In the second case, both methods similarly predict the gene expression level. Noticeably, gene distributions that do not evolve much between t 1 = 0h and t 3 = 72h display a small EMD.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 47, | |
| "text": "Out of the remaining 30 genes, for 22 genes the Bursty model-based inference method performs better. Values of EM D B and EM D D for each gene are displayed in Supplementary File (Supp. Figure 4). Noticeably, the large bar-plot of gene Sparc indicates that the distance between the two EM D is negative and rather large but, in that case, both EM D B and EM D D are also large (both methods have difficulties correctly predicting this gene dynamics). For genes Col4a2 and Lamb1, differences between EM D D and EM D B are also negative, yet smaller than for Sparc, but in that case both EM D are small (both methods perform well in predicting these gene dynamics, the Bursty model-based inference method performing better).", | |
| "entities": [ | |
| { | |
| "text": "Sparc", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 557, | |
| "end": 562, | |
| "confidence": 0.0181 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 48, | |
| "text": "One may note that we only focus here on an illustrative example where we predict the distribution at 36h by using distributions at 0h and 72h. However, the results are representative of other predictions that could be tested (not shown). Sometimes both Bursty model-based and Diffusion-based inference methods compare well (in particular in the early times, when only few genes drive the dynamics and their levels of expression do change slowly over time), nevertheless the above results show that the Bursty modelbased inference method can be applied on experimental data and performs better than the Diffusion-based method.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 49, | |
| "text": "We explored the possibility that the Bursty model of gene dynamics would represent an improved alternative to the standard diffusion model used to infer trajectories in OT theory. Based on distributions of protein counts, the Bursty model proved indeed quite efficient in capturing cells trajectories more accurately than the Diffusive model (Figure 1). This would be expected if we considered cells moving around in a Waddington's landscape [40] whose shape is constrained by the GRN structure [26,41]. In this case, diffusion could not easily represent the variables acceleration and deceleration due to valleys and hills. Noticeably, we assumed for the sake of simplicity the same number of cells in both initial and final distributions, yet our entire analysis could be extended to the case with different cell numbers by normalizing the distributions.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 50, | |
| "text": "When considering a toy model made of only one gene, we concluded that using the exact time-dependent distribution of the process was the best alternative to diffusion. Although expected, this result indicates that it is possible to very efficiently infer cell trajectories in a transcriptomic space. However, the exact distribution is only explicit for one gene. Extending the reference coupling RB ref to GRN with multiple genes seems a very challenging task, probably impossible to do for complex gene interactions. An approximation of this distribution could be considered, yet the approximation might rapidly become coarse.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 51, | |
| "text": "For a one-gene GRN, the computation of RB ref and D ref couplings is almost instantaneous on a personal computer (64 GB RAM and 13th Gen Intel(R) Core(TM) i7-13700H), whereas the computation of B ref takes approximately 7s (value averaged over 100 simulations). Increasing the number of genes does not impact the computational time for D ref , but it increases for B ref . The computational time is proportional to the number of genes in the GRN with almost no time-variability from one simulation to another, for two-gene up to six-gene GRNs where genes self-activate and inhibit the other genes (similar to the toggle switch model). For two genes, the computational time is 12s, for 3 genes, it is 13s, for 4 genes, 14s, for 5 genes 15.5s and for 6 genes 17s per simulation of the B ref coupling, on a computational server (High-Throughput Computing plateform with 1 CPU and 2 GiB memories). For most applications, such computational time would be acceptable. If needed, computations of B ref coupling can be parallelized, for instance when the number of simulations, the number of cells, or the number of genes is large. For the application to experimental data in Section Application to Semrau et al. [38] dataset, the computation of the B ref coupling took 23min using parallelization and 168min without parallelization.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 52, | |
| "text": "The full two-state model provides both mRNA and protein counts. Since experimentally available scRNA-seq datasets provide mRNA counts, it would have been appropriate to focus only on mRNA counts. However, it is not possible to reduce the Bursty model to an mRNA-only model and to preserve mathematical results, due to the fact that mRNA are typically much less stable than proteins and thereby would not preserve the Markovian property of the GRN. We can however reduce the Bursty model to a protein-only model that is well defined, along with quasi-steady-state distributions of mRNA counts conditionally on proteins [34,42]. Moreover, Baradat and Ventre [43] showed that for such a model it is still possible to define and solve a Schrödinger problem. We therefore decided to focus on distributions of protein counts when bench-marking on simulated datasets.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 53, | |
| "text": "In Section Application to experimental data from Semrau et al. [38], we applied our Busty model-based inference method on experimental data consisting of mRNA counts. We showed first that despite the above arguments in favor of preferentially using protein data, the Bursty model can also be applied to simulated mRNA data and provides in this case similar performances. To use our approach on an experimental dataset, it is first required to get a GRN, associated with the data, that is used to compute the reference coupling. This motivated the use of the dataset from [38], since Ventre et al. [36] had already inferred a GRN from this dataset using CARDAMOM. The inference and calibration of the mechanistic model with CARDAMOM are based on a two-step procedure, with bursting parameters being identified in the first step [36,41]. Noticeably, other algorithms have been proposed to infer the bursting kinetics from single-cell RNA sequencing data [44,45,46,47], the more recent ones benefiting from AI-based methods to infer kinetics and the method proposed in Huang et al. [47] relying on the use of the Telegraph model, as does our method. Importantly, even if the underlying GRN structure is not perfectly known, which is expected from using inference methods to obtain the GRN, we established in Section Analysis for other GRNs (see for instance Figure 5 and Supplementary File) that the Bursty model-based inference method is not very sensitive to parameter values of the GRN and correctly catches the trajectories even though the GRN used to compute the reference coupling is not correct.", | |
| "entities": [ | |
| { | |
| "text": "CARDAMOM", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 653, | |
| "end": 661, | |
| "confidence": 1.0 | |
| }, | |
| { | |
| "text": "CARDAMOM", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 723, | |
| "end": 731, | |
| "confidence": 1.0 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 54, | |
| "text": "When trying to predict an intermediate observation with Semrau et al. [38], the interpolation made with the Bursty model-based inference method performed better than the one made with Diffusion-based inference method. It is however worth noticing that the Diffusion-based inference method works well when observations are made on a short time period and when gene activity is limited. When observations are made on longer periods of time and genes exhibit important activity, the Bursty model-based inference method performs systematically better. Altogether the results presented in this article thus pave the way to the extension of GRN inference algorithms to methods combining GRN and trajectory inference in an integrated framework.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 55, | |
| "text": "We model gene expression using the framework introduced by Herbach et al. [34], and later simplified in the bursty regime [42,48,41]. In this model, a gene is characterized by its amount of mRNA and its level of proteins: mRNA is synthesized through stochastic transcriptional bursts, and degraded at rate d 0 ; proteins are synthesized from mRNA at rate s 1 and degraded at rate d 1 . The Bursty model is a piecewise-deterministic Markov process, with deterministic part given by", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 56, | |
| "text": "where M (t) and P (t) are continuous approximations of the quantity of mRNA molecules and the amount of protein, respectively, at time t, and a stochastic part consisting of mRNA bursts occuring at random times with rate k on and random sizes that follow an exponential distribution [48,41]. Within a cell, genes interact with each other via a GRN. Hence, a gene can be activated or inhibited by itself but also by other genes in the GRN. A burst is assumed to occur for gene i at a rate k on,i (P 1 , . . . , P G ), where G is the number of genes in the GRN and (P 1 , . . . , P G ) is the vector of protein levels of all genes in the GRN. The burst rate of gene i accounts for a basal activity parameter β i and the influence of other genes in the GRN, expressed via a gene-gene interaction matrix θ = {θ ji } j,i∈{1,...,G} [48,49], so that", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 57, | |
| "text": "where k 0,i and k 1,i correspond respectively to the minimum and maximum burst frequencies of gene i, and θ ji is the intensity of gene j action on gene i [34]. Parameters θ ji can be positive or negative, depending on whether gene j activates or inhibits gene i (Figure 10).", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 58, | |
| "text": "We consider a standard GRN made of two genes, the so-called \"toggle switch\" model. The two genes self-activate and inhibit each other, i.e. θ 11 = θ 22 ≥ 0 and θ 12 = θ 21 < 0.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 59, | |
| "text": "This GRN is illustrated in Figure 10, while Figure 11 illustrates the temporal evolutions of mRNA and proteins for a toggle switch GRN where gene dynamics obey the Bursty model, computed with parameter values given in Table 1. A deterministic toggle switch model would favor one gene over the other, depending on initial conditions of the system.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 60, | |
| "text": "In this stochastic version of the toggle switch, one observes that one gene activity always prevails, yet the dominant gene switches stochastically. More generally, by modifying the values of the interaction parameters θ ji in model (1)-(2), any two-gene GRN can be implemented (see, for example, Figure 6). Extensions to three-gene GRNs can be straightforwardly implemented (see, for example, Figure 7). Simulations of protein dynamics. Computations of the temporal evolution of protein and mRNA counts were performed using the Python package Harissa [49], which implements an exact simulation procedure for the Bursty model [42] (Figure 11). This allows to simulate one GRN, hence simulating N times the evolution of a GRN results in N independent cell trajectories (Figure 12).", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 61, | |
| "text": "In order to mimic biological experiments aimed at measuring a GRN activity at given observation times, we simulate N times the GRN with initial conditions IC M and IC P (see Table 1), select a simulation time t, and extract all mRNA and proteins counts for all genes.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 62, | |
| "text": "Throughout this work, we focus on distributions of protein counts only, yet the case of mRNA distributions is discussed in the last section. We denote by µ the distribution of G protein counts in a population of N cells at time t = t 1 , and ν the distribution of G 1.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 63, | |
| "text": "model is defined by [49,50]:", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 64, | |
| "text": "where E is the exponential distribution characterizing the random size of each burst (with mean value η), k on is the burst rate and d 1 the degradation rate of proteins, as before. Second, we consider a particular 1-dimensional case of model ( 3) with a constant k on : this setting can be interpreted either as a single gene with no feedback, or part of a GRN within a time range sufficiently short so that k on (P 1 , . . . , P G ) does not vary significantly during this period. This model turns out to be analytically tractable (see Supplementary File, Section 1). The time-dependent distribution of P (t) knowing P (0) = P 0 is given by", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 65, | |
| "text": "where δ e -d 1 t P 0 is the Dirac measure localized at e", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 66, | |
| "text": "d 1 t P 0 , and ν t has a density function g t (x|P 0 ) given by g t (x|P 0 ) = f t (x -e -d 1 t P 0 ), (5) where f t (x) = 1 η (e d 1 t -1)k on (e kont -1)d 1 exp -x η e d 1 t Φ k on d 1 + 1; 2; (e d 1 t -1) x η 1 {x≥0} , and Φ denotes the confluent hypergeometric (Kummer's) function of the first kind.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 67, | |
| "text": "For numerical applications, we regularize the singular part in (4) by replacing it with a normal distribution with small standard deviation ε ≪ 1,", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 68, | |
| "text": "so that the equality holds when ε → 0.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 69, | |
| "text": "Let us recall that experimental data would be collections of cell positions in the protein space at distinct time points, and that the objective of cell trajectory inference is to reconstruct cell trajectories in this space. Hence, we restrict the problem to two consecutive time points t 1 and t 2 . Given two collections X = (x 1 , . . . , x N ) ∈ (R G + ) N and Y = (y 1 , . . . , y N ) ∈ (R G + ) N , we search to construct a matrix Q = (Q ij ) ∈ R N ×N + of probabilities Q ij that a cell at position x i at time t 1 reaches the state y j at time t 2 , in the most realistic way. In the following, the matrix Q will be called a coupling between the collections X and Y .", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 70, | |
| "text": "By solving the Schrödinger problem (see Section Schrödinger problem), we look for the coupling that is the closest to a certain reference coupling R = (R ij ) ∈ R N ×N + , obtained using the Bursty model ( 1)-( 2), while being compatible with the data X and Y . For either the model ( 1)-( 2) or the Diffusive model, the reference coupling R is built by associating to the model a function being an exact or an approximated version of the transition kernel of the underlying theoretical process.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 71, | |
| "text": "Coupling for the reduced Bursty model. In the case of a single gene following model (3), we use the regularized analytical solution (6) to compute the probability density of reaching state y starting from x, that is p t (y|x) defined by p t (y|x) = e -kont h t (y|x) + (1 -e -kont )g t (y|x) where g t is defined in (5) and", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 72, | |
| "text": "Considering this transition kernel p t (y|x) and the uniform distribution p(x) ≡ 1/N over initial states (x 1 , . . . , x N ), the related reference coupling RB ref (x, y), for \"reduced Bursty reference coupling\", is defined as:", | |
| "entities": [ | |
| { | |
| "text": "RB ref", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 158, | |
| "end": 164, | |
| "confidence": 0.2453 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 73, | |
| "text": "where the normalizing term c(x) = 1≤j≤N p t 2 -t 1 (y j |x) is defined so that p t 2 -t 1 (y|x)/c(x) is a discrete probability distribution over (y 1 , . . . , y N ).", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 74, | |
| "text": "Coupling for the non-reduced Bursty model. Given a GRN whose dynamics are given by a Bursty model ( 1)-(2), we construct a corresponding approximated coupling B ref (x, y), denoted most of the time B ref , an approximation of its coupling between two collections at two different time points. It is the reference coupling defined by the probability of reaching y when starting from x, computed from the Bursty model ( 1)-( 2). We estimate B ref using a Gaussian method (illustrated in Figure 13). For each cell i in the initial distribution µ at t = t 1 , we simulate the Bursty model ( 1)-(2) n times with the cell i's coordinates as an initial condition with the same parameters than those used to compute the simulated data detailed in Table 1 (with one exception, when we test the influence of GRN parameter values in Section Unspecificity of the entropy indicator and the distribution parameters: in this case, the parameter values used to compute the couplings are the ones of the tested GRN, not the ones used to generate simulated data). We obtain a distribution denoted ν at time t = t 2 , representing the potential fates of cell i. An estimated density of ν is computed using Gaussian approximation, i.e. we create Gaussian distributions centered on cells in the distribution ν, and we compute the Gaussian kernel of this distribution using the default parameters of the gaussian kde() python function. The result is then divided by the total sum to ensure the resulting distribution is a probability distribution.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 75, | |
| "text": "Coupling for a Diffusive process. When the underlying model is a Diffusive process, we consider the natural discrete counterpart used in [6]. Specifically, the reference coupling denoted by D ref is defined by", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 76, | |
| "text": "where σ 2 is the diffusion coefficient, ∥x -y∥ is the Euclidean distance (here, in protein space) between x and y, and Dref is the normalizing constant given by Dref = 1≤i,j≤N exp -∥x i -y j ∥ 2 2σ 2 (t 2 -t 1 ) . The value of the diffusion coefficient was chosen to be both relevant and close to the value used in Schiebinger et al. [5]. The challenge was to consider a σ large enough to ensure a fast convergence of the Sinkhorn algorithm, while still capturing sufficiently accurate couplings (D ref (x, y) > 0).", | |
| "entities": [ | |
| { | |
| "text": "Sinkhorn", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 422, | |
| "end": 430, | |
| "confidence": 1.0 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 77, | |
| "text": "In its modern formulation, the Schrödinger problem, first introduced in 1932 [51], aims at minimizing the relative entropy of a coupling with respect to a reference coupling, with marginal constraints. This problem is deeply linked with OT: when the reference coupling is the joint law of a reversible Brownian motion with a given diffusion coefficient between times t 1 and t 2 , then the Schrödinger problem converges towards quadratic OT as the diffusion coefficient goes to 0 [52,53]. Let Π and Ψ be two discrete and finite spaces, e.g. Π = {x 1 , ..., x N } and Ψ = {y 1 , ..., y N } with N ∈ N, the number of cells. We denote by R ∈ P(Π × Ψ), the reference process, where P(Π × Ψ) is the space of probability measures on Π × Ψ.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 78, | |
| "text": "Since Π and Ψ are discrete spaces of size N , then R is a N ×N matrix. We denote by u ∈ P(Π) and v ∈ P(Ψ) two uniform vectors of size N (each cell has the same weight), and we search for the process Q solution of", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 79, | |
| "text": "where H is the relative entropy of Q with respect to R (also called Kullback-Leibler divergence of Q from R) defined by", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 80, | |
| "text": "This quantity will be called entropy throughout this paper. The Schrödinger problem (7) can be solved with the Sinkhorn algorithm [18,19]. We solve the Schrödinger problem for the reference couplings introduced in Section Coupling between time points. When R = RB ref , we denote by RB sch its solution, RB sch := Sch(RB ref ; u, v); for the reference coupling R = B ref , we denote by B sch its solution, B sch := Sch(B ref ; u, v); and when R = D ref , then we denote D sch its solution, D sch = Sch(D ref ; u, v).", | |
| "entities": [ | |
| { | |
| "text": "Sinkhorn", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 111, | |
| "end": 119, | |
| "confidence": 1.0 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 81, | |
| "text": "In order to compare the performances of the Bursty model-based inference method with the Diffusion-based inference method, we rely on the computation of a standard criterion, the entropy (8) between solutions of the Schrödinger problem (7). However, to ensure that results do not depend on the criterion, we also test other criteria measuring the difference between two stochastic processes.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 82, | |
| "text": "For one-gene GRN, we define H B,R := H(B sch |RB sch ) and H D,R := H(D sch |RB sch ), i.e. the entropy between B sch or D sch and RB sch . These values characterize the ability of a Diffusive process or a Bursty model versus RB sch to predict cell trajectories correctly. The process RB sch is considered as the reference process because it provides an approximation of the exact probability distribution. In parallel, we compute as a control measure H R := H(RB ref |RB sch ) in order to estimate the error made when performing stochastic simulations. For two-or-more-gene GRN, we define", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 83, | |
| "text": "the entropy between D sch and B sch , in order to evaluate the difference between the ability of a Diffusive process versus a Bursty model to correctly predict cell trajectories. Similarly to the one-gene GRN case above, we introduce", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 84, | |
| "text": "to estimate the error made when performing stochastic simulations.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 85, | |
| "text": "Values of H B should be smaller than H D,B values if B ref and B sch are closer than D sch and B sch . If the Bursty model-based reference process better predicts individual cell trajectories, the difference between H B and H D,B will be important.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 86, | |
| "text": "Due to stochasticity, we perform 100 iterations to compute these quantities and focus on the mean and standard deviation.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 87, | |
| "text": "Other difference indicators. To certify that our results do not depend on the choice of the entropy, we also compute the following criteria, for two processes Q and R:", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 88, | |
| "text": "-the Total Variation Distance:", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 89, | |
| "text": "-the Jensen-Shannon Divergence:", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 90, | |
| "text": "where W = (Q + R)/2;", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 91, | |
| "text": "-the Frobenius norm:", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 92, | |
| "text": "We use a time-stamped in vitro dataset from Semrau et al. [38] made of mRNA counts obtained from scRNA-seq experiments on mouse embryonic stem cells. The dataset comprises 41 genes whose levels of expression were measured at 9 observation times (0h, 6h, 12h, 24h, 36h, 48h, 60h, 72h and 96h) with approximately hundreds of cells obtained at each observation time. Genes are listed in Figure 9. Ventre et al. [36] inferred a GRN based on the Bursty model driving the dynamics observed in the dataset. We use this GRN and the associated parameter values to compute the reference coupling B ref , similarly to what has been done with protein inputs but using mRNA counts.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 93, | |
| "text": "Compared to the work done on simulated protein data, we use a larger number of simulations when constructing the approximated reference process, n = 10, 000, while other parameter values are given in Table 1. Given two experimental distributions µ at time t 1 and ν at time t 3 , we hereafter describe how we predict the distribution at time t 2 ∈ (t 1 , t 3 ).", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 94, | |
| "text": "By simulating the Bursty model with the GRN from [36], we first compute B ref (t 1 , t 3 ), the reference coupling between times t 1 and t 3 , and a reference distribution ρ ref at time t 2 . Second, we solve the Schrödinger problem and determine the coupling B sch (t 1 , t 3 ). We sample cell pairs (x, y) based on the coupling B sch (t 1 , t 3 ): the goal is to generate a certain number of pairs for each cell x in µ, where the probability of selecting a cell y in ν is determined by B sch (t 1 , t 3 ). Third, we build an interpolated distribution at t 2 , rooted in distributions µ and ν, as described below.", | |
| "entities": [ | |
| { | |
| "text": "B sch", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 488, | |
| "end": 493, | |
| "confidence": 0.0116 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 95, | |
| "text": "We define the estimated, interpolated distribution at time t 2 as", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 96, | |
| "text": "with T α * t 2 : (x, y) → α * x + (1 -α * )y for each pair (x, y), α * ∈ R G is an optimized vector and # is the push-forward function. Details of the optimization procedure leading to the estimation of α * are provided in Supplementary File. We compare the estimated distribution ρ est with the experimental distribution, using the Earth mover's distance (EMD), denoted by EM D B . We also compute the coupling between the distributions µ at t 1 and ν at t 3 using regularised OT (similarly to WOT [5]), denoted by D sch (t 1 , t 3 ). In this case, we use a log-transform of mRNA counts, we do not account for growth and death since this is not relevant in this case, and we interpolate a distribution at t 2 as in [5] by using α * = (t 2 -t 1 )/(t 3 -t 1 ) = 0.5 in (9) for all genes. We then compare the interpolated distribution and the experimental distribution by computing the EMD, denoted by EM D D . Finally, we compare the two distances, EM D B and EM D D , by computing ∆ EM D = EM D D -EM D B .", | |
| "entities": [ | |
| { | |
| "text": "OT", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 478, | |
| "end": 480, | |
| "confidence": 1.0 | |
| }, | |
| { | |
| "text": "WOT", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 495, | |
| "end": 498, | |
| "confidence": 1.0 | |
| }, | |
| { | |
| "text": "D sch", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 516, | |
| "end": 521, | |
| "confidence": 1.0 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 97, | |
| "text": "We thank the computational center of IN2P3 (Villeurbanne/France), where part of the computations were performed.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 98, | |
| "text": "This work has received financial support from the CNRS through the MITI interdisciplinary programs through its exploratory research program (TROPIC project under Action 80PRIME 2023). This work was supported by the PEPR Santé Numérique under Project AI4scMED MultiScale AI for Single Cell-Based Precision Medicine (ANR-22-PESN-0002).", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 99, | |
| "text": "All results included in this manuscript can be reproduced using the codes mentioned below (Code Availability). Experimental data used in this work have been provided by Semrau et al. [38] and are accessible at https://www.ncbi.nlm.nih.gov/geo/ under accession number GSE79578.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 100, | |
| "text": "All python codes are available at https://github.com/fourniecl/inference_trajectories_PDMP.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 101, | |
| "text": "The authors declare no competing financial or non-financial interests.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 102, | |
| "text": "CF performed the analyses and numerical simulations, and wrote the initial version of the manuscript. OG and FC supervised the work. All authors designed the methodology, and read and approved the final manuscript.", | |
| "entities": [ | |
| { | |
| "text": "FC", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 109, | |
| "end": 111, | |
| "confidence": 1.0 | |
| } | |
| ] | |
| }, | |
| { | |
| "index": 103, | |
| "text": "protein counts in the same population at time t = t 2 , with t 2 > t 1 . They are given by:", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 104, | |
| "text": "where δ is the Dirac distribution, x i = (P i g (t 1 )) g=1,...,G is the level of gene g's protein in cell x i at t 1 and y j = (P j g (t 2 )) g=1,...,G is the level of gene g's protein in cell y j at t 2 . For the sake of simplicity, when using simulated data cells have the same index in distributions µ and ν (i.e., the i-th component of µ is associated to the same cell than the i-th component in ν). Additionally, the two distributions are jointly normalized (between 0 and 1). Default parameter values are given in Table 1.", | |
| "entities": [] | |
| }, | |
| { | |
| "index": 105, | |
| "text": "In addition to the model introduced in the previous section, for which couplings will be estimated from simulations, we consider a simplified model (reduced Bursty model ) that will provide closed-form couplings. This simplification is achieved in two steps. First, a protein-only model is derived from a time scale separation of mRNA and proteins corresponding to d 0 ≫ d 1 (i.e., proteins assumed to be more stable than mRNAs). This", | |
| "entities": [] | |
| } | |
| ], | |
| "mentions": [ | |
| { | |
| "text": "Optimal Transport", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 176, | |
| "end": 193, | |
| "confidence": 0.9722 | |
| }, | |
| { | |
| "text": "WOT", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 195, | |
| "end": 198, | |
| "confidence": 0.9942 | |
| }, | |
| { | |
| "text": "Sinkhorn", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 475, | |
| "end": 483, | |
| "confidence": 1.0 | |
| }, | |
| { | |
| "text": "scEGOT", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 613, | |
| "end": 619, | |
| "confidence": 1.0 | |
| }, | |
| { | |
| "text": "OT", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 194, | |
| "end": 196, | |
| "confidence": 0.9984 | |
| }, | |
| { | |
| "text": "Sparc", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 557, | |
| "end": 562, | |
| "confidence": 0.0181 | |
| }, | |
| { | |
| "text": "CARDAMOM", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 653, | |
| "end": 661, | |
| "confidence": 1.0 | |
| }, | |
| { | |
| "text": "RB ref", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 158, | |
| "end": 164, | |
| "confidence": 0.2453 | |
| }, | |
| { | |
| "text": "B sch", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 488, | |
| "end": 493, | |
| "confidence": 0.0116 | |
| }, | |
| { | |
| "text": "D sch", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 516, | |
| "end": 521, | |
| "confidence": 1.0 | |
| }, | |
| { | |
| "text": "FC", | |
| "type": "Tool", | |
| "attributes": [], | |
| "begin": 109, | |
| "end": 111, | |
| "confidence": 1.0 | |
| } | |
| ], | |
| "duration_seconds": 33.34, | |
| "model": "csebe_scibert_bioinfo_voc_iteration4_seed100.pt" | |
| } |