GeRnika
is an R
package for the simulation
of tumor clonal data and the visualization and comparison of tumor
phylogenies. In this document we delve into the data simulation
functionality, so that advanced users can have a higher level of control
over the way instances are generated.
Each instance simulated by GeRnika
consists of 4
numerical matrices \(\boldsymbol{F}\),
\(\boldsymbol{F_{true}}\), \(\boldsymbol{U}\) and \(\boldsymbol{B}\), that relate to each other
so that:
\[\begin{equation} \boldsymbol{F_{true}} = \boldsymbol{U} · \boldsymbol{B} \end{equation}\]
This equation arises within the Variant Allele Frequency Factorization Problem (VAFFP) formulation of the Clonal Deconvolution and Evolution Problem (CDEP), and essentially states that the \(n\) mutation frequencies in a series of \(s\) tumor samples (matrix \(\boldsymbol{F_{true}}\)) are the result of the combination of the tumor clonal structure, represented by a tree (matrix \(\boldsymbol{B}\)), and the clone proportions captured in each sample (matrix \(\boldsymbol{U}\)). Specifically, these matrices are:
\(\boldsymbol{F_{true}} \in [0, 1]^{s \times n}\) matrix: It is a matrix \(s \times n\), and it contains the mutation frequency (VAF) values of the \(n\) mutations present in \(s\) tumor biopsies or samples.
\(\boldsymbol{B} \in \{0, 1\}^{n \times n}\) matrix: It is a binary square matrix of size \(n\) where \(b_{ij} = 1\) means that clone \(i\) contains the mutation \(j\). It represents the phylogeny of the tumor.
\(\boldsymbol{U} \in [0, 1]^{s \times n}\) matrix: It is a matrix \(s \times n\) where \(u_{ij}\) is the fraction of clone \(j\) in sample \(i\).
Now, all this holds in a noise-free scenario. However, we do know
that in real life there are many factors that make the VAF values noisy,
which essentially means that the VAF values of the mutations do no
longer exactly correspond to the sum of the sample fractions of all
those clones that contain that mutation. In this R
package
we have only considered the noise added by DNA sequencing procedures to
the VAF values and we have translated it into the \(\boldsymbol{F} \in [0, 1]^{s \times n}\)
matrix. Thus, we have that the product of the matrices \(\boldsymbol{U}\) and \(\boldsymbol{B}\) may not equal \(\boldsymbol{F}\):
\[\begin{equation} \boldsymbol{F} \neq \boldsymbol{U} · \boldsymbol{B} \end{equation}\]
The VAFFP formulation works under a few assumptions. The first one is that tumors have a monoclonal origin, i.e., they arise from a single abnormal cell that that at some point began to divide uncontrollably and founded the tumor. The second one is the infinite sites assumption (ISA), according to which a given mutation arises at most once over the course of evolution of the tumor, and mutations can not disappear. Any data generated using this package will adhere to those assumptions.
Importantly, even if the data simulation functionality in
GeRnika
was primarily devised for creating instances for
the CDEP, the output or the byproducts of the data models on it can
definitely be used in any other application in which this data may be
useful. With this in mind, we have implemented the procedures in such a
way that they are highly customizable by the user. We explain this usage
throughout this document.
GeRnika
allows the user to set parameters related to the
evolution of a tumor, the sampling procedure and the data acquisition
process, and include the number of clones in the tumor, number of
biopsies taken from the tumor, evolution model, evolutionary tree
topology and sequencing depth. More details can be found in the
following sections.
The general function to simulate an instance is
create_instance
. This function has the following
parameters:
Parameter | Description | Type |
---|---|---|
n |
Number of mutations/clones (\(n\)). | Natural number |
m |
Number of samples (\(s\)). | Natural number |
k |
Topology parameter that controls for the linearity of the topology. | Positive rational number |
selection |
Evolution model followed by the tumor. | Categorical: “positive” or “neutral” |
noisy |
Add sequencing noise to the values in the \(F\) matrix. | Boolean |
depth |
(only if noise = TRUE ) Average number of
reads that map to the same locus, also known as sequencing depth. |
Natural number |
seed |
Seed for the pseudo-random topology generator | Real number |
For instance, if we want to create a noise-free instance with 4 samples obtained from a tumor with 10 clones, evolving under neutral evolution and with a uniformly random topology, we can type the following:
I1 <- create_instance(n = 10, m = 4, k = 1, selection = "neutral", noisy = FALSE, seed = 1)
I1
#> $F
#> mut1 mut2 mut3 mut4 mut5 mut6 mut7 mut8 mut9 mut10
#> sample1 0.00 0.06 0.00 0.06 0.00 0.94 0.06 0.06 1 0.00
#> sample2 0.12 0.12 0.12 0.43 0.00 0.45 0.42 0.07 1 0.06
#> sample3 0.19 0.03 0.06 0.16 0.03 0.00 0.03 0.00 1 0.05
#> sample4 0.99 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1 0.00
#>
#> $B
#> mut1 mut2 mut3 mut4 mut5 mut6 mut7 mut8 mut9 mut10
#> clone1 1 0 0 0 0 0 0 0 1 0
#> clone2 0 1 0 1 0 0 1 0 1 0
#> clone3 1 0 1 0 0 0 0 0 1 0
#> clone4 0 0 0 1 0 0 0 0 1 0
#> clone5 0 1 0 1 1 0 1 0 1 0
#> clone6 0 0 0 0 0 1 0 0 1 0
#> clone7 0 0 0 1 0 0 1 0 1 0
#> clone8 0 1 0 1 0 0 1 1 1 0
#> clone9 0 0 0 0 0 0 0 0 1 0
#> clone10 1 0 1 0 0 0 0 0 1 1
#>
#> $U
#> clone1 clone2 clone3 clone4 clone5 clone6 clone7 clone8 clone9 clone10
#> sample1 0.00 0.00 0.00 0.00 0.00 0.94 0.0 0.06 0.00 0.00
#> sample2 0.00 0.05 0.06 0.01 0.00 0.45 0.3 0.07 0.00 0.06
#> sample3 0.13 0.00 0.01 0.13 0.03 0.00 0.0 0.00 0.65 0.05
#> sample4 0.99 0.00 0.00 0.00 0.00 0.00 0.0 0.00 0.01 0.00
#>
#> $F_true
#> mut1 mut2 mut3 mut4 mut5 mut6 mut7 mut8 mut9 mut10
#> sample1 0.00 0.06 0.00 0.06 0.00 0.94 0.06 0.06 1 0.00
#> sample2 0.12 0.12 0.12 0.43 0.00 0.45 0.42 0.07 1 0.06
#> sample3 0.19 0.03 0.06 0.16 0.03 0.00 0.03 0.00 1 0.05
#> sample4 0.99 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1 0.00
In this case, \(\boldsymbol{F}\) equals \(\boldsymbol{F_{true}}\).
If we instead want the instance to be noisy, we just need to adjust
the noisy
and depth
parameters:
I1 <- create_instance(n = 10, m = 4, k = 1, selection = "neutral", noisy = TRUE,
depth = 100, seed = 2)
I1
#> $F
#> mut1 mut2 mut3 mut4 mut5 mut6 mut7
#> sample1 0.3583333 0.22222222 0.0000000 0.01234568 1 0.4549763 0.00000000
#> sample2 0.2649573 0.05084746 0.0000000 0.01470588 1 0.3809524 0.00000000
#> sample3 0.5555556 0.00000000 0.6368715 0.60416667 1 0.1052632 0.09302326
#> sample4 0.9655172 0.00000000 0.9291339 0.98039216 1 0.0000000 0.00000000
#> mut8 mut9 mut10
#> sample1 0.4927536 0.00 0.12765957
#> sample2 0.4565217 0.00 0.05555556
#> sample3 0.0000000 0.06 0.22033898
#> sample4 0.0000000 0.00 0.08421053
#>
#> $B
#> mut1 mut2 mut3 mut4 mut5 mut6 mut7 mut8 mut9 mut10
#> clone1 1 0 0 0 1 0 0 0 0 0
#> clone2 0 1 0 0 1 0 0 0 0 1
#> clone3 1 0 1 1 1 0 0 0 0 0
#> clone4 1 0 0 1 1 0 0 0 0 0
#> clone5 0 0 0 0 1 0 0 0 0 0
#> clone6 0 0 0 0 1 1 0 0 0 0
#> clone7 0 0 0 0 1 1 1 0 1 0
#> clone8 0 0 0 0 1 1 0 1 0 0
#> clone9 0 0 0 0 1 1 0 0 1 0
#> clone10 0 0 0 0 1 0 0 0 0 1
#>
#> $U
#> clone1 clone2 clone3 clone4 clone5 clone6 clone7 clone8 clone9 clone10
#> sample1 0.36 0.20 0.00 0.02 0.00 0.00 0.00 0.42 0.00 0.00
#> sample2 0.27 0.09 0.00 0.01 0.23 0.07 0.01 0.32 0.00 0.00
#> sample3 0.00 0.00 0.61 0.00 0.02 0.00 0.12 0.00 0.04 0.21
#> sample4 0.00 0.00 0.94 0.00 0.00 0.00 0.00 0.00 0.00 0.06
#>
#> $F_true
#> mut1 mut2 mut3 mut4 mut5 mut6 mut7 mut8 mut9 mut10
#> sample1 0.38 0.20 0.00 0.02 1 0.42 0.00 0.42 0.00 0.20
#> sample2 0.28 0.09 0.00 0.01 1 0.40 0.01 0.32 0.01 0.09
#> sample3 0.61 0.00 0.61 0.61 1 0.16 0.12 0.00 0.16 0.21
#> sample4 0.94 0.00 0.94 0.94 1 0.00 0.00 0.00 0.00 0.06
This time, the matrices \(\boldsymbol{F}\) and \(\boldsymbol{F_{true}}\) are no longer the same.
The previous section describes how to generate instances in an easy
and rapid way. However, some users require more precise control over the
data, and this section is directed towards them. To begin with, we
provide an overview of the structure of Gernika
’s data
simulation functionality. Subsequently, we delve into each element in
detail and see what functions can be used to control the simulation of
those aspects.
In order to simulate the \(\boldsymbol{B}\) and \(\boldsymbol{U}\) matrices,
Gernika
relies in two models: a tumor model that simulates
the evolutionary history and current state of the tumor and a sampling
model that represents the tumor sampling process. Once that these two
matrices have been calculated, the matrix \(\boldsymbol{F_{true}}\) is obtained by
multiplying them. There is a last model which is a sequencing noise
model that can optionally be used if noisy data is desired; this adds
sequencing noise to the VAF values in the \(\boldsymbol{F_{true}}\) matrix, producing
the \(\boldsymbol{F}\) matrix. We
analyze each of these models in detail in the coming lines.
The tumor model generates a clonal tree \(T\) and an associated matrix \(\boldsymbol{B}\), together with the clone proportions \(\boldsymbol{c}\) and tumor blend at the moment of sampling. Briefly, the clonal tree \(T\) that represents the development of a tumor with a set of mutations \(M\) so that |\(M\)| = \(n\), is a rooted tree on an \(n\)-sized vertex set \(V_n = \{v_{1}, \dots, v_{n} \}\) and an \(n-1\)-sized edge set \(E_T = \{e_{ij} \}\), where \(v_{i}\) corresponds to the vertex or clone \(i\) and an edge \(e_{ij}\) between two vertices \(v_{i}\) and \(v_{j}\) represents a direct ancestral relationship.
In the first place, an \(n\)-sized tree \(T\) with a random topology is created in the following manner. First, the root node of \(T\), \(\mathcal R(T)\), is set and a random mutation \(M_i \in M\) is assigned to it. Then, for each of the remaining \(M_j \in M - \{M_i\}\) mutations, a new node \(v_j\) is created; then the mutation \(M_j\) is assigned to it and the node is attached as a child of an existing node in \(T\). In order to meet the ISA model, each of the newly added nodes inherits all the mutations present in its parent node.
The attachment of nodes to the tree is not uniformly at random; instead, the nodes in the \(T\) that is being built have different probabilities of being chosen as parent nodes for the new nodes, depending on the number of ascendants, \(\mathcal A(v_i)\), they have. Mathematically speaking, \(\forall M_j \neq \mathcal R(T)\), its parent node \(\mathcal P(M_j)\) is sampled from \(V_n\) following a multinomial distribution:
\[\begin{equation} \mathcal P(M_j) \sim M(n = \textrm{1}, \boldsymbol{p} = \boldsymbol{p}(v_i, k)); v_i \in V_n \label{eq:pk} \end{equation}\]
where
\[\begin{equation} \boldsymbol{p}(v_i, k) \propto k^{(|\mathcal A(v_i) | + 1)}; v_i \in V_T \end{equation}\]
\(k \in (0, +\infty)\) is the topology parameter that determines if the topology is more branched, with a decreasing probability for increasing numbers of ascendants (\(k\) \(<\) 1), or more linear, with an increasing probability for increasing numbers of ascendants (\(k\) \(>\) 1). At the same time, setting \(k\) to 1 assigns equal probabilities to all the nodes and generates a completely random topology. The relationship between \(k\) and the topology can graphically be seen in the following figure:
This procedure to obtain a tree \(T\) and one of the possible \(\boldsymbol{B}\) matrices associated to it
is implemented in the function create_B
. This function has
no parameters other than n
and k
described in
the section Basic instantiation so we do not
recommend changing it; we suggest to just adjust the value of
k
depending on the desired topology.
Once \(\boldsymbol{B}\) is obtained, the proportions of the clones in the tumor \(\boldsymbol{c}\) are simulated. We stress that these proportions \(c_i\) refer to the moment of the sampling because, due to the dynamic nature of tumor growth and evolution, these proportions need not be the same at any two time instants. It is also worth mentioning here that these proportions are not the ones that show in the \(\boldsymbol{U}\) matrix, which are the clone proportions, but the ones that do not depend on any sampling.
These clone proportions \(\boldsymbol{c}\) are calculated by sampling a Dirichlet distribution in each multifurcation of \(T\). For instance, for a node \(v_i\) with children \(\mathcal K(v_i)\) = {\(v_j\), \(v_k\)}, we draw one sample \(\{x_i, x_j, x_k\}\) that represents the proportions of the parent clone and its two children from \(Dir(\alpha_i, \alpha_j, \alpha_k)\), respectively. When this sampling is performed in an internal node of the tree, i.e., not in the root node, these proportions are scaled to the original proportion of the parent clone so that once all the multifurcations have been visited, the proportions of all the clones in \(T\) sum up to one.
The parameters of the Dirichlet distribution depend on the evolution
model of the tumor. In GeRnika
we consider two basic cases:
positive selection-driven evolution or neutral evolution. According to
positive selection-driven evolution, some mutations provide growth
advantage whereas most of them do not, so the former clones get over
other existing clones and take up the biggest part of the tumor. This
results in tumors dominated by few clones, while the rest of clones are
observed in minor proportions. Under neutral evolution, instead, there
is not a significant number of mutations that provide fitness advantage
and clones accumulate just out of tumor progression, so all clones are
present in similar proportions
Based on this, the \(\alpha\) parameters for the Dirichlet distribution for positive selection-driven evolution have been set to \(\boldsymbol{\alpha}\) = 0.3, and for the neutral evolution to \(\alpha_{p}\) = 5 and \(\boldsymbol{\alpha}_{c}\) = 10 . We use different alpha values for parent and children nodes in the neutral evolution so that clones that result from multiple multifurcations do not end up with smaller proportions just because of their position in the topology, ruining the expected clone proportion distribution for this type of evolution model. These values have been selected empirically and their effect is depicted in the following figure, which shows how 5,000 random samples from the mentioned Dirichlet distributions (for the particular case of 3 dimensions, i.e., one parent and two children nodes) would be distributed on a 2-simplex. As can be seen, for positive selection (\(\boldsymbol{\alpha}\) = 0.3), the \(\{x_i, x_j, x_k\}\) values are pushed towards the corners of the simplex (\(\alpha_i\) is \(<\) 1) in an uniform manner (all \(\alpha_i\) are equal); in other words, the samples tend to be sparse; usually one component has a large value and the rest have values close to 0. Instead, when the neutral selection is adopted (\(\boldsymbol{\alpha}\) = [5, 10, 10]), the values in \(\{x_i, x_j, x_k\}\) concentrate close to the middle of the simplex (all \(\alpha_i\) are \(>\) 1) but tend to deviate to those components with larger \(\alpha\). This means that samples \(\{x_i, x_j, x_k\}\) are less sparse, with larger values for \(x_2\) and \(x_3\) in this case, which represent the children nodes.
In short, we have that the proportion of the clone \(v_i \in V_T | v_i \neq \mathcal R(T)\) in the tumor at the moment of sampling, given by \(c_i\), follows this distribution:
\[\begin{equation} c_i \sim c_{\mathcal P(v_i)} \cdot \gamma_i \cdot \gamma_i^\prime \end{equation}\]
where
\[\begin{equation} \gamma_i \sim Beta(\alpha_{c}, \alpha_{p} + \alpha_{c} \cdot (|\mathcal K(\mathcal P(v_i))| - 1)) \end{equation}\]
and
\[\begin{align} \gamma_i^\prime = 1 \quad \text{if } |\mathcal K(v_i)| = 0 \\ \gamma_i^\prime \sim Beta(\alpha_{p}, \alpha_{c} \cdot |\mathcal K(v_i)|) \quad \text{if } |\mathcal K(v_i)| \neq 0 \end{align}\]
In turn, if \(v_i = \mathcal R(T)\):
\[\begin{equation} c_i \sim Beta(\alpha_{p}, \alpha_{c} \cdot |\mathcal K(v_i)|) \end{equation}\]
The calculation of the clone proportions is implemented in the
functions .distribute_freqs
and
calc_clone_proportions
. The former function samples the
Dirichlet distribution in the node corresponding to the clone designed
by the clone_idx
parameter and scales the resulting values
to the original proportion of the parent clone, unless
clone_idx
is the root clone. The
calc_clone_proportions
function is responsible for
calculating the clone proportions throughout the tree \(T\), so it calls the
.distribute_freqs
function at each multifurcation.
Although we have seen that the parameter values used in the Dirichlet
distributions are fairly good representatives of the underlying
evolution models, it is true that, as we have said, the determination of
these values has been done in an empirical way. If the user so wishes,
they can try to vary these values by simply changing the values of the
dirich_params
variable in the
.distribute_freqs
function (the first row of the tibble in
that variable corresponds to the clone clone_idx
, whereas
the second row corresponds to the children clones of
clone_idx
:
.distribute_freqs
#> function (B, clone_idx, clone_proportions, selection)
#> {
#> dirich_params <- dplyr::tibble(positive = rep(0.3, 2), neutral = c(5,
#> 10))
#> children_clone_idx <- get_children_idx(B = B, node_idx = clone_idx)
#> if (!is.null(children_clone_idx)) {
#> params <- dplyr::pull(dirich_params, eval(selection))
#> dirich_values <- rdir(1, c(params[1], rep(params[2],
#> length(children_clone_idx))))
#> norm_dirich_values <- dirich_values * clone_proportions[clone_idx]
#> clone_proportions[c(clone_idx, children_clone_idx)] <- norm_dirich_values
#> }
#> return(clone_proportions)
#> }
#> <bytecode: 0x0000022631d85370>
#> <environment: namespace:GeRnika>
To finish with the tumor model, the tumor blend is simulated, i.e. how physically mixed are the clones between them. For doing so, we simplify the spatial distribution to one dimension and we model the tumor as a Gaussian mixture model of \(n\) components, where each component \(G_i\) represents a tumor clone and the mixture weights are given by \(\boldsymbol{c}\). The variance for all components is set to 1, while their mean value is a random variable. Specifically, a random clone is selected in the first place and the mean value of its component is set to 0. Then, the mean values of the \(n - 1\) remaining components are calculated in a sequential manner by summing \(d\) units to the mean value of the immediately prior component. \(d \in \{0, 0.1, \ldots,4\}\) follows a Multinomial distribution so that:
\[\begin{equation} p(d) \propto Beta(d; \alpha=1, \beta=5) \end{equation}\]
For \(d\) = 0, the two clones are completely mixed; for \(d\) = 4, they are physically far from each other. This upper limit for \(d\) has been set empirically, after observing that with this value, the overlapping area between the two clones is already minimal, as shown in the following figure:
The tumor blend simulation is implemented in the
place_clones_space
function:
place_clones_space
#> function (B)
#> {
#> n <- nrow(B)
#> . <- NULL
#> max_sep <- 4
#> mean_diffs <- seq(from = 0, to = max_sep, by = 0.1)
#> clone_order <- sample(1:n, n)
#> clone_mean_diff <- sample(x = mean_diffs, size = n - 1, prob = stats::dbeta(x = seq(0,
#> 1, length.out = length(mean_diffs)), shape1 = 1, shape2 = 5),
#> replace = TRUE)
#> clone_means <- c(0, cumsum(clone_mean_diff))
#> clone_sd <- 1
#> x <- seq(min(clone_means - 3 * clone_sd), max(clone_means +
#> 3 * clone_sd), 0.01)
#> y <- purrr::map(1:n, function(z) stats::dnorm(x, mean = clone_means[z],
#> sd = clone_sd)) %>% do.call(cbind, .) %>% dplyr::as_tibble(.name_repair = ~vctrs::vec_as_names(...,
#> repair = "unique", quiet = TRUE)) %>% magrittr::set_colnames(paste0("clon",
#> clone_order))
#> spatial_coords <- y %>% dplyr::mutate(idx = x)
#> return(list(spatial_coords = spatial_coords, x = x))
#> }
#> <bytecode: 0x000002262e46c3e0>
#> <environment: namespace:GeRnika>
The user can vary, e.g., increase or decrease, fix to a certain
value, the range of values \(d\) can
take to try other blending patterns. This can easily be done by changing
the values of the vector mean_diffs
, by modifying the
minimum and maximum values of the sequence, or even the increment of the
sequence.
Likewise, the user can try to vary the parameters in the Beta
distribution with the same purpose. The current parameters \(\alpha\) = 1 and \(\beta\) = 5 have been adjusted manually so
that small \(d\) values are favoured
and thus the samples obtained in the following steps do not become too
sparse. If sparsity if desired, instead, they can, as said, increase the
values in the mean_diffs
variable or, equivalently, adjust
the parameter values and give \(\beta\)
a value larger than \(\alpha\) in th
clone_mean_diff
variable. In any case, once the role of the
Beta distribution regarding the tumor blend is clear, the user can make
informed decisions on its parameter values.
So far we have explained how the clones of a tumor are modelled by the tumor model. However, in real practice there is no easy way of observing these tumor global properties; instead, we only have the information that samples or biopsies can provide. This means, among many other things, that the real clone proportions \(\boldsymbol{c}\) can not be obtained. Instead, we can only get the clone proportions, which depend on certain sampling procedure; most of the time, the sampled clone proportions will not match the real ones. These sampled clone proportions are, in fact, the \(\boldsymbol{u}_{i.}\) elements in the matrix \(\boldsymbol{U}\).
The sampling model in Gernika
models the physical
sampling of the tumor and, at the same time, allow us build the \(\boldsymbol{U}\) matrix of the problem.
This model actuates over the data simulated with the tumor model;
specifically, it simulates some sampling performed in a grid-manner over
the tumor Gaussian mixture model described in the previous subsection.
Let \(G_i\) and \(G_n\) be the components with the lowest and
largest mean values, respectively. Then, the domain is limited to \([\mu_{Gi} - 3 \cdot \sigma_{Gi}, \mu_{Gn} + 3
\cdot \sigma_{Gn}]\) with \(\sigma_{Gi}
= \sigma_{Gn} =\) 1, and it is divided into \(m\)+1 equal-sized bins by \(m\) cutpoints that represent the sampling
sites. The 1\(^{st}\) and \(m^{th}\) cutpoints are always set to \(\mu_{Gi} - 3 \cdot \sigma_{Gi} + 20\) and
\(\mu_{Gn} + 3 \cdot \sigma_{Gn} -
20\), and the remaining \(m\)-2
cutpoints are calculated accordingly.
The probability density of the components of the mixture model on those sites is normalized so that their sum sums up to 1 and the resulting values \(\boldsymbol{p}_i\) are taken as the tumor clone composition values on the sampling spot \(i\):
\[\begin{equation} \forall i \quad p_{ij} = \frac{c_j \cdot p_j(i)}{\sum_{j = 1}^{n} c_j \cdot p_j(i)} \label{eq:pij_2} \end{equation}\]
Finally, the effect of the cell count on the samples is considered so that the final tumor clone composition values in each sample, i.e., \(\boldsymbol{u}_{i.}\), are modeled using a multinomial distribution:
\[\begin{equation} \boldsymbol{u}_{i.} \sim \frac{M(n = \textrm{100}, p = \boldsymbol{p}_i)}{100} \end{equation}\]
The sampling model is implemented in the and create_U
function:
create_U
#> function (B, clone_proportions, density_coords, m, x)
#> {
#> n_cells <- 100
#> idx <- value <- proportion <- weights <- sampled_weight <- scaled_weight <- . <- NULL
#> clone_order <- setdiff(colnames(density_coords), "idx")
#> cutpoints_idx <- seq(20, length(x) - 20, length.out = m)
#> cutpoints <- x[cutpoints_idx]
#> U <- purrr::map(cutpoints, function(j) {
#> reshape2::melt(density_coords, id = "idx") %>% dplyr::filter(idx ==
#> j) %>% dplyr::mutate(weights = value/sum(value)) %>%
#> dplyr::left_join(clone_proportions, by = c(variable = "clone_idx")) %>%
#> dplyr::mutate(scaled_weight = weights * proportion/sum(weights *
#> proportion)) %>% dplyr::mutate(sampled_weight = stats::rmultinom(n = 1,
#> size = n_cells, prob = scaled_weight)/n_cells) %>%
#> dplyr::select(sampled_weight)
#> }) %>% do.call(cbind, .) %>% magrittr::set_rownames(clone_order) %>%
#> magrittr::set_colnames(as.character(cutpoints)) %>% t()
#> U <- U[, order(as.numeric(gsub("clon", "", colnames(U))))]
#> return(U)
#> }
#> <bytecode: 0x0000022630b0b858>
#> <environment: namespace:GeRnika>
The user can again play with this function according to their needs.
They could, for instance, vary the sampling sites; for doing so, they
would just need to change the variables cutpoints_idx
and/or cutpoints
in the create_U
function, as
these variables contain the sampling sites in the Gaussian mixture model
domain, which spans, it is worth recalling, from \(\mu_{Gi} - 3 \cdot \sigma_{Gi}\) to \(\mu_{Gn} + 3 \cdot \sigma_{Gn}\).
It may also be interesting to adjust the \(n\) value in the multinomial distribution,
which accounts for the number of cells sampled. Choosing a relatively
low value for this parameter reduces the number of decimals in \(\boldsymbol{u}_{i.}\), so that clones with
frequencies several orders of magnitude lower than 1 can be modeled as
absent in the sample, i.e., with composition values equal to 0. This is
indeed way more realistic than truly observing them in so low
frequencies. We set the value of \(n\)
to 100 because it seemed a reasonable value and the resulting VAF values
are in line with VAF values obtained in real samples; however should the
user wish to try any other value, they could simply change the value of
the variable n_cells
in the first line of the
create_U
function.
Up to this point, we have explained how the \(\boldsymbol{B}\) and \(\boldsymbol{U}\) matrices of an instance are simulated. In case we are simulating noise-free data, the instance is complete once Equation (1) has been applied on them in order to obtain both \(\boldsymbol{F_{true}}\) and \(\boldsymbol{F}\) matrices.
As a brief reminder, each element \(f_{t-ij}\) in \(\boldsymbol{F_{true}}\) denotes the frequency or VAF of the mutation \(M_j\) on sample \(i\) or, in other words, the proportion of sequencing reads that harbour the mutation \(M_j\) on that sample. This also means that 1 - \(f_{t-ij}\) of the reads on that sample do not observe the mutation but the normal or reference nucleotide.
As the user may know, there exist numerous factors that alter the VAF value in an artificial manner so that this value deviates from the real ratio between the mutation and the reference nucleotides, including the noise that is introduced by the DNA sequencing process itself. The ways noise is introduced in this process are mainly two: through an incorrect nucleotide read of a fragment due to limitations of the sequencing instrument (e.g., a position that contains an A is read as a T) and through the production of a biased number of reads for a site that arises due to chemical reaction particularities or simply because not all the fragments get to be sequenced. This limitation can however be mitigated up to a certain level, as it has been shown that the higher the depth of coverage (\(\mu_{sd}\)), i.e., the average number of reads that cover each position, the most accurate is the VAF value. In order to account for the effect of such noise in the data instances, we have designed a sequencing noise model. This model adds noise to the \(\boldsymbol{F_{true}}\) matrix and builds a noisy \(\boldsymbol{F}\) matrix, so that \(\boldsymbol{F} \neq \boldsymbol{U} · \boldsymbol{B}\). Specifically, the model simulates the noise at the level of the sequencing reads and then recalculates the new \(f_{ij}\) values, as follows.
The sequencing depth \(r\) at the genomic position where \(M_j\) occurs in sample \(i\) is distributed as a negative binomial distribution so that: \[\begin{equation} r_{ij} \sim NB(\mu = \mu_{sd}, \alpha = 5) \end{equation}\]
The number of reads supporting the alternate allele \(r^{a}_{ij}\) is modeled by a binomial distribution: \[\begin{equation} r^{a}_{ij} \sim B(n = r_{ij}, p = f_{t-ij}) \label{eq:ra} \end{equation}\]
The use of probability distributions for sampling the sequencing depth and the number of reads instead of using fixed values is the manner in which stochasticity is introduced in the read numbers and thus in the \(f_{ij}\) values.
Illumina mismatch errors are known to be around 0.1%. For simulating its effect on the VAF values, the number of reads \(r^{a\prime}_{ij}\) that, despite originally supporting the alternate allele, contain a different allele as a result of a sequencing error are simulated as:
\[\begin{equation} r^{a\prime}_{ij} \sim B(n = r^{a}_{ij}, p = 0.001) \end{equation}\]
We also need to consider the situation where the reads contain the reference nucleotide but are read with the alternate allele as a result of this error. This is better understood with an example. Let us imagine that there is one genomic position where the normal cells have a T but in some cells there is a mutation and the T has changed to an A. For the normal cells, with probability 0.1\(\%\) a sequencing error will occur and instead of a T, one of C, G or A will be read, all with equal chance. Thus, \(\frac{0.001}{3}\)% of the time, reads with the mutation of interest will arise from normal reads:
\[\begin{equation} r^{r\prime}_{ij} \sim B(n = r_{ij} - r^{a}_{ij}, p = \frac{0.001}{3}) \end{equation}\]
Thus, taking all these into consideration, the final noisy VAF values \(f^{n}_{ij}\) are simulated as:
\[\begin{equation} f_{ij} = \frac{r^{a}_{ij} - r^{a\prime}_{ij} + r^{r\prime}_{ij}}{r_{ij}} \end{equation}\]
In the following figure we have depicted for a collection of noisy instances, the density of the mean absolute error between the \(\boldsymbol{F}\) and its corresponding noise-free \(\boldsymbol{F_{true}}\) matrix. As it can be seen, the higher \(\mu_{sd}\), the lower the error we introduce to the \(\boldsymbol{F}\) matrix. This is an expected behaviour since \(r^{a}_{ij}\) values follow a binomial distribution (Equation 14) where the number of trials is conditioned by \(\mu_{sd}\) (Equation 13) and the event probability is the \(f_{ij}\) value.
The addition of sequencing noise to the data is controlled by the
boolean noisy
parameter in the create_instance
function, and its quantity by the depth
parameter in the
same function. As explained in the previous paragraph, noise will
increase with decreasing depth
values. The default value
for this parameter is 30 (30X); this value was chosen as most whole
genome sequencing (WGS) data is sequenced at this depth. Again, the user
can choose any other value that caters their needs.
The negative binomial distribution is a distribution commonly used to
model sequencing depth. In such context, the mean of the distribution
\(\mu\) represents the expected number
of reads for a position, whereas the overdispersion parameter \(\alpha\) controls the degree of variability
in the number of reads across different samples of the distribution. It
is then easy to see that we can use this parameter to control for the
noise introduced in the data. If we increase the value of \(\alpha\), the range of values of the
samples drawn from the distribution will increase too, with some samples
having much higher or lower depth values than the mean sequencing depth.
In other words, increasing the value of \(\alpha\) will increase the noise in the
data. Conversely, if we decrease the value of \(\alpha\), there will be less variability
across samples and the sequencing depth values will be more consistent
across samples, i.e., data will be less noisy. The user can play with
this value in the create_instance
function, by changing the
value of the overdispersion
variable as we have just
explained.
create_instance
#> function (n, m, k, selection, noisy = TRUE, depth = 30, seed = Sys.time())
#> {
#> if (!is.character(selection)) {
#> stop("\n selection must be a character")
#> }
#> if (selection != "neutral" & selection != "positive") {
#> stop("\n selection must be neutral or positive")
#> }
#> if (depth < 1 || !identical(round(depth), depth)) {
#> stop("\n depth must be a natural number greater than or equal to 1 ")
#> }
#> set.seed(seed)
#> overdispersion <- 5
#> B <- create_B(n, k)
#> clone_proportions <- calc_clone_proportions(B, selection)
#> clones_space <- place_clones_space(B)
#> density_coords <- clones_space$spatial_coords
#> domain <- clones_space$x
#> U <- create_U(B = B, clone_proportions = clone_proportions,
#> density_coords = density_coords, m = m, x = domain)
#> F_true <- calc_F(U = U, B = B, heterozygous = FALSE)
#> if (noisy) {
#> F_matrix <- add_noise(F_matrix = F_true, depth = depth,
#> overdispersion = overdispersion)
#> }
#> else {
#> F_matrix <- F_true
#> }
#> clone_names <- purrr::map(1:n, function(x) paste0("clone",
#> x))
#> mutation_names <- purrr::map(1:n, function(x) paste0("mut",
#> x))
#> sample_names <- purrr::map(1:m, function(x) paste0("sample",
#> x))
#> rownames(F_matrix) <- sample_names
#> colnames(F_matrix) <- mutation_names
#> rownames(B) <- clone_names
#> colnames(B) <- mutation_names
#> rownames(U) <- sample_names
#> colnames(U) <- clone_names
#> rownames(F_true) <- sample_names
#> colnames(F_true) <- mutation_names
#> return(list(F = F_matrix, B = B, U = U, F_true = F_true))
#> }
#> <bytecode: 0x000002263185c9a8>
#> <environment: namespace:GeRnika>
We would like to make two final notes regarding the sequencing noise model. The first one is that, even if our simulated data procedure follows the ISA, it is possible that due to the process of adding noise, resulting data breaks this assumption. The second one is a reminder that the sequencing noise model is an optional model only to be used if noisy data is to be generated.
To close up, the following figure provides a graphical summary of the whole process and illustrates an example: