Abstract

In determining regulatory interactions within cells, models based on ordinary differential equations (ODE) are a popular choice. Here, we present a reimplementation of one such model in Stan. The model features a spline fit where the resulting spline is a parameter of the ODE. For practical reasons, the model avoids the Stan’s ODE solver in favor of a custom solution. We also discuss why the model as traditionally formulated is not well identified and introduce reparametrizations that improve identifiability and make it easier to specify reasonable default priors.

This notebook shows first usable version of the model, but there are still several outstanding issues which we highlight. We nevertheless hope that the modelling approach we took is of interest to the Stan community and can provide inspiration for other models.

A runnable version of this Rmd file and all supporting code can be found at https://github.com/cas-bioinf/genexpi-stan/tree/stancon2018

Biological Background

In a very simplified form the central dogma of molecular biology may be paraphrased as follows: The DNA contains all of the instructions needed to run a cell and can be divided into coding and non-coding regions. The coding regions of DNA can be transcribed to form messenger RNA (mRNA) molecules containing a mirror-copy of the coding region. The messenger RNA is then translated to create protein. Proteins perform most of the actual functions of the cells, including control of transcription and translation. This process is regulated at all stages and involves many feedback loops: most importantly, the rate of transcription of a gene is controlled by the abundance of regulatory proteins binding to specific sequences of the DNA near the coding region. The rate of translation and degradation of mRNA can also be regulated separately by other proteins and the rate of degradation of proteins themselves is also affected by other proteins. There are many exceptions where the above simplification does not hold, but the vast majority of cellular life can be described in these terms.

Understanding what are the regulations taking place is thus an important step in understanding how cells work. However, our ability to observe what happens in the cell is very limited: The only commonly available method that can capture information about all genes in a single experiment is measuring the concentration of mRNA which is in turn strongly related to expression (how many mRNAs are transcribed from the gene in a unit of time). Gathering expression data remains relatively expensive and except for the most studied organisms, at most several dozen whole genome expression experiments have been published.

It has been shown that mRNA and protein concentration are mostly correlated, however this relationship is imperfect (Maier, Güell, and Serrano 2009,Gygi et al. (1999)). Nevertheless, expression data is the best proxy for protein concentration available at the whole genome level and expression data are thus widely used to infer interactions that control transcription of genes.

Scope

The model presented in this notebook aims at modelling and identifying transcriptional regulations from time series data of gene expression. While the model aims to be more general, our particular interest is on determining for which genes does a known sigma factor initiate transcription.

The present work was motivated by reading (Titsias et al. 2012) who developed a fully Bayesian model of transcriptional regulation, extending a classical regulation model (Vohradský 2001) with Gaussian Processes. Titsias et al. used a custom Metropolis-within-Gibbs sampler and to our knowledge, there is no other fully Bayesian treatment of the model available. Initially, we tried to directly rephrase the Titsias et al. model in Stan, but we ran into large computational difficulties, including several non-identifiabilities, that we were not able to resolve without modifying the model. In this notebook, we present a model that resulted from resolving those issues and a slight simplification of the original model, as our use case didn’t need to model protein dynamics separately. We further show how we determine model fit and use a simple auxiliary expression model to filter out genes whose expression does not identify the parameters of the full model.

Data

The data we will work with in this notebook is a time series of 14 microarray measurements of expression of 4008 genes in the bacterium Bacillus subtilis during germination (transition from dormant spores to normal metabolism). The data is depositioned in the Gene Expression Omnibus under GSE6865 (Keijser et al. 2007, https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE6865). The individual samples were taken from a single fermentor where the bacteria grew and represent averages of a large number of cells. The cells are expected to be mostly synchronized by germination initiation at t=0 and (hopefully) stayed in sync for the duration of the experiment.

While recent methods allow expression measurement for individual cells, such data are expensive, usually gathered only at a single time point, contain large amount of noise and introduce additional analytical challenges (I tried and failed). To our knowledge, no succesful inference of transcriptional regulations has ever been made from single cell data at the time of this writing.

#Try to download the data, if it was not already present on the computer

data_dir <- here("local_data")
if(! dir.exists(data_dir)) {
  dir.create(data_dir)
}
data_file <- here("local_data","GSE6865_series_matrix.txt.gz")
if(!file.exists(data_file)) {
  download.file("ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE6nnn/GSE6865/matrix/GSE6865_series_matrix.txt.gz", data_file)
}

# Read and preprocess the data
gse6865_raw_df = read.delim(gzfile(data_file), comment.char = "!") #Intermediate data frame representation

# Raw profile data. We scale by 1000 to get to more wieldy values
gse6865_raw = as.matrix(gse6865_raw_df[,2:15]) / 1000
rownames(gse6865_raw) = gse6865_raw_df$ID_REF

#Times (in minutes) for the individual samples
gse6865_raw_time = c(0,5,10,15,20,25,30,40,50,60,70,80,90,100)
colnames(gse6865_raw) <- sapply(gse6865_raw_time,  FUN = function(x) { paste0(x,"min")})

#There are a few NA values in 1st and 2nd measurement, which we will ge trid of.

# For the first measurement, we can expect the value to be 0 (the series are from germination)
gse6865_raw[is.na(gse6865_raw[,1]),1] = 0

# For the second measurement we will go with a linear interpolation
na2 = is.na(gse6865_raw[,2])
gse6865_raw[na2,2] = 0.5 * gse6865_raw[na2,1] + 0.5 * gse6865_raw[na2,3]

#cleanup the intermediate results
rm(gse6865_raw_df)
rm(na2)
rm(data_dir)
rm(data_file)

Below is a sample of the expression profiles of 20 random genes, notice that most genes have almost zero expression.

set.seed(2345277)
genes_to_show <- sample(1:nrow(gse6865_raw), 20)
ggmatplot(gse6865_raw_time, t(gse6865_raw[genes_to_show,]), main_geom = geom_line(alpha = 0.6), x_title = "Time [min]", y_title = "expression")

Basics of The Model

In all of the following, xi is the true (unknown) expression of a target gene, yj the true protein concentration of a regulator and ρi the regulatory input for the target gene. Both xi, yj and ρi are functions of time while all other parameters are constants. The regulatory input is a linear combination of protein concentrations of the regulators:

ρi=∑jwi,jyj+bi

where F is the regulatory response function, in our case the logistic sigmoid.

Our goal is to use Stan to infer the parameters of this model (si,wi,j,bi,di), given the observed expression of the target genes and regulators over time. The model assumes constant degradation (di) over time which is unrealistic, but necessary as most datasets do not contain enough information to identify varying degradation.

There are two forms of expression data available with widely different noise models: microarray and RNA-seq. The former produces positive continuous measurements with approximately normal error while the latter produces count data with negative-binomial distribution. The datasets we are interested in come from microarray experiments so we will focus on those, but using RNA-seq means just changing the observation model. For microarrays, the observation model can be treated as a truncated normal:

~xi(t)∼N(xi(t),σabs+xi(t)σrel)|~xi(t)>0

The two error terms represent the fixed absolute error (σabs) which is the result of technical noise in the microarray platform which is constant (Klebanov and Yakovlev 2007) and the relative error (σrel) wich represents the uncaptured biological variation which tend to be proportional to the expresssion of the genes.

For the regulator(s) there are two possible regimes: a) the expression of the regulators is observed and the protein concentration is assumed to be identical to the true expression or b) the y is estimated only through its influence on x. In case a), the exact same error model applies to regulators.

In the following, we treat both σabs and σrel as given. The model can treat the error terms as parameters, but with the dataset we work with, the resulting σrel values tended to be unrealistically high and most fits unexpectedly poor as a consequence. We are yet to investigate why this is the case.

Based on our previous experience with similar models, we expect σabs=0.1 and the biological variation to be around 20% of the expression, suggesting σrel=0.1.

Modelling Regulator Expression

To solve the regulation ODE numerically, we need to have yj available at much finer intervals than the available measurements (1-2 minute intervals have proven sufficient). Initially, we considered Gaussian Processes (as Titsias et al. 2012)), but those introduced computational issues and we switched to B-splines which have less theoretical appeal, but introduce fewer parameters and were easier to fit.

We however need to ensure that yj are positive. After some early setbacks with the `log1pexp transform (x+=ln(1+ex)), we settled to simply subtract the minimum of the spline from all points¹, i.e. given the spline basis B we get:

¯yj=Bαjyj=cscale(¯yj−min¯yj)+βjαj∼N(0,1)βj∼HalfNormal(0,σβ)

Where αj is a column vector of spline coefficients and βj serves as intercept, all are treated as parameters. The scaling constant cscale and σβ reflect the range of the data and are given by the user. When the regulator proteins are estimated from the effect on targets only (no expression measurements for regulator), the spline intercept βj is ignored, as it cannot be separated from bi.

Solving the ODE

It is a bit tricky to use Stan’s built-in ODE solver with a spline as one of the parameters driving ODE evolution as it needs the value of yi at arbitrary time points. The most straightforward way - linearly interpolating between precomputed discrete values of yi requires some hacks and breaks the solver. In principle, αi could be given to the ODE solver and the spline basis computed for each timepoint the solver needs, but this seemed cumbersome. Instead we decided to follow the approach of (Titsias et al. 2012). First we can observe that the regulation ODE is linear in xi and can be solved for t≥0:

xi(t)=ηie−dit+si∫t0F(ρi(u))e−di(t−u)du

χi(0)=−12F(ρi(0))χi(t+1)=(χi(t)+F(ρi(t)))dixi(t)=ηie−dit+si(χi(t)+12F(ρi(t)))

Identifiability

The model as given above is biologically relevant, but has multiple identifiability issues, all of which can be demonstrated with just a single regulator and a single target, so we will drop parameter indicies for the rest of this section. The first non-identifiability arises with w and b if all values of ρ fall on the tail of the sigmoid, which is approximately linear. In this case, the walue of F(ρ) becomes insensitive to linear transformations of w and b:

time <- seq(0,2,length.out = 100)
regulator <- sin(time * 4) + 1

params1 <- c(degradation = 0.5, bias = -3, sensitivity = 1, weight = 5, basal_transcription = 0, protein = approxfun(time, regulator, rule=2));
target1 <-  ode( y = c(x = 0), times = time, func = target_ODE, parms = params1, method = "ode45")[,"x"];

params2 <- c(degradation = 0.5, bias = -6, sensitivity = 1, weight = 10, basal_transcription = 0, protein = approxfun(time, regulator, rule=2));
target2 <-  ode( y = c(x = 0), times = time, func = target_ODE, parms = params2, method = "ode45")[,"x"];

params3 <- c(degradation = 0.5, bias = -30, sensitivity = 1, weight = 50, basal_transcription = 0, protein = approxfun(time, regulator, rule=2));
target3 <-  ode( y = c(x = 0), times = time, func = target_ODE, parms = params3, method = "ode45")[,"x"];

params_to_legend <- function(params) {
  paste0("w = ", params["weight"], ", b = ", params["bias"])
}

data.frame(time = time, regulator = regulator, target1 = target1, target2 = target2, target3 = target3) %>%
  gather("profile","expression", -time) %>%
  ggplot(aes(x = time, y = expression, color = profile)) + 
  geom_line() + 
  scale_color_hue(labels = c("regulator", params_to_legend(params1), params_to_legend(params2), params_to_legend(params3))) + ggtitle("Non-identifiability due to w and b","All targets have s = 1, d = 0.5")

Transformations of s and d together may also have very minor effect on the resulting expression. While the resulting expression profiles are more different than in the previous case, those differences are still very small compared to the noise observed in real data. In the following figures, the red dots represent measurements that have approximately equal likelihood for each of the shown true target profiles.

time <- seq(0,2,length.out = 100)
regulator <- sin(time * 4) + 1

params1 <- c(degradation = 5, bias = -1, sensitivity = 10, weight = 1, basal_transcription = 0, protein = approxfun(time, regulator, rule=2));
target1 <-  ode( y = c(x = 1), times = time, func = target_ODE, parms = params1, method = "ode45")[,"x"];

params2 <- c(degradation = 10, bias = -1, sensitivity = 19, weight = 1, basal_transcription = 0, protein = approxfun(time, regulator, rule=2));
target2 <-  ode( y = c(x = 1), times = time, func = target_ODE, parms = params2, method = "ode45")[,"x"];

params3 <- c(degradation = 100, bias = -1, sensitivity = 180, weight = 1, basal_transcription = 0, protein = approxfun(time, regulator, rule=2));
target3 <-  ode( y = c(x = 1), times = time, func = target_ODE, parms = params3, method = "ode45")[,"x"];

params_to_legend <- function(params) {
  paste0("s = ", params["sensitivity"], ", d = ", params["degradation"])
}

measured_data = data.frame(profile = "measured", time = seq(0,2,length.out = 9), 
                           expression = c(1,1.7,1.4,0.85, 0.83,0.35,0.7,1.4,1.5))


data.frame(time = time, regulator = regulator, target1 = target1, target2 = target2, target3 = target3) %>%
  gather("profile","expression", -time) %>%
  ggplot(aes(x = time, y = expression, color = profile)) + geom_line() + 
  geom_point(data = measured_data, color = "#ba1b1d", size = 3) +
  scale_color_hue(labels = c("regulator",params_to_legend(params1), params_to_legend(params2), params_to_legend(params3))) + ggtitle("Non identifiability due to s and d","All targets have w = 1, b = -1")

There is also a non-identifiability when w=0 as s and b become redundant. Even worse, some of the solutions with w=0 might not be distinguishable form solutions with |w|≫0. This might also induce non-identifiability in the initial condition η as it is much less constrained by data. Once again the expression profiles are not identical, but are close enough that given the amount of noise, non-negligible posterior mass can be found for all of them:

time <- seq(0,2,length.out = 100)
regulator <- sin(time * 4) + 1

params1 <- c(degradation = 3, bias = -1, sensitivity = 3, weight = 5, basal_transcription = 0, protein = approxfun(time, regulator, rule=2));
target1 <-  ode( y = c(x = 1.5), times = time, func = target_ODE, parms = params1, method = "ode45")[,"x"];

params2 <- c(degradation = 3.4, bias = 0, sensitivity = 6, weight = 0, basal_transcription = 0, protein = approxfun(time, regulator, rule=2));
target2 <-  ode( y = c(x = 1.6), times = time, func = target_ODE, parms = params2, method = "ode45")[,"x"];

params3 <- c(degradation = 10, bias = 0, sensitivity = 16, weight = 0, basal_transcription = 0, protein = approxfun(time, regulator, rule=2));
target3 <-  ode( y = c(x = 2.5), times = time, func = target_ODE, parms = params3, method = "ode45")[,"x"];


params_to_legend <- function(params) {
  paste0("s = ", params["sensitivity"],", w = ", params["weight"], ", b = ", params["bias"], ", d = ", params["degradation"])
}

measured_data = data.frame(profile = "measured", time = seq(0,2,length.out = 9), 
                           expression = c(1.8,0.9,1,0.66, 1.11,0.6,0.7,0.96,1.03))


data.frame(time = time, regulator = regulator, target1 = target1, target2 = target2, target3 = target3) %>%
  gather("profile","expression", -time) %>%
  ggplot(aes(x = time, y = expression, color = profile)) + geom_line() + 
  geom_point(data = measured_data, color = "#ba1b1d", size = 3) +
  scale_color_hue(labels = c("regulator",params_to_legend(params1), params_to_legend(params2), params_to_legend(params3))) + ggtitle("Non identifiability between w=0 and w > 0")

Last but not least, it is also possible to get similar solutions with different sign of w:

time <- seq(0,2,length.out = 100)
regulator <- c(0.3296698,0.6667181,1.0083617,1.3518170,1.6943005,2.0330289,2.3652186,2.6880862,2.9988482,3.2947213,3.5729219 ,3.8306665,4.0651718,4.2736542,4.4533304,4.6014168,4.7154822,4.7961600,4.8456619,4.8662125,4.8600365,4.8293585 ,4.7764032,4.7033952,4.6125592,4.5061198,4.3863016,4.2553294,4.1154277,3.9688212,3.8177346,3.6643924,3.5110195 ,3.3598403,3.2130795,3.0725728,2.9385992,2.8110488,2.6898117,2.5747779,2.4658374,2.3628803,2.2657967,2.1744767 ,2.0888102,2.0086873,1.9339981,1.8646326,1.8004810,1.7414331,1.6873792,1.6382092,1.5938132,1.5540812,1.5189034 ,1.4881697,1.4617702,1.4395950,1.4215341,1.4074776,1.3973155,1.3909378,1.3882347,1.3890962,1.3934123,1.4010732 ,1.4119687,1.4259843,1.4428964,1.4623727,1.4840759,1.5076691,1.5328149,1.5591763,1.5864163,1.6141975,1.6421830 ,1.6700355,1.6974179,1.7239932,1.7494242,1.7733737,1.7955046,1.8154798,1.8329622,1.8476146,1.8590998,1.8670809 ,1.8712205,1.8711817,1.8666272,1.8572200,1.8426229,1.8224987,1.7965104,1.7643208,1.7255927,1.6799891,1.6271729 ,1.5668068) * 0.5

params1 <- c(degradation = 10, bias = -8, sensitivity = 15, weight = 10, basal_transcription = 0, protein = approxfun(time, regulator, rule=2));
target1 <-  ode( y = c(x = 0), times = time, func = target_ODE, parms = params1, method = "ode45")[,"x"];

params2 <- c(degradation = 1, bias = 3, sensitivity = 100, weight = -10, basal_transcription = 0, protein = approxfun(time, regulator, rule=2));
target2 <-  ode( y = c(x = 0), times = time, func = target_ODE, parms = params2, method = "ode45")[,"x"];


params_to_legend <- function(params) {
  paste0("s = ", params["sensitivity"],", w = ", params["weight"], ", b = ", params["bias"], ", d = ", params["degradation"])
}

measured_data = data.frame(profile = "measured", time = seq(0,2,length.out = 9), 
                           expression = c(0.05,1.52,1.44,0.83, 0.96,1.01,0.72,1.07,0.65))


data.frame(time = time, regulator = regulator, target1 = target1, target2 = target2) %>%
  gather("profile","expression", -time) %>%
  ggplot(aes(x = time, y = expression, color = profile)) + geom_line() + 
  geom_point(data = measured_data, color = "#ba1b1d", size = 3) +
  guides(color=FALSE) + ggtitle("Non identifiability between w < 0 and w > 0")

There are probably many more non-identified scenarios, especially when multiple targets are involved, but the above are of the type we have actually encountered with real data.

Workflow

The assumed workflow when using the model to infer novel regulations proceeds as follows:

1. Gather putative and known targets of the regulators of interest
2. Fit the constant synthesis model to all known & putative targets
3. Assess model fit and discard targets that are fit well
4. Optional: Use the known targets to constrain the true expression of the regulators
5. Fit the main model to each putative target separately
6. Compare the fit with the main model to the fit of the constant synthesis model

In the following example, we will use 5 known and 4 putative targets of the sigA regulator. We use cubic spline with 6 degrees of freedom, cscale=5 as the basis for the main model.

#Globally used params for the algorithm
measurement_times = gse6865_raw_time + 1
smooth_time  <- 1:101#seq(0,100, by = 1)
expression_data <- gse6865_raw

spline_df <- 6
spline_basis <- bs(smooth_time, degree = 3, df = spline_df)
default_spline_params <- spline_params(
  spline_basis = spline_basis,
  scale = 5
)

default_params_prior <-  params_prior(
  initial_condition_prior_sigma = 2,
  asymptotic_normalized_state_prior_sigma = 2,
  degradation_prior_mean = -2,
  degradation_prior_sigma = 1,
  mean_regulatory_input_prior_sigma = 5,
  sd_regulatory_input_prior_sigma =5,
  intercept_prior_sigma = 2
)

default_measurement_sigma = measurement_sigma_given(0.1,0.1)

putative_targets <- c("kinE","purT", "yhdL", "codV")

#The known targets are those predicted and biologically validated in our previous work with the dataset (https://doi.org/10.1016/j.bbagrm.2017.06.003)
known_targets <- c("acpA","fbaA","rpmGA","ykpA","yyaF")


plot_profiles <- function(expression_data, targets) {
  expression_data[targets,,drop = FALSE] %>% as.data.frame() %>% 
    rownames_to_column("gene") %>% 
    gather("time","expression",-gene) %>%
    mutate(time = as.integer(gsub("min","",time, fixed = TRUE))) %>%
    ggplot(aes(x = time, y = expression, color = gene, linetype = gene)) + geom_line()
}

plot_profiles(expression_data, "sigA") + ggtitle("Regulator")

plot_profiles(expression_data, known_targets) + ggtitle("Known targets")

plot_profiles(expression_data, putative_targets) + ggtitle("Putative targets")

csynth_model <- stan_model(file = here('stan','constant_synthesis.stan'))

targets <- c(putative_targets)
data_csynth <- list()
fits_csynth <- list()
loo_csynth <- list()
for(t in 1:length(targets)) {
  data_csynth[[t]] <- list(
    num_measurements = length(measurement_times),
    measurement_times = measurement_times,
    expression = expression_data[targets[t],],
    measurement_sigma_absolute = default_measurement_sigma$sigma_absolute_data[1],
    measurement_sigma_relative = default_measurement_sigma$sigma_relative_data[1],
    initial_condition_prior_sigma = default_params_prior$initial_condition_prior_sigma,
    asymptotic_normalized_state_prior_sigma = default_params_prior$asymptotic_normalized_state_prior_sigma,
    degradation_prior_mean = default_params_prior$degradation_prior_mean,
    degradation_prior_sigma = default_params_prior$degradation_prior_sigma
  )
  fits_csynth[[t]] <- sampling(csynth_model, data_csynth[[t]])
  loo_csynth[[t]] <- get_loo_csynth(fits_csynth[[t]])
}

plots <- list()
for(t in 1:length(targets)) {
  fit <- fits_csynth[[t]]
  plot1 <- fitted_csynth_plot(fit, data_csynth[[t]], name = targets[t])
  plot2 <- fitted_csynth_observed_plot(fit, data_csynth[[t]], name = targets[t])
  plot_grid(plot1, plot2, ncol = 2) %>% print()
}

get_ic_estimate <- function(x) {
  x %>% map_dbl(function(x) { x$estimates["looic","Estimate"]})
}


csynth_table <- tibble(target = targets, looic_csynth = get_ic_estimate(loo_csynth))
csynth_table 

## # A tibble: 4 x 2
##   target looic_csynth
##   <chr>         <dbl>
## 1 kinE          -9.24
## 2 purT          53.9 
## 3 yhdL          56.1 
## 4 codV         136.

Estimating regulator expression from known targets

A problem with the main model is that the data constrain the true expression quite weakly. This is not an issue for the putative targets, but becomes problematic for the expression of the regulators. When trying to determine new regulations, only a single target at a time needs to be fit, because the model assumes that all the regulations actually take place and thus a single regulation that is not, in fact, correct may shift the parameter values (especially the spline params α) considerably. While it would be in principle possible to marginalize over the power set of targets to get estimates of probabilities of the individual regulations, this is computationally infeasible. The downside of fitting each target separately is that the estimated expression of regulators may not be consistent across targets, possibly leading to false positives.

Similarly to (Titsias et al. 2012), our model is able to use known regulations to reduce the uncertainty in regulator expression, hopefully elminating any gross inconsistencies between individual fits for putative targets. This is a simple side-effect of the Bayesian treatment, where after fitting the model with all the known targets at once, we can extract the posterior for y and use it as input when fitting putative targets. This step is however optional and if no targets are known with high certainty, it can be skipped.

First, let us see the posterior samples from fitting the regulator with 0 targets (equivalent to just fitting a spline). The red dots are the measured values:

#Load the model
regulated_model <- stan_model(file = here('stan','regulated.stan'))

set.seed(4127785)
source <- "sigA"

data <- regulated_model_params(
  measurement_times = measurement_times,
  regulator_expression = expression_data[source,],
  measurement_sigma = default_measurement_sigma,
  spline_params = default_spline_params,
  params_prior = default_params_prior
)


fit_regulator_spline_only <- sampling(regulated_model, data = data, control = list(adapt_delta = 0.95))

fitted_regulator_plot(fit_regulator_spline_only, data, name = paste0(source, " spline only"))

While the exact scaling of the regulator expression should not affect model results much, we see that there is also qualitative uncertainty, for example in the number of local minima and maxima after the 60 minute mark.

And this is what the posterior looks like when both regulator expression and 5 known targets are used. Note that unlike using only splines, the profiles are now very similar especially in that local minima and maxima are at almost the same time point:

set.seed(751235428)


data <- regulated_model_params(
  measurement_times = measurement_times,
  regulator_expression = expression_data[source,],
  target_expression = t(expression_data[known_targets,,drop = FALSE]),
  regulation_signs = matrix(1, 1, length(known_targets)),
  measurement_sigma = default_measurement_sigma,
  spline_params = default_spline_params,
  params_prior = default_params_prior
)


fit_regulator_all_info <- sampling(regulated_model, data = data, control = list(adapt_delta = 0.95))

fitted_regulator_plot(fit_regulator_all_info, data, name = paste0(source, " all information"))

# for(target in 1:length(training_targets)) {
#   fitted_target_plot(fit_regulator_all_info, data, target = target, name = training_targets[target]) %>% print()
# }

pairs(fit_regulator_all_info, pars = "coeffs")

samples_coeffs <- rstan::extract(fit_regulator_all_info,"coeffs")$coeffs[,,1]
mvn_test_results <- mvn(samples_coeffs, mvnTest = "mardia", desc = FALSE, multivariatePlot = "qq")

mvn_test_results$multivariateNormality 

##              Test        Statistic               p value Result
## 1 Mardia Skewness 1080.51030587331 1.36639895742596e-189     NO
## 2 Mardia Kurtosis 15.1620214384148                     0     NO
## 3             MVN             <NA>                  <NA>     NO

n_samples <- 100
mvn_unscaled <- coeffs_prior_from_fit(fit_regulator_all_info, covariance_scale = 1)
cov_scale <- 0.5

means_array <- t(array(
  rep(mvn_unscaled$coeffs_prior_mean,n_samples),c(length(mvn_unscaled$coeffs_prior_mean), n_samples)))


samples_learned <- array(rnorm(length(mvn_unscaled$coeffs_prior_mean) * n_samples, 0, 1), c(n_samples, length(mvn_unscaled$coeffs_prior_mean))) %*% chol(mvn_unscaled$coeffs_prior_cov[1,,]) + means_array

samples_learned_scaled <- array(rnorm(length(mvn_unscaled$coeffs_prior_mean) * n_samples, 0, 1), c(n_samples, length(mvn_unscaled$coeffs_prior_mean))) %*% chol(mvn_unscaled$coeffs_prior_cov[1,,] * cov_scale) + means_array


limits <- ylim(c(0,6.5))
plot1 <- fitted_regulator_plot(fit_regulator_all_info, data, name = paste0(source, " all information"), num_samples = n_samples) + limits

measured_geom <- geom_point(data = data.frame(x = data$measurement_times,  y = data$regulator_expression), aes(x=x, y=y), inherit.aes = FALSE, color = "#ba1b1d", size = 3)

plot2 <- ggmatplot(1:data$num_time, t(samples_learned %*% t(data$spline_basis)) * data$scale, main_geom = default_expression_plot_main_geom) + measured_geom + ggtitle("MVN approximation unscaled") + limits

plot3 <-ggmatplot(1:data$num_time, t(samples_learned_scaled %*% t(data$spline_basis)) * data$scale, main_geom = default_expression_plot_main_geom) + measured_geom  + ggtitle(paste0("MVN approximation, scale = ", cov_scale)) + limits

plot_grid(plot1, plot2, NULL, plot3, nrow = 2)

targets <- c("purT", "yhdL", "codV")

data_regulated <- list()
fits_regulated <- list()
loo_regulated <- list()
coeffs_prior <- coeffs_prior_from_fit(fit_regulator_all_info, covariance_scale = 1)
for(t in 1:length(targets)) {
  data_regulated[[t]] <- regulated_model_params(
    measurement_times = measurement_times,
    target_expression = t(expression_data[targets[t],,drop = FALSE]),
    regulation_signs = matrix(1, 1, 1),
    measurement_sigma = default_measurement_sigma,
    spline_params = default_spline_params,
    params_prior = default_params_prior,
    coeffs_prior = coeffs_prior
  )  
  fits_regulated[[t]] <- sampling(regulated_model, data_regulated[[t]])
  loo_regulated[[t]] <- get_loo_genexpi(fits_regulated[[t]], target = 1)
}

for(t in 1:length(targets)) {
  fit <- fits_regulated[[t]]
  plot1 <-  fitted_target_plot(fit, data_regulated[[t]], name = targets[t])
  plot2 <- fitted_target_observed_plot(fit, data_regulated[[t]], name = targets[t])
  plot_grid(plot1, plot2, ncol = 2) %>% print()
}

data_free <- list()
fits_free <- list()
loo_free <- list()
for(t in 1:length(targets)) {
  data_free[[t]] <- regulated_model_params(
    measurement_times = measurement_times,
    target_expression = t(expression_data[targets[t],,drop = FALSE]),
    regulation_signs = matrix(1, 1, 1),
    measurement_sigma = default_measurement_sigma,
    spline_params = default_spline_params,
    params_prior = default_params_prior
  )  
  fits_free[[t]] <- sampling(regulated_model, data_free[[t]])
  loo_free[[t]] <-  get_loo_genexpi(fits_free[[t]], target = 1)
}

for(t in 1:length(targets)) {
  fit <- fits_free[[t]]
  plot1 <-  fitted_target_plot(fit, data_regulated[[t]], name = targets[t])
  plot2 <- fitted_regulator_plot(fit, data_regulated[[t]], name = targets[t])
  plot_grid(plot1, plot2, ncol = 2) %>% print()
}

regulated_table <- tibble(target = targets, looic_regulated = get_ic_estimate(loo_regulated), looic_free = get_ic_estimate(loo_free)) %>%
  left_join(csynth_table, by = "target")
regulated_table  

## # A tibble: 3 x 4
##   target looic_regulated looic_free looic_csynth
##   <chr>            <dbl>      <dbl>        <dbl>
## 1 purT              50.8       25.0         53.9
## 2 yhdL              22.1       27.1         56.1
## 3 codV              42.1       62.3        136.

Exploratory analysis with multiple regulators

Since purT is likely not regulated by sigA, we might try a combination of multiple regulators. Unless the dataset is much larger than the one we are using here, this is highly speculative: many combinations of two regulators are able to fit a majority of all genes quite well.

regulators = c("sigA", "sigB")
targets = "purT"

data_two_reg <- regulated_model_params(
    measurement_times = measurement_times,
    regulator_expression = t(expression_data[regulators,,drop = FALSE]),
    target_expression = t(expression_data[targets,,drop = FALSE]),
    regulation_signs = matrix(1, 2, 1),
    measurement_sigma = default_measurement_sigma,
    spline_params = default_spline_params,
    params_prior = default_params_prior
  )  

fit_two_reg <- fit_regulated(data_two_reg, regulated_model)

Inspecting the fit visually and comparing the looic in this case we see that the regulation by sigA and sigB is plausible.

plot1 <- fitted_target_plot(fit_two_reg, data_two_reg)
plot2 <- fitted_target_observed_plot(fit_two_reg, data_two_reg)
plot_grid(plot1, plot2, ncol = 2)

regulated_table %>% filter(target == "purT") %>% mutate(looic_two_reg = get_loo_genexpi(fit_two_reg)$estimates["looic","Estimate"]) 

## # A tibble: 1 x 5
##   target looic_regulated looic_free looic_csynth looic_two_reg
##   <chr>            <dbl>      <dbl>        <dbl>         <dbl>
## 1 purT              50.8       25.0         53.9          28.6

Conclusions

While the model as presented already coversthe full workflow to infer novel regulations, there are some issues that need to be ironed out before we can rely on it. Most notably, we are yet to perform a thorough evaluation and comparison with other tools for the same task. Initial results were actually comparable to a simpler maximum likelihood variant of the model with separate splining and parameter fitting phases (as presented in Modrák and Vohradský 2018). We are currently working to understand why there is not a marked improvement, as we know that some of the incorrect results of the simpler version stemmed from separate splining.

Further workflow steps we are not happy with are the way we fit a multivariete normal distribution to the posterior of regulator expression, the arbitrary way we use looic to make decisions, and convergence issues that sometimes arise when running the model without observed regulator expression on real data. There is also substantial work to be done in encapsulating the model as a package and providing a cleaner interface.

Acknowledgements

Huge thanks belong to the Stan development team and the Stan community. Without the help I got from their tranining materials and answers on the forums, I would not be able to move this project forward.

This work was supported by C4Sys research infrastructure project (MEYS project No: LM20150055).

Session Info

sessionInfo()

## R version 3.4.4 (2018-03-15)
## Platform: x86_64-w64-mingw32/x64 (64-bit)
## Running under: Windows 10 x64 (build 17134)
## 
## Matrix products: default
## 
## locale:
## [1] LC_COLLATE=English_United States.1252 
## [2] LC_CTYPE=English_United States.1252   
## [3] LC_MONETARY=English_United States.1252
## [4] LC_NUMERIC=C                          
## [5] LC_TIME=English_United States.1252    
## 
## attached base packages:
## [1] parallel  splines   stats     graphics  grDevices utils     datasets 
## [8] methods   base     
## 
## other attached packages:
##  [1] bindrcpp_0.2.2     MVN_5.5            here_0.1          
##  [4] forcats_0.3.0      stringr_1.3.1      dplyr_0.7.5       
##  [7] purrr_0.2.5        readr_1.1.1        tidyr_0.8.1       
## [10] tibble_1.4.2       tidyverse_1.2.1    cowplot_0.9.2     
## [13] genexpiStan_0.1.0  loo_2.0.0          truncnorm_1.0-8   
## [16] deSolve_1.21       rstan_2.17.3       StanHeaders_2.17.2
## [19] ggplot2_3.0.0      knitr_1.20        
## 
## loaded via a namespace (and not attached):
##   [1] colorspace_1.3-2      mvoutlier_2.0.9       modeltools_0.2-21    
##   [4] class_7.3-14          rio_0.5.10            mclust_5.4           
##   [7] rprojroot_1.3-2       pls_2.6-0             rstudioapi_0.7       
##  [10] roxygen2_6.0.1        cvTools_0.3.2         flexmix_2.3-14       
##  [13] mvtnorm_1.0-8         lubridate_1.7.4       xml2_1.2.0           
##  [16] codetools_0.2-15      sROC_0.1-2            mnormt_1.5-5         
##  [19] robustbase_0.93-0     jsonlite_1.5          robCompositions_2.0.8
##  [22] broom_0.4.4           kernlab_0.9-26        cluster_2.0.6        
##  [25] rrcov_1.4-4           compiler_3.4.4        httr_1.3.1           
##  [28] backports_1.1.2       Matrix_1.2-12         assertthat_0.2.0     
##  [31] lazyeval_0.2.1        cli_1.0.0             htmltools_0.3.6      
##  [34] tools_3.4.4           gtable_0.2.0          glue_1.2.0           
##  [37] reshape2_1.4.3        Rcpp_0.12.17          carData_3.0-1        
##  [40] trimcluster_0.1-2     cellranger_1.1.0      zCompositions_1.1.1  
##  [43] sgeostat_1.0-27       nlme_3.1-131.1        fpc_2.1-11           
##  [46] lmtest_0.9-36         psych_1.8.4           laeken_0.4.6         
##  [49] openxlsx_4.1.0        rvest_0.3.2           devtools_1.13.5      
##  [52] DEoptimR_1.0-8        zoo_1.8-1             MASS_7.3-49          
##  [55] scales_0.5.0          VIM_4.7.0             hms_0.4.2            
##  [58] inline_0.3.15         RColorBrewer_1.1-2    yaml_2.1.19          
##  [61] curl_3.2              NADA_1.6-1            memoise_1.1.0        
##  [64] gridExtra_2.3         reshape_0.8.7         stringi_1.1.7        
##  [67] nortest_1.0-4         pcaPP_1.9-73          e1071_1.6-8          
##  [70] energy_1.7-4          boot_1.3-20           zip_1.0.0            
##  [73] prabclus_2.2-6        rlang_0.2.1           pkgconfig_2.0.1      
##  [76] commonmark_1.5        moments_0.14          matrixStats_0.53.1   
##  [79] evaluate_0.10.1       lattice_0.20-35       bindr_0.1.1          
##  [82] labeling_0.3          tidyselect_0.2.4      GGally_1.4.0         
##  [85] plyr_1.8.4            magrittr_1.5          R6_2.2.2             
##  [88] pillar_1.2.3          haven_1.1.1           foreign_0.8-69       
##  [91] withr_2.1.2           survival_2.41-3       abind_1.4-5          
##  [94] sp_1.3-1              nnet_7.3-12           modelr_0.1.2         
##  [97] crayon_1.3.4          car_3.0-0             utf8_1.1.4           
## [100] KernSmooth_2.23-15    rmarkdown_1.9         grid_3.4.4           
## [103] readxl_1.1.0          data.table_1.11.4     diptest_0.75-7       
## [106] vcd_1.4-4             digest_0.6.15         stats4_3.4.4         
## [109] munsell_0.4.3         viridisLite_0.3.0     kableExtra_0.9.0