\usepackage[table,xcdraw]{xcolor}

\usepackage{hyperref}

\usepackage{svg}

\begin{document}

The main advantage of considering epigenetics data for the task of TF binding prediction is that the number of false positive predictions can be reduced \cite{pmid21106904}.

One way of incorporating epigenetics data is to reduce the genomic search space to a few candidate regions of TF binding.

As shown before, genome-wide candidate sites for TF binding can be determined by open-chromatin experiments \cite{pmid25294828,pmid25086003,pmid22072382,pmid23424114}, e.g. peaks or footprints in DNase1-seq data,

and/or by considering Histone marks \cite{pmid25489339,pmid25086003}, e.g. H3K4me3.

A major advantage of affinity based predictions compared to hit-based methods like Fimo \cite{Grant16022011} is that

low-affinity binding sites can be included \cite{pmid27899623,pmid17098775}. Using the \textit{TEPIC} method, we compute TF gene scores by aggregating TF predictions calculated for a user defined set of candidate regions.

The scores, either per peak/region or gene, can be interpreted as a quantitative measurement of TF binding.

We obtained \textit{Position Count Matrices (PCMs)} from JASPAR \cite{pmid26531826}, which is also including data from Uniprobe \cite{pmid25378322}, HOCOMOCO \cite{pmid23175603} and the Kellis Lab ENCODE Motif database \cite{pmid24335146}.

There are three folder containing Position specific energy matrices (PSEMs): Our current collection of PSEMs \textit{PWMs/2.1}. The previously used motifs are provided in the folders \textit{PWMs/2.0} and \textit{PWMs/1.0}.

\begin{itemize}

\item 579 PSEMs for vertebrates

\item 176 PSEMs for fungi

Files holding the length of the PSEMs are provided too.

As mentioned above, \textit{TRAP} computes TF affinities that are based on a biophysical model of TF binding.

Therefore \textit{PCMs} have to be converted to \textit{Position Specific Energy Matrices (PSEMs)} such that they can be used in \textit{TRAP}.

Intuitively, \textit{PSEMs} represent the mismatch energy of a given motif. For a detailed explanation and motivation of the energy based score, please check \cite{pmid17098775}.

\end{itemize}

In all other cases, a default GC-content of $0.42$ is used.

that could be found in the reference genomes of the respective species and

overlap with a window of user defined size $w$ that is centered at the most $5'$ TSS of all annotated genes in the considered organism.

The contribution of the individual sites is weighted by their distance to the selected TSS with an exponential decay function \cite{pmid19995984}.

Formally, the TF gene score $a_{g,i}$ for gene $g$ and TF $i$ is computed as

\begin{align}

where $s_p$ is the per base signal in peak $p$. This computation can be done with and without length normalisation of the affinities.

The workflow of TEPIC is depicted in Figure \ref{workflowFig}.

\begin{figure}[h!]

\begin{center}

\includegraphics[width=\textwidth]{Workflow.png}

\label{workflowFig}

\end{figure}

To compute TF gene scores a user needs to specify:

\begin{itemize}

\end{itemize}

Special care should be taken for \textit{caenorhabditis elegans}, as Roman digits are used for enumeration of chromosomes.

\begin{enumerate}

\end{enumerate}

Epigenetics data contains a wealth of information on gene regulation. It was shown that especially

data on open-chromatin is well suited to build predictive models of gene-expression \cite{pmid27899623,pmid22955983,pmid25231769,pmid22954627}.

Interpreting these models allows the inference of regulators that may play a key role in gene-expression regulation.

\item Learning a linear regression model to predict gene expression from TF gene scores computed in (1).

\end{enumerate}

In order to learn about potentially important regulators, we build a linear, interpretable regression model,

comparable to methods proposed in \cite{pmid27899623,pmid22955983,pmid25231769,pmid22954627}.

Here, we use TF gene scores computed with \textit{TEPIC} as features in a linear regression setup to predict gene expression.

Details on the learning setup and on the available regularization methods are provided in the next section.

We offer three different regularization techniques:

\begin{itemize}

\item Lasso:

It resolves the correlation between features by distributing the feature weights among them, and simultaneously leads to sparse and stable models \cite{Zou05regularizationand}.

However, learning a model using elastic net penalty is slower than using either only Lasso or Ridge regularization.

The data matrix $X$, containing TF gene scores, and the response vector $y$, containing gene expression values, are log-transformed,

Regression coefficients are computed in a inner cross validation,

the $\alpha$ parameter of elastic net regularization is optimized with a default step size of $0.1$.

\end{enumerate}

All parameters mentioned in this section can be changed by the user. The learning process is sketched in Figure \ref{learningFig}.

In addition to the input required for the computation of TF gene scores in TEPIC, a file containing gene expression data must be provided.

This file should be structured such that column $1$ contains the gene identifiers and column $2$ holds expression values.

The user is always provided with the following files:

\begin{itemize}

\item a list of regression coefficients computed on the entire data set,

\end{figure}

Although a variety of methods have been proposed to generate genome-wide TF binding predictions [\cite{pmid27899623,pmid22072382,pmid23424114,pmid25086003}] and to establish \textit{TF to tissue}

associations [\cite{pmid22955983,pmid19995984,pmid27899623}], systematic, feasible, and easy to use ways of linking TFs to distinct genes are rare.

In addition to the \textit{INVOKE} analysis, we propose a method to infer the most likely transcriptional regulators for a set of differentially expressed genes.

We use TF scores, computed using \textit{TEPIC}, and logistic regression to identify TFs that have explanatory power to distinguish between up- and down-regulated genes.

To run \textit{DYNAMITE}, a user most provide candidate regions of TF binding for two groups of samples, $A$ and $B$, e.g. control and disease.

These can be derived, for example, by open chromatin experiments such as DNase-seq.

It is essential that the candidate regions reflect the characteristics of chromatin organization in the analysed tissues.

In addition, a list of differential expressed genes between two groups as well as log2 fold changes of the expression are needed.

Our method consits of two parts: (1) gene-TF score computation, and (2) identification of key TFs.

\subsection*{Step 1: Computing Gene-TF Scores}

Using TEPIC, we compute gene-TF scores $g_{ij}$ for all differentially expressed genes $i$ and distinct TFs $j$ considering the provided candidate regions for all replicates $a$ of group $A$ and for all replicates $b$ of group $B$.

\caption{Computation of differential TF features between two groups.}

\label{TF-Gene-Score_Computation}

\end{figure}

To identify those TFs that can explain the differential expression state of as many genes as possible, we build a logistic regression classifier.

We use matrix $R_{AB}$ computed in Step $1$ as the feature matrix $X$, and a binary vector of gene expression changes as response $y$.

An example is shown in Figure \ref{Log-Reg-Example}.

same learning paradigm that is described for the \textit{INVOKE} analysis (Figure \ref{learningFig}). We use the entire dataset for model training and to interpret the regression coefficients.

TFs that correspond to features with a non-zero regression coefficient can be seen as being essential to explain the observed expression differences and should be further investigated.

Model performance is reported in a \textit{txt} file and visually in a bar plot using mean test and training accuracy as well as the F1 measure.

A heatmap shows the regression coefficients in the outer cross validation folds.

Additionaly, we report confusion matrices for the outer cross validation folds.

\end{figure}

\textit{EPIC-DREM} is a combination of \textit{TEPIC} and the \textit{Dynamic Regulatory Events Miner (DREM)} \cite{pmid17224918}.

Instead of using static ChIP-seq data, which is provided in \textit{DREM 2.0}, we suggest to use time-point specific TF binding predictions

based on time-dependent epigenomic profiles. Thereby, \textit{DREM} can infer regulators that can be linked to expression changes at distinct points in time.

\label{epicdrem}

\end{figure}

In some applications it is required to make a binary decision whether a factor is binding or not.

To infer this information from TF affinities, \textit{TEPIC} allows the computation of a TF specific affinity threshold by calculating TF affinities on a randomly selected set of genomic regions. When selected by TEPIC, these regions show similar characteristics compared to the provided regions (GC content and length).

Alternatively a set of background regions can be provided by the user.

By applying a user defined p-value on the distribution of affinities computed on the random regions, a threshold is chosen.

Per TF, all affinities that are smaller than the selected threshold, are set to zero, thus a sparse matrix with TF-gene interactions can be generated.

In addition to the input mentioned above, a reference genome in 2bit format is required.

Optionally, the user can provide a bed file containing background regions.

These replace the automated generation of background sequences.

The following output files are generated in addition:

\begin{enumerate}

\item TF affinities for all selected \textit{PSEMs} in the regions provided by the user that passed the filtering step, where all affinities below the TF specific thresholds are set to 0.

\end{enumerate}

Either (2) or (3) can be combined with RNA-seq data and used as input for \textit{DREM}.

\bibliographystyle{plain}

\bibliography{Description}

\end{document}