\@writefile{lof}{\contentsline{figure}{\numberline{5.1}{\ignorespaces\textbf{SMiLE-seq pipeline :}\textbf{a} Schematic representation of the experimental setup. A snapshot of three units of the microfluidic device is shown. In vitro transcribed and translated bait TF, target dsDNA, and a nonspecific competitor poly-dIdC are mixed and pipetted in one of the wells of the microfluidic device. The mixtures are then passively pumped in the device (bottom panel). Newly formed TF\IeC{\textendash}DNA complexes are trapped under a flexible polydimethylsiloxane membrane, and unbound molecules as well as molecular complexes are washed away (upper panel). Left, schematic representation of three individual chambers. Right, corresponding snapshots of an individual chamber taken before and after mechanical trapping. \textbf{b} Data processing pipeline. The bound DNA is eluted from all the units of the device simultaneously and collected in one tube. Recovered DNA is amplified and sequenced. The sequencing reads are then demultiplexed, and a seed sequence is identified for each sample. This seed is then used to initialize a probability matrix representing the sequence specificity model for the given TF. The model parameters are then optimized using a Hidden Markov Model-based motif discovery pipeline. Figure and legend taken and adapted from \citep{isakova_smile-seq_2017}.\relax}}{68}{figure.caption.29}}
\newlabel{smile_seq_pipeline}{{5.1}{68}{\textbf{SMiLE-seq pipeline :}\textbf{a} Schematic representation of the experimental setup. A snapshot of three units of the microfluidic device is shown. In vitro transcribed and translated bait TF, target dsDNA, and a nonspecific competitor poly-dIdC are mixed and pipetted in one of the wells of the microfluidic device. The mixtures are then passively pumped in the device (bottom panel). Newly formed TF–DNA complexes are trapped under a flexible polydimethylsiloxane membrane, and unbound molecules as well as molecular complexes are washed away (upper panel). Left, schematic representation of three individual chambers. Right, corresponding snapshots of an individual chamber taken before and after mechanical trapping. \textbf{b} Data processing pipeline. The bound DNA is eluted from all the units of the device simultaneously and collected in one tube. Recovered DNA is amplified and sequenced. The sequencing reads are then demultiplexed, and a seed sequence is identified for each sample. This seed is then used to initialize a probability matrix representing the sequence specificity model for the given TF. The model parameters are then optimized using a Hidden Markov Model-based motif discovery pipeline. Figure and legend taken and adapted from \citep{isakova_smile-seq_2017}.\relax}{figure.caption.29}{}}
\citation{isakova_smile-seq_2017}
\citation{isakova_smile-seq_2017}
\citation{weirauch_evaluation_2013}
\@writefile{lof}{\contentsline{figure}{\numberline{5.2}{\ignorespaces\textbf{Example of a Hidden Markov model :} initial HMM representation with a seed sequence 'ATGCC'. The upper Markov chain models + strand motif containing sequences, the middle one - strand motif containing sequences and the lower zero motif occurrence sequences. The FB, FE, RB and RE positions represents positions in the sequence that occur before and after the binding site on the forward and reverse strand. For these nodes, a self transition exist to allow the binding site to occur at a variable distance from the beginning and the end of the sequence. Once transiting toward the 1st position of the binding site, the next transition is forced toward the 2nd position in the binding site, and so on until the end of the binding site. The + strand and - strand Markov chains emission parameters are paired together (they have the same values), as represented by the grey dashed lines. The transition probabilities in red are not subjected to the Baum-Welch training. Finally, a binding model represented as a probability matrix is composed of the emission probabilities at the binding site positions. Figure and legend taken and adapted from \citep{isakova_smile-seq_2017}\relax}}{69}{figure.caption.30}}
\newlabel{smile_seq_hmm}{{5.2}{69}{\textbf{Example of a Hidden Markov model :} initial HMM representation with a seed sequence 'ATGCC'. The upper Markov chain models + strand motif containing sequences, the middle one - strand motif containing sequences and the lower zero motif occurrence sequences. The FB, FE, RB and RE positions represents positions in the sequence that occur before and after the binding site on the forward and reverse strand. For these nodes, a self transition exist to allow the binding site to occur at a variable distance from the beginning and the end of the sequence. Once transiting toward the 1st position of the binding site, the next transition is forced toward the 2nd position in the binding site, and so on until the end of the binding site. The + strand and - strand Markov chains emission parameters are paired together (they have the same values), as represented by the grey dashed lines. The transition probabilities in red are not subjected to the Baum-Welch training. Finally, a binding model represented as a probability matrix is composed of the emission probabilities at the binding site positions. Figure and legend taken and adapted from \citep{isakova_smile-seq_2017}\relax}{figure.caption.30}{}}
\@writefile{toc}{\contentsline{section}{\numberline{5.2}Hidden Markov Model Motif discovery}{69}{section.5.2}}
\newlabel{section_smileseq_hmm}{{5.2}{69}{Hidden Markov Model Motif discovery}{section.5.2}{}}
\@writefile{lof}{\contentsline{figure}{\numberline{5.3}{\ignorespaces\textbf{Predictive power of SMiLE-seq :}\textbf{A} the motifs compared to that of previously reported motifs that are retrievable from the indicated databases. For each motif, the AUC-ROC values on the 500 top peaks of the ENCODE ChIP-seq data sets for the corresponding TF was computed. The heatmap represents the AUC values computed for each method on the respective ChIP-seq data sets that were selected based on the highest mean AUC values among all five models. \textbf{B} the predictive performances of MAX and YY1 binding models were assessed using subsets of binding sites of decreasing affinities. Inside each peak list, the peaks were ranked by score and subsets of 500 peaks were selected. Peaks 1-500 have the highest affinity, then peaks 501-1000, and so on. The boxplots indicate the distribution of AUC-ROC obtained over all available peak-lists.\relax}}{72}{figure.caption.31}}
\newlabel{smileseq_auc}{{5.3}{72}{\textbf{Predictive power of SMiLE-seq :}\textbf{A} the motifs compared to that of previously reported motifs that are retrievable from the indicated databases. For each motif, the AUC-ROC values on the 500 top peaks of the ENCODE ChIP-seq data sets for the corresponding TF was computed. The heatmap represents the AUC values computed for each method on the respective ChIP-seq data sets that were selected based on the highest mean AUC values among all five models. \textbf{B} the predictive performances of MAX and YY1 binding models were assessed using subsets of binding sites of decreasing affinities. Inside each peak list, the peaks were ranked by score and subsets of 500 peaks were selected. Peaks 1-500 have the highest affinity, then peaks 501-1000, and so on. The boxplots indicate the distribution of AUC-ROC obtained over all available peak-lists.\relax}{figure.caption.31}{}}