geneXplain platform

Machine Learning and AI methods

TRANSFAC integrates advanced tools for the prediction of transcription factor binding sites (TFBS) and their combinatorial patterns across the genome using machine learning and AI-based approaches.

These capabilities include:

• AI-driven combinatorial TFBS analysis using the Composite Module Analyst algorithm
• Combinatorial modeling based on sparse logistic regression (MEALR)

AI-driven combinatorial TFBS analysis using Composite Module Analyst algorithm

Introduction to Composite Modules

Composite modules represent combinations of multiple TFBS that co-occur within regulatory DNA sequences. Identifying these modules enables the discovery of regulatory patterns that are significantly overrepresented in a target dataset compared to a background set.

Within the geneXplain platform, composite modules are identified using an in-house implementation of a genetic algorithm known as the Composite Module Analyst

Composite Module Analysis Workflow

The Composite Module Analyst operates on the output generated from TFBS site search analyses. Two complementary analysis workflows are available:

1. Construct Composite Modules (Gene Set-Based)

This method analyzes promoter sequences defined relative to transcription start sites (TSS) for a given gene set.

• Input: Results from Site search on gene set analysis
• Application: Identification of TFBS combinations in promoter regions
• Output: Composite modules that distinguish a target gene set (Yes-set) from a background set (No-set)

2. Construct Composite Modules on Tracks (Genome Coordinate-Based)

This method operates on DNA sequences defined by absolute genomic coordinates and is particularly suited for high-throughput sequencing data.

• Input: Results from Site search on track analysis
• Typical use case: Analysis of ChIP-seq fragments
• Output: Composite modules distinguishing a target track (Yes-track) from a background track (No-track)

Key Differences Between the Methods

The two approaches differ primarily in:

• Type of input sequences
  • Gene-based promoter regions vs. genomic coordinate-defined tracks
• Input data format
  • Gene sets vs. genomic tracks

Access in geneXplain Platform : 

Both analysis workflows are available in:

Analyses → Methods → Site analysis

Summary of Composite Module Construction

• The gene set-based method identifies regulatory TFBS combinations in promoter regions, distinguishing between Yes-set and No-set gene groups.
• The track-based method identifies TFBS combinations in genomic regions such as ChIP-seq data, distinguishing between Yes-track and No-track datasets.

Both analyses require:

• Precomputed site search results
• Defined Yes/No datasets
• A selected matrix profile

Additional Information

Hierarchical Structure of Composite Modules

Before interpreting the results, it is important to understand how composite modules are structured and visualized.

Composite modules are organized in a two-level hierarchy:

1. Site models
2. Modules

At the highest level, multiple modules together form a promoter model, representing the regulatory architecture of a gene.

Site Models (Level 1)

Site models represent individual transcription factor binding models derived from PWMs used in the selected profile. Each site model includes:

• A threshold value optimized by the genetic algorithm
• An N value representing the number of best TFBS matches used for scoring

Modules (Level 2)

Modules group multiple site models into combinatorial regulatory units. Each module:

• Contains several site models
• Is defined by a module width (DNA region where TFBS cluster)

Promoter Model

Multiple modules form a promoter model describing gene regulation. Users can define:

• Number of modules
• Number of site models per module
• Number of TFBS considered

Results Visualization

Below we provide the results visualisation of the Construct composite modules analysis obtained for the demo input data set. The input parameters of the method that were used for the analysis launch were as follows:

As a result, the method constructed two tables (Model visualization on Yes set and Model visualization on No set), two tracks (Yes track and No track), and one histogram. 

In the Model visualization on Yes set table the the primary results of the analysis are presented: the identified composite modules are shown in the promoters of the Yes set:

Each row in this table corresponds to one gene of the Yes set, and for each gene the Ensembl ID and the gene symbol are shown in the two first columns. The column Model displays a symbolic map of the gene promoter taken for the analysis, in this case -500/+100 relative to the TSS. 

Arrows of different colors correspond to individual TFBSs, and a gradient in grey corresponds to the statistical density of the identified composite modules. The most intensive grey color corresponds to the center of a composite module. Each individual TFBS on this map is clickable, and upon a click information is displayed in the Info box (bottom left corner in the tool). 

As an example, one blue arrow is selected on the promoter of the top gene in the screenshot above, and for this selected TFBS the following details are shown in Info box:

The last column in the table, Score, shows a score calculated for each promoter depending on the number of modules, site models, sites, their scores and other statistical parameters. The higher the score for a promoter, the better the differentiation of this promoter from the promoters of the No set. The column Score is used for default sorting of the table, with the highest scores on top.

In addition to that, at the bottom part of the tool in the Model visualization on Yes set table you can also see the schematic representation of the hierarchical structure of the identified composite module, as well as a comprehensive set of its statistical characteristics:

The Yes track provides essential information about the regulation of individual promoters and is therefore important to be included in the visualization of individual promoters by the genome browser.

The schematic visualization can be comfortably extended to a more detailed visualization for each individual promoter:

For a selected promoter, you can see a more detailed map, including the names of the matrices and the numbers of individual modules, M1 through M4. Each element of this interactive map has a corresponding check box. Unchecked elements will not be displayed on the map. De-selection is applied simultaneously to both: the detailed view of one promoter, and the table with the schematic representation of all promoters.

The table Model visualization on No set shows a visualization of the identified composite modules in the promoters of the No set.
The structure of this table is the same as that of the Model visualization on Yes set table, described above.

The function of the No track is to provide a possibility for a detailed visualization of no promoters in a way similar to that of the Yes track.

The distribution of scores for individual promoters is shown as a Histogram, where the promoter score value is shown on X axis and the percentage of promoters (% sequences) having this score is shown on the Y axis:

This histogram can be further interpreted applying the statistical characteristics described above.

The center, a vertical grey line, corresponds to the average score value and is equal to 3.44 in this example. Promoters from the No set with a score above 3.44 are shown in the histogram as blue bars to the right of the center, and they are referred to as false positives. In this example, the false positive rate is 16.82 %.

Promoters from the Yes set with a score below 3.44 are shown in the histogram as red bars to the left of the center, and they are referred to as false negatives. In this example, the false negative rate is 23.42 %.

A visual analysis of the histogram suggests that the Yes promoters with a score above 4.5 are very well separated from the No promoters, which means that for this part of the promoters the composite model constructed is most suitable. In this example there are 38 promoters with the score value >4.5; they can be saved as a separate gene set, and for them the model obtained works best.

Score calculation of the composite models

The figure below demonstrates the calculation of the score value for the composite modules in the promoter sequences. The TSS is shown as a thin arrow on the right side of the figure. Four thick arrows exemplify four sites found in this promoter. The color of the arrows exemplifies the site model which these sites belong to (three site models – red, green and blue).

A promoter model consists of K modules. The score of each module Mk(Score(Mk),k= 1, …,K) is calculated according to this formula:

Here, Site Score (t,i) is the site score for the sites found in the promoter, which is calculated by the Match algorithm.

mt– the number of sites of the site model t found in the promoter.

Tk– the number of site models in the module Mk, and

The final promoter score is calculated as the sum of the module scores Mk.

Standard deviation (σ) of the normal distribution is subject of optimization by the genetic algorithm and represents the width of the module in the output of the composite module analysis.

Further details on the composite modules construction can be found in the respective chapter of the geneXplain platform user guide.

Combinatorial modeling based on sparse logistic regression (MEALR)

Combinatorial regulation analysis of genomic or custom sequences

This workflow analyzes DNA sequences to identify combinations of transcription factor (TF) binding regions using MEALR models from TRANSFAC®.

Best suited for:

• Single sequence analysis
• Identifying regulatory patterns in genomic regions

Workflow steps:

The workflow proceeds through the following main steps:

1. Prediction of potential binding locations using the TRANSFAC® MEALR combinatorial regulation analysis
2. Extraction of TRANSFAC® PWMs represented by MEALR model hits using the Extract TRANSFAC® PWMs from combinatorial regulation analysis
3. Preparation of a cutoff profile with extracted PWMs for subsequent MATCHTM search using the Create profile from site model table
4. Prediction of binding sites represented by PWMs in input sequences using the TRANSFAC® MATCHTM for tracks
5. The tools Filter track by condition and Intersect tracks are applied to derive filtered model predictions and intersections of predicted TF binding site and combinatorial model locations.

Challenges:

Predicting TF binding sites is challenging because:

• Binding sequences are very short (10–20 base pairs)
• They vary significantly in sequence patterns
• Binding depends on biological context (cell type, chromatin state, TF cooperation)

Key advantage:
With almost 1000 human TFs, over 300 cell types and more than 50 tissue types, the TRANSFAC® library of MEALR models provides the first comprehensive collection of TF binding models that account for combinatorial TF-DNA complexes comprising multiple DNA-binding specificities as well as cellular and tissue-related contexts.

These models can therefore deliver predictions with increased accuracy, so that subsequent searches for binding sites that mark the locations of individual TFs are able to focus on relevant PWMs. 

This is an important step to select from the large number of PWMs curated by TRANSFAC® and to prioritize binding sites of interest. Within longer sequences the MEALR model predictions furthermore suggest subregions where binding by certain TFs most likely occurs.

A much more detailed description of this workflow can be found here.

MEALR combinatorial regulation analysis

This analysis applies combinatorial regulatory models (CRMs) based on the MEALR affinity score [1] to classify or scan sequences for occurrences of combinations of transcription factor binding sites represented by TRANSFAC® PWMs. The models are taken from the MEALR library whose training data originate from the TRANSFAC® collection of high-throughput sequencing experiments.

The method can be launched either in a classification or in the scan mode. The Classification mode evaluates input sequences as a whole, whereas the scan mode analyzes sequence windows separated by the given step size (sliding window). In scan mode, the Best hit method reports the best scoring sequence window disregarding a cutoff and the Cutoff method reports the best non-overlapping windows satisfying the specified cutoff.

The output folder encompasses a table and sequence track with information about model hits. The output table contains sequence start and end points of hits, model ids, match probabilities as well as other values as described below. For input sequences derived from genomic regions (instead of imported as custom sequences) the table includes in addition a sequence id generated for a region as well as the genomic sequence id, start and end coordinates.

[1] Katie Lloyd, Stamatia Papoutsopoulou, Emily Smith, Philip Stegmaier, Francois Bergey, et al., The SysmedIBD Consortium; Using systems medicine to identify a therapeutic agent with potential for repurposing in inflammatory bowel disease. Dis Model Mech 1 November 2020; 13 (11): dmm044040.

A much more detailed description of this method can be found here.

Extract TRANSFAC® PWMs from combinatorial regulation analysis

This tool extracts TRANSFAC® PWMs from a result table generated by the MEALR combinatorial regulation analysis. The PWMs represent transcription factor binding specificities that constitute the combinatorial module predicted by the MEALR model.

The output contains the TRANSFAC® PWMs extracted from MEALR models according to specified cutoffs. This table can further be applied in several analyses, e.g. to extract corresponding transcription factors using the tool or to create a profile for binding site predictions ( Create profile from site model table) with MATCHTM.

A much more detailed description of this method can be found here.

Search for discriminative sites with TRANSFAC® (MEALR)

The tool MEALR finds combinations of TFBS matrices that discriminate between two sets of sequences (denoted asYes and No sets). TheYes set may consist of genomic regions identified in a ChIP-seq experiment. No sequences are often other non-coding genomic regions not overlapping with the peaks.

MEALR differs from other tools in the following points:

• No cutoff or threshold is used on matrix scores to determine potential binding sites. Instead, MEALR calculates threshold-free sequence scores.
• MEALR builds a discriminative model for classification which is well-established and widely applied in statistical analysis called Sparse Logistic Regression. The model consists of a linear model that estimates the probability that a sequence belongs to the Yes set based on its binding site features.
• The sparseness constraint enables MEALR to select a subset of matrices relevant for classification of Yes and No sequences from a possibly large matrix library. Therefore MEALR’s output differs from other tools by presenting a focused set of matrices.
• While other site enrichment tools provided in the platform evaluate enrichment separately for each matrix, the model used in MEALR assesses the importance of matrices for discrimination in combination with other matrices of the library. Therefore, MEALR suggests (linear) combinations of transcription factor motifs.

For each input sequence x MEALR calculates a score for each PWM by the following equation where W denotes the number
of windows scored by the PWM and LSmax(xw) is the high log-odds score of window w :

f(x) = log( Σw exp(LSmax(xw)) / W )

Each sequence is therefore associated with a vector of scores, one from each matrix, and a class (Yes, No).

Let us present an example analysis for a ChIP-seq data set consisting of 500 peak regions and 1000 sequences randomly sampled from regulatory regions across the human genome.

Yes set  is the set of sequence intervals that you want to analyze, for example these can be ChIP-seq peak regions.

No set is the set of background intervals (control set).

In the input profile (collection of PWMs – binding motifs) the cutoffs will be ignored, because MEALR calculates whole sequence scores.

In the results of analysis the summary table would look like this:

A row of this table contains a matrix identifier and its logistic regression coefficient. The larger the coefficient value, the more important the corresponding matrix was for discriminating between Yes and No sequences. In our example, three of the five top matrices represent members of the transcription factor subfamily C/EBP.

For further details on this method please refer to the respective chapter of the geneXplain platform user guide.