FABIAN-variant

Documentation

Contents

About

FABIAN-variant predicts the effect of DNA variants on transcription factor (TF) binding in human (Homo sapiens) and mouse (Mus musculus) genomes. Given a non-coding or regulatory variant, FABIAN-variant evaluates whether the variant is likely to disrupt, create, or alter transcription factor binding sites (TFBS). The analysis uses 40,315 models for 1,530 human TFs and 35,990 models for 1,123 mouse TFs, plus 1,259 optional deep learning models. Results are reported as a score between −1 (predicted binding loss) and +1 (predicted binding gain).

To get started, visit https://fabianapp.org/variant26/ and enter the DNA variant into the SNV or indel field, select one or more transcription factors and click on Analyse. If you do not know the location but have a reference and a variant sequence, you can still enter them by clicking on Enter sequences directly.

Interfaces

Search interface

FABIAN-variant search interface

Results page

FABIAN-variant results page

Detailed results page

FABIAN-variant detailed results page

Click on an image to enlarge it.

Variant input

FABIAN-variant supports five input modes for variants. In each mode the supported formats can be displayed by clicking on the link 'Format info' below the input field.

Transcription factors

FABIAN-variant 2026 integrates 40,315 prediction models for 1,530 human transcription factors plus 1,420 heterodimer complexes (e.g., AHR::ARNT, BACH1::MAFK). For mouse, 35,990 models for 1,123 transcription factors and 1,355 heterodimer complexes are supported.

The models were pooled from nine publicly accessible PWM databases together with TFFMs from JASPAR 2024. Models are assigned to each organism by matching TF gene symbols against reference lists from HOCOMOCO and CIS-BP. Models matching both species are included in both sets. Per-source model counts for each organism are shown below.

Source Type Models for human TFs Models for mouse TFs
JASPAR 2026 PWM 3,357 2,903
HOCOMOCO v14 PWM 6,307 5,044
CIS-BP 3.00 PWM 6,473 5,580
SwissRegulon PWM 1,316 1,313
Factorbook PWM 15,061 13,682
SELEX collection PWM 4,635 4,414
UniPROBE PWM 1,403 1,403
HOMER PWM 333 318
hPDI PWM 140 117
JASPAR 2024 TFFM 1,290 1,216
Total 40,315 35,990

TF names from different databases are mapped to their canonical gene symbol (HGNC for human, MGI for mouse). For example, models listed under the protein name P53 in HOCOMOCO are displayed under the gene symbol TP53.

The TF selection list has three heterodimer modes, cycled by clicking the button below the list: Heterodimers excluded (default) hides all heterodimer entries; Heterodimers merged into TFs joins heterodimers into their component TFs, so that selecting ARNT automatically includes models for heterodimers containing ARNT, such as AHR::ARNT; Heterodimers shown separately displays heterodimers as individual entries that can be selected independently.

1,290 transcription factor flexible models (TFFMs) from JASPAR 2024 are included. For each transcription factor, FABIAN-variant combines the results of multiple models into a final prediction of the binding affinity change.

In addition, 1,259 deep learning models (BPNet) covering 240 human TFs from the JASPAR 2026 Deep Learning collection predict binding changes from the sequence context around the variant, available when genomic coordinates are provided. See Deep learning models (BPNet) for details.

The data sources are covered by different licenses, listed in the Acknowledgements.

Known transcription factor binding sites

On the results page, FABIAN-variant highlights known binding sites for transcription factors with a black rectangle around the score. The number of transcription factors that each source covers, per assembly, is shown below (a dash means the source does not cover that assembly):

Source hg19 hg38 mm10 mm39
ENCODE 491 922 39 -
ReMap 875 875 414 414
ChIP-Atlas 1,071 1,071 469 -
UniBind - 267 268 -
Ensembl Regulation 42 42 - 1
Distinct TFs (union) 1,161 1,205 515 414

On the search page, the databases used for the 'Known TFBSs' TF filter can be selected individually (ENCODE is selected by default). On the results page, hovering over a cell with a known TFBS shows which databases confirmed the binding site.

Please note that this function is only available if you entered genomic positions. Because the binding regions provided by ENCODE, ReMap and ChIP-Atlas are several hundred bases long, a highlighted position does not necessarily contain a binding site for the TF at the exact variant location. UniBind and Ensembl Regulation instead provide motif-level sites (about 15 bases), which are localised more precisely.

Evaluation of TFFMs and PWMs

TFFMs and PWMs are evaluated at every position overlapping at least one variant base, on both strands, in both the reference sequence and the variant sequence. The highest score in the variant sequence is compared with the highest score in the reference sequence. A higher Ref score than Alt score indicates weakened binding (a loss), and a higher Alt score indicates strengthened binding (a gain). For each model, FABIAN-variant generates a per-model score S between −1 (likely binding loss) and +1 (likely binding gain) as: S = tanh(F · ln 2),    F = salt − sref 1 − max(sref, salt) + α

with pseudocount α = 0.1 to avoid zero in the denominator. The figure below illustrates S for all combinations of reference and alternate scores in [0, 1]. Blue indicates a predicted binding gain, red a predicted loss. Hover over the plot to read exact values.

Per-model scores with |S| < 0.001 are labelled NA.

To obtain the combined score from multiple models, FABIAN-variant averages the per-model scores S. If both TFFMs and PWMs are available, by default only the results from TFFMs are used for the combined score (this setting can be changed by unchecking 'Options > Prefer TFFMs' on the results page).

The Ref score, Alt score, the per-model score S and the combined score are shown on the results page. For example:

Detailed results table showing the Ref score, Alt score, per-model score S, and combined score

In the detailed results table, a dot (•) next to each model indicates whether it contributes to the combined score. Models without a dot are excluded because 'Prefer TFFMs' is enabled (default) and TFFMs are available for this transcription factor.

The Ref and Alt position columns show the location of the highest-scoring motif match relative to the variant site, formatted as start,end with an arrow indicating the DNA strand (→ plus, ← minus).

The species column shows the organism(s) from which the model was derived.

Score magnitude reflects the predicted strength of the binding change. Large positive or negative combined scores more reliably indicate a gain or loss of binding, whereas scores close to 0.0 are less certain and should be interpreted with caution. In a benchmark against experimental SNP-SELEX data, restricted to variant–TF pairs with a statistically significant binding preference (P < 0.05), agreement between the predicted direction and the measured preference rose with the combined |score|, from about 60% near 0.0 to about 99% near 1.0. By default, the 'All TFs' search hides transcription factors whose combined |score| is below 0.05.

A C++ implementation of the forward-backward algorithm evaluates TFFMs. See this article to learn more about TFFMs:

Mathelier A, Wasserman WW. The next generation of transcription factor binding site prediction. PLoS computational biology. 2013 Sep 5;9(9):e1003214. https://doi.org/10.1371/journal.pcbi.1003214

There are two types of TFFMs: detailed models and first-order models. Detailed models are listed as JASPAR2024_DetailedTFFMs and first-order models as JASPAR2024_FirstOrderTFFMs in the database field in the results table. The model ID field starts with TFFM (e.g., TFFM0040.1). Note: JASPAR 2026 did not include TFFMs, so FABIAN-variant uses the JASPAR 2024 TFFMs.

Position count matrices (PCMs) were converted to position weight matrices (PWMs) using the method described in:

Bucher P. Weight matrix descriptions of four eukaryotic RNA polymerase II promoter elements derived from 502 unrelated promoter sequences. Journal of molecular biology. 1990 Apr 20;212(4):563-78. https://doi.org/10.1016/0022-2836(90)90223-9

A custom C++ implementation computes the scores.

The database field in the results table shows the model source:

Deep learning models (BPNet)

FABIAN-variant scores variants against 1,259 deep learning models from the JASPAR 2026 Deep Learning collection, trained on ENCODE ChIP-seq data for 240 human TFs. Unlike PWMs, which detect motif disruption within a short window, DL models predict binding changes from a 2114 bp sequence context around the variant. Results are cell-type-specific (each model was trained on data from a specific cell line).

DL scoring is available when genomic coordinates are provided. For mouse genomes, predictions are based on human models and may be less precise. For VCF files with more than 100 variants, DL scoring is limited to the first 100 variants. For each variant, the scorer reports:

The per-model prediction (loss, gain, NA) and its red/blue shading come from logFC alone. NA means |logFC| < 0.001. Otherwise, the sign gives the direction and the shade depth the magnitude. JSD has no direction: a variant can shift the binding profile (high JSD) while barely changing total binding (logFC near zero), or the reverse.

DL results are shown in the detail view for each TF and as indicators in the overview table. For example:

Deep learning models table showing the cell type, JSD, prediction, and logFC for each BPNet model

See the JASPAR 2026 publication for details on the BPNet models:

Ovek Baydar D, et al. JASPAR 2026: expansion of transcription factor binding profiles and integration of deep learning models. Nucleic Acids Research. 2025 Dec 2;gkaf1209. https://doi.org/10.1093/nar/gkaf1209

Results page

Each cell of the results table shows the combined score for a variant and transcription factor as a colour, with binding loss in red and gain in blue. Deeper shades represent a stronger loss or gain. Known TFBSs are displayed with a border around the cell.

Moving the mouse pointer over a coloured cell reveals the individual model scores. Clicking on the table cell shows the detailed results page.

Variants have the format chr1:159204893T>C.1 or AGTCCTTGGCTCTTAT>AGTCCTTGGCTCTTAC*.1. In both cases, .1 is the line number of the variant in the input. * indicates that some bases of a long sequence are not displayed. Clicking on a variant opens the corresponding location in the UCSC Genome Browser. Clicking on a transcription factor opens Ensembl.

The results table can be filtered and sorted in the browser using the checkboxes and radio buttons in the header of the page. Filtering and sorting only consider the currently visible TFs and variants. Filters combine per variant, so a transcription factor is kept only where the selected conditions are met at the same variant, for example a notable effect at a known binding site.

Results retention

Results are accessible via unique URLs and retained on the server for 14 days after the analysis is complete. After this time, they are automatically deleted.

Results can also be manually deleted before the retention period expires by clicking the 'delete' link in the computation log. Deleting results removes all information about the search parameters and uploaded variants from the server. Deleted results cannot be restored.

To extend the retention period, click the 'save' link in the computation log on the results page (visible under 'Options > Show log'). A verification email will be sent to the provided address. Once confirmed, the project is excluded from automatic deletion. Saved projects can still be deleted manually at any time using the link in the verification email.

Download format

The full download of all results has the following columns:

variant tf model_id database ref_score alt_score start_ref end_ref start_alt end_alt strand_ref strand_alt prediction score

The summary download mirrors the results table and reflects the filters and sorting currently applied on the results page. Scores for a known TFBSs are marked with *.

Programmatic access

FABIAN-variant provides a RESTful JSON API for programmatic access, and it is the recommended interface for new integrations. The API supports variant submission, job-status polling, retrieval of results, and job deletion, returns structured JSON responses, and can be tested directly in the browser. The API documentation includes ready-to-run examples for Python, Perl, and cURL.

The legacy form-based interface also remains available for compatibility with scripts written against earlier versions. On Unix-based systems, cURL can be used to submit variants and retrieve results. Analysis runs asynchronously, so the results file is polled until it becomes available. A single variant can be submitted and its results fetched with:

printf "($(date +%T)) Submitting " && \
FABIANID=$( curl -sLD - -o /dev/null \
-F "mode=single" \
-F "single_variant=chr1:159204893T>C" \
-F "genome=hg38" \
-F "tfs_filter=all" \
-F "sources=JASPAR-PWM JASPAR-TFFM HOCOMOCO" \
https://fabianapp.org/variant26/analyse \
| grep -im 1 "location: " | grep -o "\([0-9]\+_[0-9]\+\)" ) && \
i=1; until curl -sfo fabian.data_${FABIANID} \
https://fabianapp.org/variant26/temp/${FABIANID}/fabian.data; \
do printf "\r($(date +%T)) Waiting for $FABIANID"; \
[ $i == 30 ] && sleep $i || sleep $((i++)); done && \
printf "\r($(date +%T)) Saved file fabian.data_${FABIANID}\n"

Some parameters are specific depending on the mode and which transcription factors you are looking for. A few examples are listed below.

If the request is correct, cURL polls the server until results are available, which are then saved under a project-specific name (e.g., fabian.data_1650751034_19489). Please note that the polling loop can wait indefinitely in case of an error. The status can always be checked at the project-specific URL (e.g., https://fabianapp.org/variant26/1650751034_19489).

VCF submissions follow the same pattern, except the input is uploaded as a file. Submit with -F "mode=vcf" and -F "filename=@TinyExample38.vcf" in place of the single-variant fields. VCF result tables can be large, so they are also written as a compressed archive (fabian.data.zip), which is smaller to download. Poll for it the same way:

until curl -sfo fabian.data_${FABIANID}.zip \
https://fabianapp.org/variant26/temp/${FABIANID}/fabian.data.zip; do sleep 2; done

Please do not run more than three automated requests at the same time! If you require more processing slots, please send a short email with details of your request.

Standalone version

For users who require local offline execution or need to process large amounts of data, a standalone version of FABIAN-variant can be run on local infrastructure. The standalone tool is a single C++ executable with prediction models from openly-licensed sources embedded, requiring no additional dependencies.

See the standalone version documentation for technical details, command-line options, and download links.

History

FABIAN-variant is an update of the ePOSSUM software.

The initial version of FABIAN-variant was released on 20 December 2021.

FABIAN-variant 2026 is the second major version, featuring updated model databases, mouse genome support, and deep learning models. See the what's new and changelog pages for details.

Citation

If you use FABIAN-variant in your work, please cite:

Steinhaus R, Robinson PN, Seelow D. FABIAN-variant 2026: improved prediction of the effects of DNA variants on transcription factor binding. Nucleic Acids Research. 2026;gkag449. https://doi.org/10.1093/nar/gkag449

The original version is described in:

Steinhaus R, Robinson PN, Seelow D. FABIAN-variant: predicting the effects of DNA variants on transcription factor binding. Nucleic Acids Research. 2022;50:W322–9. https://doi.org/10.1093/nar/gkac393

Acknowledgements

FABIAN-variant relies on data from multiple public databases. Each source is governed by its own license terms.

Team

FABIAN-variant 2026 has been developed at the BIH (Berlin Institute of Health @ Charité – Universitätsmedizin Berlin) by

Contact

If you have suggestions about this software, please email robin.steinhaus (at) bih-charite.de. If you discover a bug, please submit a ticket via email using this link.

Imprint / privacy