TFXcan

TFXcan was developed to test transcription factor (TF) binding-GWAS trait associations using SNP-based predictors of TF binding.

These SNP-based predictors are developed using Enformer, a sequence-to-function deep learning model. These predictors are called Enpact predictors. Enpact weights are linear combinations of epigenomic features and are stored here.

Version:

TFXcan v4.0

Usage/Command:

Using Colab

This is a version that runs a minimal TFXcan. You can edit and input your own GWAS.

Access to a computing cluster

If you have access to a computing cluster, you could download the colab notebook, connect to your GPUs (if you have that resource), and run.

Otherwise if you want results faster, you can submit a snakemake job as below:

1. conda activate <<conda environment>>
2. snakemake -s snakefile.smk --configfile config/pipeline.asthma.children.yaml --profile profiles/simple/ --resources load=45

The --resources load=45 flag makes sure that the PredictDB part of the pipeline does not run more than 9 jobs at a time on midway3 i.e 9*5. Any number could have been used but I chose multiples of 5. If your cluster allows you to run more than 100 jobs at a time, you can up this number.

To use screen [preferred]:

1. screen
2. conda activate << conda environment >> (see software section)
3. snakemake -s snakefile.smk --configfile config/job_config.yaml --profile profiles/simple/ --resources load=45

Within the profiles/ directory is the slurm configuration. You many need to modify this to suit your computing cluster.

Software:

This pipeline depends on a number of software to do the following:

1. Predict with Enformer (this dependency is optional) (Enformer, GPUs, pytorch)
2. Train models of TF binding that is linear on SNPs (Nextflow, predictDB)
3. Test TF binding-GWAS trait association (PrediXcan, Summary-PrediXcan, MetaXcan)

Conda

We suggest the following to have a hitch-free environment:

1. Use conda to create an environment and install the software with the environment files in there.

All of these software are self-contained in this repository. You only need to install the conda environment.

Apptainer/singularity

If you are familiar with how to use a container, the pipeline also allows that. There is a definition file that can be used to build the environment.

Input:

In general, the pipeline expects:

1. A yaml config or parameters file. There is a template configuration file that can be edited. There is also a minimal yaml file.
2. The GWAS summary statistics to run, listed under gwas: in the yaml file (phenotype: file name, with the files in input.directory). See the template and the minimal yaml file. Older configs use a separate metadata CSV (metadata:), which still works.
3. A number of files needed in the yaml config file. These can be downloaded from here. This is a direct download link. You will need to decompress this archive.

The GWAS summary statistics file should have the following columns (others headers are allowed but will be ignored):

|chrom|pos|variant_id|ref|alt|pval|zscore|beta|se|
|---|---|---|---|---|---|---|---|---|
|1|134|1_134_A_G|A|G|0.0001|0.1|0.7|0.1|

- chrom: (character or string) 1,2,3, e.t.c (No chromosomes X, Y, or M e.t.c)

- pos: (numeric) 134 (bp coordinates)

- variant_id: chrom_pos_ref_alt

- pval: (numeric) GWAS pvalues

- zscore: GWAS zscores; you can pre-calculate this from the beta and standard errors (beta/se)

• The framework assumes that genomic coordinates are in hg38 coordinates

• Weights file: You will need this in the config yaml file (see enpact_weights). The weights file is a dataframe of TF binding predictors. You can find and use examples from here.

Output:

The output of the pipeline is the association results of the GWAS trait with the TF binding, and it can be found in the data/.../output folder. The output is a summary ***.TFXcan.csv file of the association results file with the following columns:

|tfbs|zscore|effect_size|pvalue|var_g|pred_perf_r2|pred_perf_pval|pred_perf_qval|n_snps_used|n_snps_in_cov|n_snps_in_model|

Notes:

Updates:

[X] Fri Oct 9 2026: Removed the need to supply a separate metadata txt file. GWAS file(s) can be pointed directly in the input yaml.

[X] To predict TF/tissue binding, the pipeline takes in a dataframe of weights.

feature	TF/tissue1	TF/tissue2	...	TF/tissueN
f1	0.1	0.2	...	0.1
f2	0.1	0.2	...	0.1
...	...	...	...	...
f5313	0.1	0.2	...	0.1

[X] SNPs are matched with the reference panel and uses the matched SNPs for the PredictDB training. This is to ensure that the SNPs used for the PredictDB training are the same as the SNPs used for the GWAS.

[X] All software necessary for TFXcan are shipped with the pipeline. You only need to install the conda environment.