Comparison of bioinformatics workflows: nf-core/tfactivity, TF-Prioritizer, & diffTF

Bang Thanh Huynh

Mentor: Dr. Markus Hoffmann, Georgetown University.

Date/Time: August 25th, 2026 at 2:00 PM.

Abstract: It is important to study transcription factor (TF) activity because it can provide insight into the regulation of gene expression. Several computational methods can predict TF activity, but they differ in their approaches, data requirements, reproducibility, and ease of use. TF-Prioritizer combines chromatin accessibility and gene expression data, using HINT to identify candidate TF footprints and TEPIC to estimate binding affinities and link regulatory regions to genes. The nf-core/tfactivity pipeline follows a similar approach but uses STARE to analyze accessible regions directly and link them to genes. It is built with Nextflow, a workflow system designed to improve reproducibility and portability. In contrast, diffTF estimates TF activity directly from accessibility changes around predicted TF-binding sites. Because these pipelines use different methods, it was not known whether they would identify similar TFs when looking at the same biological comparison. The aim of this project was to compare the TF rankings and activity predictions produced by the three pipelines and evaluate their practical differences. The pipelines were compared using ENCODE datasets from K562 and MCF7 cells. Five K562 and five MCF7 DNase-seq samples and their corresponding peak files were used. RNA-seq count data from the same cell types were also included in TF-Prioritizer and nf-core/tfactivity, while diffTF was run in its basic mode without RNA-seq. All pipelines were run on Expanse, a high performance computing cluster at the San Diego Supercomputer Center. TF-Prioritizer was run in a Singularity container, while diffTF was run in a separate Conda environment. The three pipelines produced very different TF predictions. There was no overlap among their top 13 ranked TFs. When the complete final results were compared, TF-Prioritizer and nf-core/tfactivity still did not share any TFs. However, 18 TFs reported by nf-core/tfactivity were also significant in diffTF, compared with two TFs from TF-Prioritizer. This finding shows that nf-core/tfactivity agreed more closely with the separate accessibility based analysis performed by diffTF, although overlap alone does not prove that one method is more accurate. Regulators supported by the literature were also examined. The diffTF pipeline recovered the known K562 regulators GATA1, GATA2, and TAL1 and MCF7 regulators FOXA1 and GRHL2, while KLF1 was identified by both diffTF and nf-core/tfactivity. TF-Prioritizer did not report any of these examples. However, other known regulators, including RUNX1 for K562 and ESR1 for MCF7, were not included in the final reported results. We further investigated the practical use of the pipelines by comparing their runtime, storage requirements, outputs, setup, and ability to recover from interrupted runs. The nf-core/tfactivity pipeline had the shortest runtime, followed by TF-Prioritizer, while diffTF took the longest. The nf-core/tfactivity pipeline also used the least storage, followed by diffTF, while TF-Prioritizer used the most. TF-Prioritizer and nf-core/tfactivity produced interactive HTML reports, while the diffTF tables and PDF figures required more manual interpretation. TF-Prioritizer and diffTF required more compatibility fixes and troubleshooting, although both were able to reuse completed results after an interrupted run. Overall, nf-core/tfactivity was faster, more storage-efficient, and more portable than the other pipelines, but it did not reproduce TF-Prioritizer’s predictions. The limited agreement among the three methods shows that TF activity predictions strongly depend on the approach used. Because this benchmark only used DNase-seq data, future work should test ATAC-seq and ChIP-seq datasets to determine whether these findings apply to other chromatin accessibility and TF binding assays.