iiinova — where your digital self keeps evolving

Pierre-Luc Germain @plger.bsky.social · 26/08/2026

Based on these tools we then develop a framework to systematically screen for synergistic and antagonistic interactions between TF motifs. Joint work with Jiayi Wang, @esonder2.bsky.social , @sdomcke.bsky.social and @markrobinsonca.bsky.social

030

Pierre-Luc Germain @plger.bsky.social · 26/08/2026

We introduce 3 R packages: weightedMotifAccess implements fragment weighing strategies, betterChromVAR an analytical version of chromVAR (eliminates stochasticity, orders of magnitude faster, etc.), and epiwraps to facilitate especially visualization

110

Pierre-Luc Germain @plger.bsky.social · 26/08/2026

Sharing a new preprint about technical variation in (bulk/ #SingleCell ) #ATACseq and motif accessibility analyses, and methods to correct for them www.biorxiv.org/content/10.6...

1117

Pierre-Luc Germain @plger.bsky.social · 23/06/2026

If by "non-specific binding" you mean binding that is not primarily driven by sequence affinity, or binding that is in common between many TFs, then yes, absolutely, the model learns that because that's what experimental data looks like, and that's the whole point about promiscuity.

110

Pierre-Luc Germain @plger.bsky.social · 23/06/2026

Important point: the model prioritizes sites for given TFs, meaning that it says where it would be more likely to be found assuming it's active. As mentioned on the homepage, it will also predict when the TF is absent for the given cell type. For us filtering active TFs is done downstream.

100

Pierre-Luc Germain @plger.bsky.social · 23/06/2026

Doing this at such a scale revealed an unexpectedly high promiscuity in TF occupancy. We show that crowdedness is highly cell-type specific, functional and most likely the product of tethered binding from distal elements brought in 3D contact.

110

Pierre-Luc Germain @plger.bsky.social · 23/06/2026

Going beyond the method, we generated a massive compendium of binding predictions for 1108 TFs across 43 cell types representing all major lineages. Interpretable features yield insights into the TFs, and the predictions can be used in downstream tasks such as reverse-engineering perturbed TFs.

130

Pierre-Luc Germain @plger.bsky.social · 23/06/2026

Cell-type specific DNA accessibility profiles are instead easy to get, so we developed a scalable platform to predict ChIP peaks from ATACseq data based on >12k paired datasets. TFBlearner achieves competitive accuracy in unseen cell types in a pre-registered benchmark.

Pre-registered benchmark on unseen cell-types

110

Pierre-Luc Germain @plger.bsky.social · 23/06/2026

How do >1600 transcription factors (TFs) bind across hundreds of cell types? Experimentally profiling every combination isn't practically doable. But relying on DNA motifs isn't a great alternative -- they lack both sensitivity and specificity. #Genomics #GeneRegulation 👇

2125