iiinova — where your digital self keeps evolving

Seppe De Winter

@seppedewinter.bsky.social

Post-doctoral researcher at aertslab VIB-AI KU Leuven. seppedewinter.net

Seppe De Winter @seppedewinter.bsky.social · 15/01/2026

Finally, using topic-modelling we model TFBS co-occurrences. We validate these white-box enhancer models by performing cross-species integration of our TF-MINDI results, with a zebrafish developmental S2F model, and show that the topic models can recover zebrafish floor plate-specific regions.

Top left: schematic representation of topic modeling. 1. A count matrix (heatmap) of pattern clusters instances across genomic regions is generated. 2. Topic modeling is performed. A schematic representation is shown illustrating that both topic-pattern probabilities and region-topic probabilities are simultaneously optimised in this procedure. 3. Once the model is generated now sequence can be scored by generating a new pattern-cluster by genomic region count matrix and scoring this using the topic model.

Top right: Heatmap showing pattern-topic probabilities in organoid and embryo. Topic corresponding to floor plate cells (FP), pre-migratory neural crest (pre-mig. NC), migratory neural crest (mig. NC) and neurons (Neu) are highlighted. Under this heatmap another heatmap is shown for both organoid and embryo. This heatmap quantifies the correlation between accessibility based region topic probabilites and pattern based region topic probabilities.

Bottom left: tSNE dimensionality reduction of TF-MINDI instances from organoid, embryo and zebrafish. Instances are colored by dataset of origin and plot shows that they integrate well. TF-families are indicated with dashed lines.

bottom right top: Precision-recall curves of deepNeuralTube models, zebrafish models and pattern-topic probability models showing their accuracy of predicting floor plate specific regions in zebrafish. The pattern-topic models perform on par with the deepNeuralTube models.

bottom right bottom: Scatter plot showing that the average floor plate prediction score of the DeepNeuralTube models (x-axis) is correlated to the average floor plate pattern-region-topic probabilities (y-axis).

100

Seppe De Winter @seppedewinter.bsky.social · 15/01/2026

Implanting either two high affinity coordinator instances or a single coordinator instance together with 5 additional TF-MINDI TFBS is sufficient to generate new facial mesenchyme enhancers.

Left: Contribution score plot of six synthetic enhancers. Coordinator and non-coordinator instances are highlighted. Right: Bar chart showing luciferase activity of those synthetic enhancers. Plot shows that enhancers with either two high affinity coordinator instances or a single high affinity coordinator instances together with five additional instances are equally active. Enhancers containing only a single high affinity coordinator instance (and no additional binding sites) are less active.

100

Seppe De Winter @seppedewinter.bsky.social · 15/01/2026

To test the sufficiency of the TF-MINDI extracted enhancer code rules we turn to synthetic enhancer design in facial mesenchyme cells. A homeobox-ebox dimer motif (Coordinator) has been shown to be instrumental for this cell type. TF-MINDI identified Coordinator instances at varying affinities.

111

Seppe De Winter @seppedewinter.bsky.social · 15/01/2026

We validate the TF-MINDI instances using ChIP-seq data in PBMC. Showing that TF-MINDI is more accurate compared to traditional motif enrichment analysis tools.

A large tSNE dimensionality reduction showing PBMC TF-MINDI seqlets colored based on TF-family. This is surrounded by four smaller tSNE dimensionality reductons colored based on TF-ChIP-seq Z-score. Showing specific enrichment of TFs in TF binding sites annotated to the family of that TF. Bottom right shows ROC curve, comparing TF-MINDi based prediction of ChIP-seq signal with motif enrichment based prediction (cisTarget). This shows that TF-MINDI is more accurate.

211

Seppe De Winter @seppedewinter.bsky.social · 15/01/2026

Counting the number of TFBS instances across cell type-specific peaks reveals cell type-specific regulators in the form of a code-table. Similar codes are obtained for the organoid and embryo and we link instances to TFs based on cell type-specific TF-expression.

Figure showing three panels. Top: code table for organoid (left) and embryo (right). On the y-axis different TF-families; on the x-axis different cell type-specific regions. And a dotplot, where the dotsize represents the average number of instances of a TF-family for each cell type-specific peak and the color represents the average contribution per TF-family and cell type-specific peak. Bottom left: heatmap showing jaccard index quanitfying genomic overlap of organoid and embryo derived seqlets. Bottom right: Plot showing pearson correlation between TF-expression across cell types (x-axis) and average number of instances per cell type-specific peak (y-axis). A square is drawn whenever the pearson correlation coefficient > 0.35, and the color of the square represents the pearson correlation coefficient.

100

Seppe De Winter @seppedewinter.bsky.social · 15/01/2026

To obtain high dimensional embeddings of S2F identified motifs, annotate TFBS across cell-type specific peaks and model TFBS co-occurrences we developed a new python package named TF-MINDI. Resulting in > 400k annotated TFBS instances across the genome (each dot in the tSNE below is one instance).

172

Seppe De Winter @seppedewinter.bsky.social · 15/01/2026

We tackle this challenge in human neural development. We generate two multiome atlases: on neural tube organoids and on a 4 p.c.w. human embryo. The organoids recapitulate human embryonic development strikingly well. DeepNeuralTube S2F models recover most known cell type-specific TF binding motifs.

Two UMAP dimensionality reductions. One for neural tube organoids and the other for human embryo. The UMAP is colored based on cell type-identity.

111