Interestingly, seq2PRINT captured de novo sequence motifs resembling composite motifs involving Runx, Ets, and Gata, many of which were supported by structural data from PDB or AlphaFold3 predictions, suggesting physical cooperations between TFs.

Finally, in collaboration with the Wagers Lab at HSCRB, we examined the CRE alterations during murine hematopoietic stem cell aging. We observed global gain of Gata/AP-1/Runx/Ets/NF-I binding, loss of Ctcf/Nrf1/Yy1 binding, as well as weakened nucleosome footprints.

If we rank pseudobulks along the same differentiation lineage by their pseudo-time, we can reconstruct a “movie” of how TFs and nucleosomes reorganize during differentiation. We observed stepwise establishment of hemoglobin CREs during erythroid differentiation.

We found that instead of having only two states, open vs closed, each CRE can be bound by several distinct TF combinations across cell states/types. Individual CREs thus occupy complex regulatory states undetectable by simply quantifying overall accessibility.

The really exciting part is the combination of seq2PRINT with single cell data. By pseudobulking cells and using low-rank adaptation to tune seq2PRINT to the differences among cell states, we tracked the changes in TF binding across cell types in human hematopoiesis.

Building upon foundational work on DNA sequence models by the @anshulkundaje.bsky.social & others, we trained a deep learning model that predicts footprints from DNA sequence. We examined sequences that drive footprint predictions and saw that the model relies on the organization of TF motifs.

Here is an example region showing multi-scale footprints of CTCF and flanking nucleosomes:

Here we present PRINT. By improving the statistical approach and varying the footprinting kernel size, we detect proteins and complexes across diverse sizes (TFs and nucleosomes).