Robust footprinting with sample-specific Tn5 bias correction for bulk and single cell ATAC-seq

Footprint analysis of assay for transposase-accessible chromatin via sequencing (ATAC-seq) enables base resolution mapping of regulatory elements but is often underpowered and prone to false positives due to data sparsity and Tn5 transposase cleavage bias. We uncover substantial sample-to-sample variability in Tn5 transposase cleavage bias across samples, revealing a previously underappreciated source of batch effects and highlight the motivations for sample-specific Tn5 bias modeling. We present TraceBIND, a computational framework that corrects sample-specific Tn5 bias through mitochondria-based fine-tuning of PRINT's deep learning model, and use it to identify TF and nucleosome footprints through a dynamic flanking window statistical scan. Compared to existing approaches, TraceBIND substantially reduces false discoveries, controlling type 1 error while maintaining high sensitivity. Despite the fact that TraceBIND is completely unsupervised and does not require training on ChIP-seq data, we used multiple validation analyses to demonstrate that TraceBIND matches the power of supervised methods for detection of known TF binding sites. In scATAC-seq data from aging rat kidney, TraceBIND discovers age-associated dynamic regulatory changes, linking footprint activity to age-associated epigenetic drift. This demonstrates that footprint-informed scATACseq analysis reveals rich regulatory signals missed by conventional peak-based approaches.