Publications

We combined whole-genome sequencing with single-cell chromatin and gene-expression data from 190 donors to investigate how non-coding variants influence Parkinson’s disease risk. Across two brain regions and 4.2 million nuclei, we identified more than 53,000 variants associated with cell type-specific changes in enhancer accessibility. Deep-learning models predicted many of these effects directly from DNA sequence and showed that they often arise by altering transcription factor binding. By linking variants to their target genes, we could prioritize likely causal variants at known Parkinson’s risk loci, including in rare disease-relevant cell types such as dopaminergic neurons.

The mammalian cerebellum has experienced many evolutionary innovations, but their molecular basis remains elusive. Most phenotypic changes are thought to be driven by mutations in cis-regulatory elements (CREs) such as enhancers and promoters, which control gene expression in a cell type–specific manner. However, the fast CRE evolution and our limited understanding of how DNA sequences encode regulatory activity have hindered our ability to study regulatory innovations. Single-cell multiomics enable the mapping of CRE cell type specificity, whereas recent advances in machine learning facilitate predicting CRE accessibility from DNA sequence. We reasoned that if CRE sequence codes of cerebellar cell types are conserved across mammals, then we could use sequence-based deep learning models to reconstruct CRE evolutionary histories from genomic sequences and identify sequence changes underlying gene regulatory innovation.

S2F-deeplearning models can accurately encode enhancer, decoding these models into human-interpretable rules remains a major challenge. Here we develop a new python package and framework, TF-MINDI, that generates high dimensional embeddings of S2F identified motifs, annotate TFBS across cell-type specific peaks and model TFBS co-occurrences. We apply this framework to human neural development and extract its cel l type-specific cis-regulatory rules.