Learning CIS- and trans-regulatory drivers of chromatin accessibility at single-cell resolution

The National Institute of General Medical Sciences awarded the University of Massachusetts Medical School $460,625 on September 1, 2026, under the Biomedical Research and Research Training program (CFDA 93.859) to develop machine learning methods that model cis- and trans-regulatory drivers of chromatin accessibility from single-cell multiomic datasets.

Over the five-year period of performance through May 31, 2031, the recipient will create machine learning approaches that simultaneously account for histone and nucleotide modification when predicting chromatin accessibility and use transcription factor amino acid sequences to adapt predictions to coding mutations and aberrant splicing. The work integrates cell-state information, chromatin biology, and protein sequence into machine learning models currently blind to these aspects of cellular regulation, enabling their application in novel settings.

Work is performed in Worcester, Massachusetts.

Abstract

The binding of transcription factors (TFS) to the genome is governed by CIS-regulatory features in the genome, such as motif presence, affinity, spacing, orientation, and higher-order syntax, as well as the trans-regulatory state of the cell, including the concentration of these TFS and the presence of their CO-binders. Each binding event has the potential to influence chromatin accessibility, histone modification, nucleotide methylation, transcription initiation, splicing, degradation, and numerous other cellular processes. Aberrant binding can lead to disease and dysfunction. Hence, TF binding is an immensely important aspect of regulatory genomics. My lab aims to comprehensively understand the regulatory role of each nucleotide in the human genome through the lens of TF binding and how this role changes across every cell in our body. We do this by developing machine learning (ML) methods using public data sources. Far from being black-boxes, modern ML methods have been invaluable tools for understanding by predicting variant effects, attributing importance to motif instances in the genome, and designing novel biosequences. In the next five years, we will develop ML methods that (1) simultaneously model CIS- and trans-regulatory drivers of chromatin accessibility from single-cell multiomic data sets, (2) account for local histone and nucleotide modification when making their predictions, and (3) use amino acid sequence of the relevant TFS directly to adapt predictions to coding mutations and aberrant splicing. Together, these projects will dramatically improve ML methods by making them explicitly aware of aspects of cellular biology they are currently blind to: trans-regulatory state of cells, histone and nucleotide modification, and protein sequence. Being aware of these aspects will enable ML methods to work in completely novel settings by, for example, predicting the effect of gene knockouts on global chromatin accessibility for all cells in an experiment, characterizing the effect that methylation of each cytosine or pairs of cytosines in the genome has on TF binding, and predicting the influence of each coding mutations on protein binding genome-wide. Further, these aims are modular, meaning that they can be pursued independently and simultaneously and then used together when development is complete.