Transcription factors (TFs) are master regulators of normal cellular processes such as differentiation and development, and their dysregulation in disease have made them appealing therapeutic targets. TF binding can be inferred from chromatin accessibility assays as depletions in local signal called “footprints”, but single-cell footprinting is challenged by sparsity, enzyme bias, and cellular heterogeneity. Here, we present TF-HOUND (High-resolution Occupancy prediction using Unsupervised Network Detection), a deep-learning framework that infers statistically significant TF binding from single-cell accessibility assays. TF-HOUND corrects for complex enzymatic sequence bias and encodes a smooth latent space to counter dramatically sparse accessibility profiles, enabling robust prediction of TF binding from just a few hundred cells. Benchmarking TF-HOUND in K562 chronic myelogenous leukemia cell line demonstrates high precision, sensitivity, and superior detection of footprints. We then deployed TF-HOUND to temporally resolve site-specific TF binding and regulation in single-cell accessibility sequencing of healthy blood and immune cell differentiation (hematopoiesis). Remarkably, we find that cell lineages are defined by differential binding activity of even a single TF, as well as families of TFs that we evaluate simultaneously. Collectively, we use TF-HOUND to reveal the landscape of TF binding and dynamics in differentiating cell populations, highlighting the regulatory nature of TF activity and its implication in disease at unprecedented resolution.

Before the talk (~3:45pm), tea and coffee will be served outside 116 Weniger.

After the talk, there will be a reception with food and drink in 247 Weniger.