Cobb, K. | Search bioRxiv

bioRxiv2026

Inhibition of Dot1L Histone Methyltransferase Expands Bone Injury-Responsive CXCL12⁺ Stromal Progenitors

Adult bone marrow contains a heterogeneous network of skeletal stromal and progenitor cells (SSPC) that maintain bone homeostasis, support the hematopoietic niche, and drive the regenerative responses to injury. Despite their central role in bone regeneration, the mechanisms that regulate injury-induced SSPC activation and fate transitions remain incompletely understood. Here we identify Disruptor of telomeric silencing 1 like (DOT1L), the sole histone methyltransferase responsible for H3K79 methylation, as a critical regulator that restrains stromal activation. Using complementary genetic, pharmacologic, single-cell transcriptomic and injury models, we show that either Prrx1 lineage Dot1L haploinsufficiency or acute pharmacological Dot1L inhibition promote SSPC expansion, but through distinct programs: haploinsufficiency drives the expansion of Cxcl12+ CAR cells, whereas acute pharmacologic inhibition enriches fibroblastic-like stromal states. Single Cell Regulatory Network Inference and Clustering (SCENIC) analysis identifies DOT1L as a stabilizer of homeostatic marrow-supportive transcriptional programs, whose disruption facilitates transition to injury responsive stromal states. Partial loss of Dot1L in the Prrx1 lineage enhances injury-induced intramedullary mineralization in vivo. Collectively, these findings establish DOT1L as a gatekeeper of the stromal progenitor state that restrains lineage commitment and restrains the magnitude of the SSPC response following injury.

cell biology↗

bioRxiv2025

The reasonable effectiveness of domain adaptation for inference of introgression

Supervised machine learning approaches have proven powerful in population genetics. To use such approaches, training data with known inputs and outputs are required. Since such data are generally unavailable in population genetics, researchers typically rely on simulations under the models of interest to train machine learning algorithms. While powerful, this approach depends heavily on the models used to generate training data. Because of the variety and complexity of processes shaping genetic variation, it is inevitable that not all processes important in an empirical system will be included when generating training data. This leads to a mismatch between the data used to train a machine learning algorithm and the data to which the trained model is ultimately applied--i.e., a domain shift-- and can negatively impact inference. Here, we train a Convolutional Neural Network (CNN) to detect introgression between sister populations and demonstrate that it has near perfect accuracy when applied to data generated under the models used for training. To evaluate the impacts of domain shifts on inference, we generated new data with introgression from a third, unsampled population into one of the two focal populations (i.e., ghost introgression), and accuracy was substantially reduced on these data. Finally, we used domain adaptation, which aims to train a network that performs well in the presence of a domain shift. Notably, this requires no knowledge of the target or empirical domain. Our domain adaptation network was able to accurately detect introgression, even in the presence of unmodelled ghost introgression. We also applied this approach to empirical data to detect introgression between ABC Island brown bears and other populations of brown bears. Previous work has suggested that introgression between ABC Island bears and polar bears can mislead tests of introgression between populations of brown bears. We found that using domain adaptation reduced support for introgression between geographically isolated populations of brown bears, suggesting that our approach reduces false inferences of introgression due to ghost introgression.

evolutionary biology↗