T13 - Machine learning-based single-molecule analysis of epigenetic enzyme dynamics in live cells

Introduction: : The epigenetic state of the genome plays a pivotal role in regulating gene expression [1]. However, epigenetic regulation is inherently dynamic due to the intricate balance of multiple epigenetic enzymes, including histone deacetylase 1 (HDAC) [1]. Even within a single cell nucleus, individual epigenetic enzymes exhibit heterogeneous behaviors, with activity being dependent on local cues [2]. Therefore, to accurately measure these dynamics, live-cell imaging at single-molecule resolution is essential. Thus, we employ single particle tracking (SPT) of HDACs within live HeLa cells to precisely quantify the activity of individual enzymes at the molecular level. By tagging HDAC with HaloTag labeling, our approach allows for the analysis of molecular trajectory data by delineating tracks into 5 distinct classes of diffusive behavior: Confined, Anomalous, Brownian, Directed, and Butterfly (Fig. 1A) [3, 4]. HDACs within slow-diffusive regimes (i.e., Confined and Anomalous) exhibit stronger enzyme-DNA interactions, indicating high enzymatic activity. On the other hand, fast-diffusive regimes (i.e., Brownian and Directed) indicate inactive enzymes or target-searching states. Butterfly motion represents a hybrid state that alternates between confined and directed diffusion, reflecting complex chromatin interactions. To probe how epigenetic landscape changes affect HDAC behavior, we applied this classification scheme within the context of perturbed epigenetic states (i.e., epigenetic inhibitors). Furthermore, we trained a random forest machine learning algorithm on SPT-derived features to classify cells based on inhibitor treatment duration. This study indicates that dynamic, biologically relevant changes in enzyme activity can be effectively captured and analyzed using live-cell SPT combined with machine learning.

Materials and

Methods: : HeLa cells were transduced with a lentiviral vector encoding a fusion construct of HDAC and HaloTag proteins to establish stable protein expression for HDAC SPT. Cells were cultured on chambered coverglass slides and labeled with JF-549 HaloTag ligand (Promega) to tag HDAC proteins. To perturb epigenetic activity, cells were treated with Trichostatin A (TSA, histone deacetylase inhibitor, 200 ng/mL for 1 or 6 hours) prior to live cell imaging. Live-cell SPT imaging was performed on the Nanoimager (ONi) in phenol red-free culture media. Spot detection and tracking were performed using the ImageJ plugin TrackMate [5]. Trajectory motion types were annotated using mean squared deviation (MSD) analysis in MATLAB. Any enzymes unable to be classified based on MSD (i.e., Undefined molecules) were clustered with existing classes and annotated using a random forest approach, allowing all filtered molecules to have a motion type assigned. Machine learning (ML) classification was implemented in Python [4]. The ML algorithm was trained to distinguish TSA-treated cells from controls based on single-molecule trajectory features, with performance evaluated using the area under the receiving operator characteristic (ROC) curve (AUROC). Classification results used 5-fold cross validation averaged over 100 repetitions. Statistics were performed in GraphPad using one-way analysis of variance with Tukey’s post-hoc testing, and significance was defined as p-value < 0.05.

Results, Conclusions, and Discussions:: A stable HeLa cell line expressing HDAC-HaloTag fusion proteins was successfully generated and validated. SPT of HDAC enzymes produced trajectory maps of all particles tracked within each cell (Fig. 1A). MSD analysis enabled classification of trajectories into five motion types: Directed, Brownian, Anomalous, Confined, and Butterfly (Fig. 1A, B). Initially, a large portion of tracks remained unclassified (Undefined) (Fig. 1A). However, applying ML classification of Undefined tracks allowed for the categorization of all molecules into the 5 motion types with high precision (Fig. 1C). Each motion type reflected biologically meaningful differences in molecular behavior. For example, slow-diffusing tracks (Confined and Anomalous) feature lower confinement ratio than fast-diffusing tracks (Brownian and Directed) (Fig. 2A). Upon TSA treatment, the distribution of HDAC motion types shifted significantly, with a higher proportion of trajectories falling into fast-diffusing categories (Fig. 2B), indicating reduced chromatin engagement. Using a random forest model, we found that the inclusion of both motion type annotation and motion type estimation of Undefined tracks produced the highest AUROC values compared to analysis lacking these features (Fig. 3A, B). These findings indicate that TSA-induced epigenetic perturbations systematically alter HDAC diffusion dynamics, and motion type annotation, especially when integrated with machine learning classification, offers a sensitive and interpretable metric for assessing enzyme activity at the single-molecule level.
This single-molecule imaging technique provides novel insights into the dynamic behavior of epigenetic remodeling enzymes within living cells. SPT enables the detection of epigenetic enzyme activity as a function of local chromatin accessibility, with variations in enzyme diffusive behavior reflecting both enzymatic engagement and target-search mechanisms. By classifying all tracks into distinct motion types using machine learning, we can enhance predictive power of SPT for inferring chromatin conformation and epigenetic state. Consistent with its known role in promoting global chromatin decondensation, TSA treatment led to marked increases in fast-diffusing HDAC trajectories. Taken together, the integration of SPT-based super-resolution imaging with machine learning classification advances our understanding of epigenetic regulation and dynamic chromatin rearrangements, providing a more comprehensive and interpretable insight into areas of cellular reprogramming and gene expression.

Acknowledgements and/or References (Optional):: Acknowledgements: This work is supported by the NIH (P50 AR080581).

References: [1] Magaril+ (2023), Cell Dev Biol; [2] Carter+ (2024) Nat Rev Genet; [3] Simon+ (2024) Nat Rev Methods; [4] Lerner (2020) Cell STAR Prot; [5] Tinevez+ (2017), Sci Direct Methods.