A lineage tree-based hidden Markov model quantifies cellular heterogeneity and plasticity

Abstract

Individual cells can assume a variety of molecular and phenotypic states and recent studies indicate that cells can rapidly adapt in response to therapeutic stress. Such phenotypic plasticity may confer resistance, but also presents opportunities to identify molecular programs that could be targeted for therapeutic benefit. Approaches to quantify tumor-drug responses typically focus on snapshot, population-level measurements. While informative, these methods lack lineage and temporal information, which are particularly critical for understanding dynamic processes such as cell state switching. As new technologies have become available to measure lineage relationships, modeling approaches will be needed to identify the forms of cell-to-cell heterogeneity present in these data. Here we apply a lineage tree-based adaptation of a hidden Markov model that employs single cell lineages as input to learn the characteristic patterns of phenotypic heterogeneity and state transitions. In benchmarking studies, we demonstrated that the model successfully classifies cells within experimentally-tractable dataset sizes. As an application, we analyzed experimental measurements in cancer and non-cancer cell populations under various treatments. We find evidence of multiple phenotypically distinct states, with considerable heterogeneity and unique drug responses. In total, this framework allows for the flexible modeling of single cell heterogeneity across lineages to quantify, understand, and control cell state switching.

TNBC response to paclitaxel phenocopies interferon response which reveals cell cycle-associated resistance mechanisms

“…As chemotherapeutic drugs often have peak efficacy during a specific cell cycle phase, we next asked whether the combination of paclitaxel treatment and siRNA knockdown altered the dynamics of cell cycle progression. To that end, we trained a Markov Model on the live-cell data, which enabled us to infer transition rates and the average time spent in each of the four phases for a given treatment condition[ 42 , 46 ]. This approach uses the change in fraction of cells in each cell cycle phase over time ( Figure 5B ) to learn cell cycle-specific transition rates, which represent the fraction of cells that transition from one phase to another phase within a 1-hour timestep ( Supplemental Figure 6B ).…”

“…In this study we identified dual roles of the transcription factor ELF3 that contribute to paclitaxel tolerance by: 1) permitting cells to transition from G1 to S/G2, and 2) enabling successful division into two mononuclear daughter cells. These findings were enabled by a Markov Model of cell cycle progression built on population level cell count data which learned the transition rates between cell cycle phases and inferred cell cycle phase durations[ 42 , 46 ]. While the inferred cell cycle durations represent an accurate prediction of the population’s average behavior, they cannot inform whether this arises from a homogenous or heterogenous distribution of cell cycle durations.…”

“…An alternative approach could track individual cells and their progeny to build complete lineages with accompanying cell cycle timing information. Lineage based approaches tend to be relatively low throughput due to the computational and experimental requirements, but offer the opportunity to discern between heterogenous states of differing cycling speeds[ 46 ]. Another limitation of our Markov Model’s implementation is the assumption that transition rates are static throughout the duration of observations.…”

Quantifying Dynamic Phenotypic Heterogeneity in Resistant Escherichia coli under Translation‐Inhibiting Antibiotics

“…To gain further insights into the dynamic behaviors of the phenotypes across diverse Cm concentrations, we applied a modified Hidden Markov Model (HMM), often used for temporal pattern recognition, [ 31 ] to fit the real‐time single‐cell data trajectories of gene expression fluorescence intensity, [ 19 , 32 , 33 ] morphology, [ 34 ] and growth rate. [ 35 ] Based on the HMM fitted results, we examined observable behaviors such as residence time and switching time, as well as the non‐equilibrium dynamics and thermodynamics involved. [ 36 ] Specifically, our analysis of non‐equilibrium dynamics and thermodynamics included the potential landscape topography, state switching paths and barrier heights, flux loop values, time irreversibility, and entropy production rate.…”

The Application of Discrete Hidden Markov Model on Crosses of Diploid Plant

“…The hidden Markov model can help model pairs of events and their causes that cannot be observed directly [2]. This models have also been used in bioinformatics and biochemistry, some of them are [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13].…”