1. Poster:
Session B #114: 1pm-2pm
HopLand: Single-cell pseudotime recovery using continuous Hopfield network based modeling of Waddington’s epigenetic landscape
Jing Guo1,2 and Jie Zheng1,3,4
1Biomedical Informatics Lab, School of Computer Science and Engineering, Nanyang Technological University
2Bioinformatics Institute, Agency for Science, Technology, and Research (A*STAR)
3Genome Institute of Singapore, A*STAR
4Complexity Institute, Nanyang Technological University
ISMB/ECCB 2017

2. Pseudo-time recovery in single-cell data
What
Biological Time
Physical
Time
Input: DN*S
Tree=Mapping(D)
Output: Tree
multiple time points

3. Pseudo-time recovery in single-cell data
Why
Understand the transition of gene expression profiles
Increases the temporal resolution

4. Current methods Wanderlust (Bendall et al., 2014)
Wishbone (Setty et al., 2016)
Monocle (Trapnell et al., 2014)
Diffusion map (Haghverdi et al., 2015)
Dimensionality reduction
SCUBA (Marco et al., 2014)
Modeling system dynamics
Topslam (Zwiessele and Lawrence, 2016)
Waddington’s epigenetic landscape
Current methods…
1. Common… 2. specific one….
Summary:
Dimensionality reduction
Mapping cells to the latent space
drawbacks…..
1. common… 2. specific one….

5. The overview of HopLand
A landscape inferred from single cell sequencing data
Individual cells are placed on the surface regions corresponding to their developmental stages
The order of cells is determined by the geodesic distances in the landscape

6. Waddington’s epigenetic landscape
Figure from (Mohammad and Baylin, 2010)
Attractor system

7. Continuous Hopfield network (John Hopfield in 1984)
1) continuous variables and predicts continuous responses
2) Neurons == Genes
describe the regulatory function of a gene
What they do… W
The discrete Hopfield network (John Hopfield in 1982)
has been used to study biological systems with each neuron representing a gene (Taherian Fard et al., 2016, Maetschke and Ragan, 2014, Lang et al., 2014).
discrete variables and predicts discrete responses

8. 1. Data preprocessing Temporal single cell data
Physical
Time
How…….. edgeR, DESeq
Identify differentially expressed genes
genes are detected in only a subset of cells and such dropout events are thought to be rooted in the stochasticity of single-cell library preparation [8]
Dropout effect 
Over-dispersion
Data types: qPCR, scRNA-seq
MAST,SCDE…
(Scadden, 2014, Nature Methods)

9. 2. Parameter estimation CHN Initial parameters
Generate real trajectories
Optimize objective function
CHN W
There are several parameters in the ODE model of kinetics in Eq To
infer these parameters = fi; Ii;Ci;Wij ; i; j = 1; 2; :::;Ng from the sequencing
data, an optimization method (Algorithm 4) which ts the simulated and
observed single-cell data is proposed based on the premise that a realistic
model should be able to generate simulated data consistent with the real
data.

10. 3. Landscape construction
GP-LVM
Dimensionality reduction
Generate a grid
Get high-dimensional vectors
Calculate energies
Plot the landscape
Altitude
Comp2
Comp1
Interpolation

11. 4. Pseudotime recovery Calculate geodesic distances
Build the minimum spanning tree

12. Evaluation Synthetic dataset:
Applied to real single-cell gene expression data from different types
of biological experiments and compared against other methods, HopLand
outperformed most of the other methods in most cases. In addition, our
method could also be used to identify key regulators and interactions which
is helpful for the understanding of underlying mechanisms.

13. Dataset: mouse pre-implantation development The expression profiles of 438 cells with 48 genes per cell covering the developmental stage from the 1-cell to 64-cell stages (Guo et al., 2010).

14. (a) The Waddington’s epigenetic landscape recovered using HopLand
(a) The Waddington’s epigenetic landscape recovered using HopLand. (b) The minimum spanning tree constructed from Waddington’s epigenetic landscape. The dots are colored according to the developmental stages of the cells in the dataset of GUO2010.

15. Mapping gene expression values to the landscape using (a) FGF4, (b) GATA4, (c) CDX2 and (d) SOX2. The value decreases from dark red to white.

16. The core network found by searching GeneGo MetaCore for interactions overlapping with HopLand’s top interactions and genes.

17. Summary Poster: Session B #114: 1pm-2pm Limitations & Future work:
Continuous Hopfield network
Provide a non-linear mapping between the Waddington's epigenetic landscape and the phenotype space.
Waddington’s epigenetic landscape
Construct the landscape based on biological interactions between genes that allows to simulate real biological processes.
Geodesic distance
Use the geodesic distances in the landscape to refine the estimation of distances between cells.
Waddington's epigenetic landscape which merges the genotype-phenotype mapping with the last developments of genomics on transcriptional regulation, epigenetic modification and signalling pathways, is applied to analyze and visualize the system kinetic.
Limitations & Future work:
Supervised method ---- unsupervised cluster methods
Computationally costly ---- Stochastic gradient descent (SGD), HPC
Data noise ---- different types of sequencing technologies
Poster: Session B #114: 1pm-2pm

18. Acknowledgement
Jie Zheng
The SCE-BII joint PhD program