Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse...
Transcript of Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse...
![Page 1: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/1.jpg)
Robust Reverse Engineering of
Dynamic Gene Networks
Under Sample Size Heterogeneity
Ankur P. Parikh, Wei Wu, Eric P. Xing
Carnegie Mellon University
![Page 2: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/2.jpg)
Outline
Motivation
Challenge of sample size heterogeneity
Our solution
Results
Conclusion
2
![Page 3: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/3.jpg)
Outline
Motivation
Challenge of Sample Size Heterogeneity
Our solution
Results
Conclusion
3
![Page 4: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/4.jpg)
4
Multiple Network Estimation for Progression
and Reversion of Breast Cancer cells
T4 Revert Group 1
T4 Revert Group 2
T4 Revert Group 3
S1 (normal)
T4 (malignant)
![Page 5: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/5.jpg)
5
Stem Cell Differentiation
![Page 6: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/6.jpg)
Use Microarray Data to Infer
Network Structure
6
?
?
?
![Page 7: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/7.jpg)
A Data Integration Problem Small numbers of samples from many related cell states.
Poses opportunities and challenges.
Challenges
How to leverage similarities among cell states for more
accurate network estimation? (some of our previous work)
How to integrate data in a “correct” way that does not
compromise validity of biological conclusions? (a focus of
this work)
7
![Page 8: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/8.jpg)
Different cell states have different sample sizes
Challenge: Sample Size Heterogeneity for
network estimation
8
![Page 9: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/9.jpg)
Cell states with more samples have more edges
Biases Resulting Networks
9
![Page 10: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/10.jpg)
Sample Heterogeneity Makes
Network Analysis Problematic Unclear which network differences are meaningful
and which are an artifact of sample size heterogeneity
Comparing Macro-Topology Features
Network Density
Centrality
Comparing Gene Level Interactions
Gene neighborhoods
Local modules
10
![Page 11: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/11.jpg)
Our Contribution
Identify the novel problem of sample size heterogeneity in
network reconstruction
We focus on sparse regression methods for network
reconstruction [Meinshausen and Buhlmann 2006, Friedman et al. 2008.
Song et al. 2009, Parikh et al. 2011]
We add a regularization constraint to make these methods
robust to sample size heterogeneity
11
![Page 12: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/12.jpg)
Outline Motivation
Challenge of sample size heterogeneity
Our solution
Background
Increasing robustness to sample size heterogeneity
Sharing information across states
Results
Conclusion
12
![Page 13: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/13.jpg)
Background on Sparse Regression
Network Reconstruction[Meinshausen and Buhlmann 2006]
Network Learning with the Graphical LASSO [Meinshausen and
Buhlmann 2006]
13
Learn each neighborhood
separately
![Page 14: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/14.jpg)
14
Use the LASSO to select a sparse set of only the relevant
edges
Selects sparse set of edges
Background on Sparse Regression
Network Reconstruction[Meinshausen and Buhlmann 2006]
![Page 15: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/15.jpg)
Repeat this for every node
15
Background on Sparse Regression
Network Reconstruction[Meinshausen and Buhlmann 2006]
![Page 16: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/16.jpg)
16
Background on Sparse Regression
Network Reconstruction[Meinshausen and Buhlmann 2006]
Combine all the neighborhoods to form a network
![Page 17: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/17.jpg)
(Naïve way) run the graphical lasso for each!
More sophisticated techniques to share information exist (Song
et al. 2009, Ahmed and Xing 2009, Parikh et al. 2011)
Learning Multiple Networks
17
network 1
network 2
Graphical lasso for network 1
Graphical lasso for network 2
![Page 18: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/18.jpg)
The graphical lasso will encourage networks with more samples
to have more edges.
Simply tricks such as scaling parameters only work in theory
but not in practice.
Sample Size Heterogeneity
18
network 1
network 2
Graphical lasso for network 1
Graphical lasso for network 2
![Page 19: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/19.jpg)
Intuition Of Our Solution
Make “normalizing” assumption that absolute sum of
edge weights for all networks is equal
Incorporate this assumption into graphical lasso procedure
19
𝑠𝑢𝑚 𝑜𝑓 𝑎𝑏𝑠𝑜𝑙𝑢𝑡𝑒 𝑒𝑑𝑔𝑒𝑤𝑒𝑖𝑔ℎ𝑡𝑠 𝑜𝑓 𝑛𝑒𝑡𝑤𝑜𝑟𝑘 1
𝑠𝑢𝑚 𝑜𝑓 𝑎𝑏𝑠𝑜𝑙𝑢𝑡𝑒 𝑒𝑑𝑔𝑒𝑤𝑒𝑖𝑔ℎ𝑡𝑠 𝑜𝑓 𝑛𝑒𝑡𝑤𝑜𝑟𝑘 2= = 𝐶
![Page 20: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/20.jpg)
Intuition Of Our Solution
20
network 1
network 2
Graphical lasso for network 1
with constraint
Graphical lasso for network 2
with constraint
![Page 21: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/21.jpg)
Sharing Information Among Related
States
21
Often number of
samples per state is
very small.
“Share” samples among
related states
Song et al. 2009
Ahmed and Xing 2009
Parikh et al. 2011
These strategies can be
integrated into our robust
framework
![Page 22: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/22.jpg)
Summary of Method
Based on the graphical lasso [Meinshausen and Buhlmann 2006]
Add regularization constraint so that networks are
calibrated
Share information among related cell states by taking
weighted average of related samples
22
![Page 23: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/23.jpg)
Outline
Motivation
Challenge of sample size heterogeneity
Our solution
Results
Conclusion
23
![Page 24: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/24.jpg)
Robust algorithm produces more calibrated networks than
non robust version
Both methods produce comparable F1 score
Results on Synthetic Data
24
Mean vs Variance
Mean Number of Edges
Vari
an
ce
Robust
Non-Robust
(Scaled)
![Page 25: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/25.jpg)
Stem Cell Data
25
Hematopoietic stem cell data from Novershtern et al. 2011
38 cell states
4-7 samples per cell state
211 total samples
![Page 26: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/26.jpg)
Stem Cell Data Results
26
Granulocyte networkCommon Myleoid
Progenitor Network
Robust
Non-Robust
![Page 27: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/27.jpg)
Stem Cell Data: Biological Analysis
27
Noveshtern et al. 2011 doesn’t give
insight into module topology.
Experimentally validated
regulators (Novershtern et al.)
Experimentally validated genes
Non-experimentally validated
genes
granulocyte-subnetwork monocyte-subnetwork
![Page 28: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/28.jpg)
Stem Cell Data: Biological Analysis
28
granulocyte-subnetwork monocyte-subnetwork
Two modules with identical gene interaction patterns
large 10-gene module
small 2-gene module
7 out of these 12 genes experimentally validated by Novershtern et al.
![Page 29: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/29.jpg)
Conclusion
We addressed the novel problem of sample size
heterogeneity in the context of dynamic network
reconstruction
Our solution produces more calibrated networks allowing
for more meaningful biological comparison
We hope this can help efforts in personalized medicine and
data integration
29
![Page 30: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/30.jpg)
Acknowledgements
30
PSB Travel
Fellowship 2014
NSF Graduate
FellowshipNIH
![Page 31: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/31.jpg)
Thanks!
31
![Page 32: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/32.jpg)
Regulation of cell response to stimuli is
paramount, but we can usually only measure
(or compute) steady-state interactions
Transcriptionalinteractions
Protein—proteininteractions
Biochemicalreactions
▲ Chromatin IP
▲ Microarrays
▲ Protein coIP
▲ Yeast two-hybrid
▲ Metabolic flux measurements
Transcriptionalinteractions
Protein—proteininteractions
Biochemicalreactions
▲ Chromatin IP
▲ Microarrays
▲ Protein coIP
▲ Yeast two-hybrid
▲ Metabolic flux measurements
32
![Page 33: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/33.jpg)
Sparsity: In a mathematical sense Consider least squares linear regression problem:
Sparsity means most of the beta’s are zero.
But this is not convex!!! Many local optima, computationally
intractable.
33
𝑦
𝑥1
𝑥2
𝑥3
𝑥𝑛−1
𝑥𝑛
b1
b2
b3
bn-1
bn
𝜷 = argmin𝜷
𝒀 − 𝑿𝜷 22
𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜
𝑗=1
𝑝
𝐼 𝛽𝑗 > 0 ≤ 𝐶
![Page 34: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/34.jpg)
L1 Regularization (LASSO) [Tibshirani 1996]
A convex relaxation.
Still enforces sparsity!
34
Constrained Form Lagrangian Form
𝜷 = argmin𝜷
𝒀 − 𝑿𝜷 22
𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜
𝑗=1
𝑝
𝐼 𝛽𝑗 > 0 ≤ 𝐶
𝜷 = argmin𝜷
𝒀 − 𝑿𝜷 22 + 𝜆 𝜷 1
![Page 35: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/35.jpg)
Our Solution – Network Calibration
Calibrate networks to remove this sample size “bias”
Analogous to microarray processing to remove
systematic dye bias
Except our calibration is not a pre/post processing
step but rather a fundamental part of the method
35
![Page 36: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/36.jpg)
Sparsity!
One common assumption to make sparsity.
Makes biological sense: Genes are only assumed to
interface with small groups of other genes.
Makes statistical sense: Learning is now feasible in high
dimensions with small sample size
36
![Page 37: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/37.jpg)
Stem Cell Data: Quantitative
Results
37
![Page 38: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/38.jpg)
Intuition Of Our Solution
38
(something like this is true)
𝜷𝟏(𝒛), … , 𝜷𝒑
(𝒛)= argmin
𝜷𝟏(𝒛)
,…, 𝜷𝒑(𝒛)
𝑗=1
𝑝
𝑿𝒋(𝒛)
− 𝑿−𝒋𝟏𝜷𝒋(𝒛)
2
2
𝑗=1
𝑝
𝜷𝒋(𝒛)
1= 𝐶
𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜
𝑓𝑜𝑟 𝑎𝑙𝑙 𝑐𝑒𝑙𝑙 𝑠𝑡𝑎𝑡𝑒𝑠 𝒛
![Page 39: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/39.jpg)
Stem Cell Data Results
39
Granulocyte network
Common Myleoid Progenitor
Network
28 nodes with degree > 0 719 nodes with degree > 0
![Page 40: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/40.jpg)
40
Our Robust Approach
643 nodes with degree > 0 677 nodes with degree > 0
![Page 41: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/41.jpg)
41
Our Robust Approach
643 nodes with degree > 0 677 nodes with degree > 0
![Page 42: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/42.jpg)
Sample Size Heterogeneity for
Network Estimation
42
𝜷𝟏(𝟏), … , 𝜷𝒑
(𝟏)
= argmin𝜷𝟏(𝟏)
,…, 𝜷𝒑(𝟏)
𝑗=1
𝑝
𝑿𝒋(𝟏)
− 𝑿−𝒋𝟏𝜷𝒋(𝟏)
2
2+ 𝜆
𝑗=1
𝑝
𝜷𝒋(𝟏)
1
Increases as number of
samples increases stays constant as a
function of number of
samples
Thus applying the same 𝝀 will result in cell types with more samples
having more edges
![Page 43: Robust Reverse Engineering of Dynamic Gene …apparikh/talks/psb2014_robust...Robust Reverse Engineering of Dynamic Gene Networks Under Sample Size HeterogeneityAnkur P. Parikh, Wei](https://reader033.fdocuments.net/reader033/viewer/2022050111/5f487a325e52e5248862ff11/html5/thumbnails/43.jpg)
What About Scaling?
43
𝜷𝟏(𝟏), … , 𝜷𝒑
(𝟏)
= argmin𝜷𝟏(𝟏)
,…, 𝜷𝒑(𝟏)
𝑗=1
𝑝1
𝑆(1)𝑿𝒋(𝟏)
− 𝑿−𝒋𝟏𝜷𝒋(𝟏)
2
2
+𝜆
𝑆(1)
𝑗=1
𝑝
𝜷𝒋(𝟏)
1
Only works in theory, not in practice.
Requires restrictive modeling assumptions to hold
Requires larger sample size to kick in